More memory optimizations

README.md

@@ -119,7 +119,7 @@ const ai = new Ai({
         system: 'You are a helpful assistant.',
         compress: {max: 90_000, min: 50_000}, // Compress chat history to min tokens when max is reached
         temperature: 0.8,
-        max_tokens: 100_000,
+        maxTokens: 100_000,
         memoryModel: 'gpt-4o', // Cheap model for managing memories in background, defaults to current model
         models: {
             'claude-3-5-sonnet': {proto: 'anthropic', token: process.env.ANTHROPIC_TOKEN},
@@ -186,7 +186,7 @@ console.log(chunks);
 
 // Manually compile history into memories at end of conversation
 // Happens automatically when coverstaions are compressed
-await ai.language.updateMemory(history, memory);
+await ai.language.memorize(history, memory);
 
 // Summarize text
 const summary = await ai.language.summarize(longText, 200);

package-lock.json

@@ -1,21 +1,22 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "1.2.6",
+	"version": "1.6.6",
 	"lockfileVersion": 3,
 	"requires": true,
 	"packages": {
 		"": {
 			"name": "@ztimson/ai-utils",
-			"version": "1.2.6",
+			"version": "1.6.6",
 			"license": "MIT",
 			"dependencies": {
 				"@anthropic-ai/sdk": "^0.102.0",
 				"@huggingface/transformers": "^4.2.0",
 				"@tensorflow/tfjs": "^4.22.0",
 				"@ztimson/node-utils": "^1.0.7",
-				"@ztimson/utils": "^0.29.4",
+				"@ztimson/utils": "^0.30.8",
 				"cheerio": "^1.2.0",
 				"openai": "^6.42.0",
+				"pdf-parse": "^2.4.5",
 				"tesseract.js": "^7.0.0"
 			},
 			"devDependencies": {
@@ -56,39 +57,12 @@
 				"node": ">=6.9.0"
 			}
 		},
-		"node_modules/@emnapi/core": {
-			"version": "2.0.0-alpha.3",
-			"resolved": "https://registry.npmjs.org/@emnapi/core/-/core-2.0.0-alpha.3.tgz",
-			"integrity": "sha512-AZypUeJ/yByuxyS7BlSNRDOMLMlROYtjYdIAuBmJssVz1UJDSeYxLrdizhXCFYhedC5bqd/ASy8EuNXbVVXp9g==",
-			"dev": true,
-			"license": "MIT",
-			"optional": true,
-			"peer": true,
-			"dependencies": {
-				"@emnapi/wasi-threads": "2.0.1",
-				"tslib": "^2.4.0"
-			}
-		},
 		"node_modules/@emnapi/runtime": {
-			"version": "2.0.0-alpha.3",
+			"version": "1.11.3",
-			"resolved": "https://registry.npmjs.org/@emnapi/runtime/-/runtime-2.0.0-alpha.3.tgz",
+			"resolved": "https://registry.npmjs.org/@emnapi/runtime/-/runtime-1.11.3.tgz",
-			"integrity": "sha512-hFPAhMUjJD9BSyCANEISPOogeXC9Zo9ZQl7L6vKnaVsMkCtzznaW/naYypeyl0Gv5rYfWYsZbpixTMpjDJzQeA==",
+			"integrity": "sha512-Xz4Tpyki7XyrpbUK1jR1AhdAdaXyhhY4lZ3neLodmhpuWfy2PAQN5B46sAiU4liOXGLkHypn/qU+jvfWSCYYLA==",
-			"dev": true,
 			"license": "MIT",
 			"optional": true,
-			"peer": true,
-			"dependencies": {
-				"tslib": "^2.4.0"
-			}
-		},
-		"node_modules/@emnapi/wasi-threads": {
-			"version": "2.0.1",
-			"resolved": "https://registry.npmjs.org/@emnapi/wasi-threads/-/wasi-threads-2.0.1.tgz",
-			"integrity": "sha512-9DsSk+o5NBX0CCJT8s0EROGSGxjR/tKu6aBTaVyq+SjAEQH4XcdcRxPBRzsBLizTTJ49MJjF+jgu3qnO9GLQcQ==",
-			"dev": true,
-			"license": "MIT",
-			"optional": true,
-			"peer": true,
 			"dependencies": {
 				"tslib": "^2.4.0"
 			}
@@ -577,16 +551,6 @@
 				"url": "https://opencollective.com/libvips"
 			}
 		},
-		"node_modules/@img/sharp-wasm32/node_modules/@emnapi/runtime": {
-			"version": "1.11.3",
-			"resolved": "https://registry.npmjs.org/@emnapi/runtime/-/runtime-1.11.3.tgz",
-			"integrity": "sha512-Xz4Tpyki7XyrpbUK1jR1AhdAdaXyhhY4lZ3neLodmhpuWfy2PAQN5B46sAiU4liOXGLkHypn/qU+jvfWSCYYLA==",
-			"license": "MIT",
-			"optional": true,
-			"dependencies": {
-				"tslib": "^2.4.0"
-			}
-		},
 		"node_modules/@img/sharp-win32-arm64": {
 			"version": "0.34.5",
 			"resolved": "https://registry.npmjs.org/@img/sharp-win32-arm64/-/sharp-win32-arm64-0.34.5.tgz",
@@ -694,32 +658,209 @@
 				"@jridgewell/sourcemap-codec": "^1.4.14"
 			}
 		},
-		"node_modules/@napi-rs/wasm-runtime": {
+		"node_modules/@napi-rs/canvas": {
-			"version": "1.2.0",
+			"version": "0.1.80",
-			"resolved": "https://registry.npmjs.org/@napi-rs/wasm-runtime/-/wasm-runtime-1.2.0.tgz",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas/-/canvas-0.1.80.tgz",
-			"integrity": "sha512-kDoONqMa+VnZ4vvvu/ZUurpJ4gkZU57e7g69qpNgWhYcZFPUHZM2CEMKm+cG6ufDVALbjMvfmMjFVqaK7uEMnA==",
+			"integrity": "sha512-DxuT1ClnIPts1kQx8FBmkk4BQDTfI5kIzywAaMjQSXfNnra5UFU9PwurXrl+Je3bJ6BGsp/zmshVVFbCmyI+ww==",
-			"dev": true,
+			"license": "MIT",
+			"workspaces": [
+				"e2e/*"
+			],
+			"engines": {
+				"node": ">= 10"
+			},
+			"optionalDependencies": {
+				"@napi-rs/canvas-android-arm64": "0.1.80",
+				"@napi-rs/canvas-darwin-arm64": "0.1.80",
+				"@napi-rs/canvas-darwin-x64": "0.1.80",
+				"@napi-rs/canvas-linux-arm-gnueabihf": "0.1.80",
+				"@napi-rs/canvas-linux-arm64-gnu": "0.1.80",
+				"@napi-rs/canvas-linux-arm64-musl": "0.1.80",
+				"@napi-rs/canvas-linux-riscv64-gnu": "0.1.80",
+				"@napi-rs/canvas-linux-x64-gnu": "0.1.80",
+				"@napi-rs/canvas-linux-x64-musl": "0.1.80",
+				"@napi-rs/canvas-win32-x64-msvc": "0.1.80"
+			}
+		},
+		"node_modules/@napi-rs/canvas-android-arm64": {
+			"version": "0.1.80",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-android-arm64/-/canvas-android-arm64-0.1.80.tgz",
+			"integrity": "sha512-sk7xhN/MoXeuExlggf91pNziBxLPVUqF2CAVnB57KLG/pz7+U5TKG8eXdc3pm0d7Od0WreB6ZKLj37sX9muGOQ==",
+			"cpu": [
+				"arm64"
+			],
 			"license": "MIT",
 			"optional": true,
-			"dependencies": {
+			"os": [
-				"@tybys/wasm-util": "^0.10.3"
+				"android"
-			},
+			],
 			"engines": {
-				"node": "^20.19.0 || ^22.13.0 || >=23.5.0"
+				"node": ">= 10"
-			},
+			}
-			"funding": {
+		},
-				"type": "github",
+		"node_modules/@napi-rs/canvas-darwin-arm64": {
-				"url": "https://github.com/sponsors/Brooooooklyn"
+			"version": "0.1.80",
-			},
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-darwin-arm64/-/canvas-darwin-arm64-0.1.80.tgz",
-			"peerDependencies": {
+			"integrity": "sha512-O64APRTXRUiAz0P8gErkfEr3lipLJgM6pjATwavZ22ebhjYl/SUbpgM0xcWPQBNMP1n29afAC/Us5PX1vg+JNQ==",
-				"@emnapi/core": "^2.0.0-alpha.3",
+			"cpu": [
-				"@emnapi/runtime": "^2.0.0-alpha.3"
+				"arm64"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"darwin"
+			],
+			"engines": {
+				"node": ">= 10"
+			}
+		},
+		"node_modules/@napi-rs/canvas-darwin-x64": {
+			"version": "0.1.80",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-darwin-x64/-/canvas-darwin-x64-0.1.80.tgz",
+			"integrity": "sha512-FqqSU7qFce0Cp3pwnTjVkKjjOtxMqRe6lmINxpIZYaZNnVI0H5FtsaraZJ36SiTHNjZlUB69/HhxNDT1Aaa9vA==",
+			"cpu": [
+				"x64"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"darwin"
+			],
+			"engines": {
+				"node": ">= 10"
+			}
+		},
+		"node_modules/@napi-rs/canvas-linux-arm-gnueabihf": {
+			"version": "0.1.80",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-linux-arm-gnueabihf/-/canvas-linux-arm-gnueabihf-0.1.80.tgz",
+			"integrity": "sha512-eyWz0ddBDQc7/JbAtY4OtZ5SpK8tR4JsCYEZjCE3dI8pqoWUC8oMwYSBGCYfsx2w47cQgQCgMVRVTFiiO38hHQ==",
+			"cpu": [
+				"arm"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"linux"
+			],
+			"engines": {
+				"node": ">= 10"
+			}
+		},
+		"node_modules/@napi-rs/canvas-linux-arm64-gnu": {
+			"version": "0.1.80",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-linux-arm64-gnu/-/canvas-linux-arm64-gnu-0.1.80.tgz",
+			"integrity": "sha512-qwA63t8A86bnxhuA/GwOkK3jvb+XTQaTiVML0vAWoHyoZYTjNs7BzoOONDgTnNtr8/yHrq64XXzUoLqDzU+Uuw==",
+			"cpu": [
+				"arm64"
+			],
+			"libc": [
+				"glibc"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"linux"
+			],
+			"engines": {
+				"node": ">= 10"
+			}
+		},
+		"node_modules/@napi-rs/canvas-linux-arm64-musl": {
+			"version": "0.1.80",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-linux-arm64-musl/-/canvas-linux-arm64-musl-0.1.80.tgz",
+			"integrity": "sha512-1XbCOz/ymhj24lFaIXtWnwv/6eFHXDrjP0jYkc6iHQ9q8oXKzUX1Lc6bu+wuGiLhGh2GS/2JlfORC5ZcXimRcg==",
+			"cpu": [
+				"arm64"
+			],
+			"libc": [
+				"musl"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"linux"
+			],
+			"engines": {
+				"node": ">= 10"
+			}
+		},
+		"node_modules/@napi-rs/canvas-linux-riscv64-gnu": {
+			"version": "0.1.80",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-linux-riscv64-gnu/-/canvas-linux-riscv64-gnu-0.1.80.tgz",
+			"integrity": "sha512-XTzR125w5ZMs0lJcxRlS1K3P5RaZ9RmUsPtd1uGt+EfDyYMu4c6SEROYsxyatbbu/2+lPe7MPHOO/0a0x7L/gw==",
+			"cpu": [
+				"riscv64"
+			],
+			"libc": [
+				"glibc"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"linux"
+			],
+			"engines": {
+				"node": ">= 10"
+			}
+		},
+		"node_modules/@napi-rs/canvas-linux-x64-gnu": {
+			"version": "0.1.80",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-linux-x64-gnu/-/canvas-linux-x64-gnu-0.1.80.tgz",
+			"integrity": "sha512-BeXAmhKg1kX3UCrJsYbdQd3hIMDH/K6HnP/pG2LuITaXhXBiNdh//TVVVVCBbJzVQaV5gK/4ZOCMrQW9mvuTqA==",
+			"cpu": [
+				"x64"
+			],
+			"libc": [
+				"glibc"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"linux"
+			],
+			"engines": {
+				"node": ">= 10"
+			}
+		},
+		"node_modules/@napi-rs/canvas-linux-x64-musl": {
+			"version": "0.1.80",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-linux-x64-musl/-/canvas-linux-x64-musl-0.1.80.tgz",
+			"integrity": "sha512-x0XvZWdHbkgdgucJsRxprX/4o4sEed7qo9rCQA9ugiS9qE2QvP0RIiEugtZhfLH3cyI+jIRFJHV4Fuz+1BHHMg==",
+			"cpu": [
+				"x64"
+			],
+			"libc": [
+				"musl"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"linux"
+			],
+			"engines": {
+				"node": ">= 10"
+			}
+		},
+		"node_modules/@napi-rs/canvas-win32-x64-msvc": {
+			"version": "0.1.80",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-win32-x64-msvc/-/canvas-win32-x64-msvc-0.1.80.tgz",
+			"integrity": "sha512-Z8jPsM6df5V8B1HrCHB05+bDiCxjE9QA//3YrkKIdVDEwn5RKaqOxCJDRJkl48cJbylcrJbW4HxZbTte8juuPg==",
+			"cpu": [
+				"x64"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"win32"
+			],
+			"engines": {
+				"node": ">= 10"
 			}
 		},
 		"node_modules/@oxc-project/types": {
-			"version": "0.139.0",
+			"version": "0.144.0",
-			"resolved": "https://registry.npmjs.org/@oxc-project/types/-/types-0.139.0.tgz",
+			"resolved": "https://registry.npmjs.org/@oxc-project/types/-/types-0.144.0.tgz",
-			"integrity": "sha512-r9gHphtCs+1M7J0pw6Sn/hh/Wpa/iQrOOkrNAlVLF/gHq+/CJmHIWKKUUhdWjcD6CIa8idarspCsASiXCXvFUw==",
+			"integrity": "sha512-nuhZIOLuI6TFQ32I/WnUx+SCPY7SdSKwgnFHydAuoS1+Z4BRcaP+RRJmGzl9lw+0OFF7UmaESf7KQRXaNLHypg==",
 			"dev": true,
 			"license": "MIT",
 			"funding": {
@@ -784,9 +925,9 @@
 			"license": "BSD-3-Clause"
 		},
 		"node_modules/@rolldown/binding-android-arm64": {
-			"version": "1.1.5",
+			"version": "1.2.4",
-			"resolved": "https://registry.npmjs.org/@rolldown/binding-android-arm64/-/binding-android-arm64-1.1.5.tgz",
+			"resolved": "https://registry.npmjs.org/@rolldown/binding-android-arm64/-/binding-android-arm64-1.2.4.tgz",
-			"integrity": "sha512-lZg8fqIv2v7FF237bwMgzGZEJvGL79/s5knJ/i6FmsGF4XXlzccZ4jb+TrFIxtSSxFtIpdsgrPZeMk1I9AFcyQ==",
+			"integrity": "sha512-jHC2cnyKz5xU2fhECtFl8OZ83cYNt13GZQD+0uMJ/X3o+ijmd56okHhTUwxVSHPx1IRVIJEZ1/1pPzeLCU6XKA==",
 			"cpu": [
 				"arm64"
 			],
@@ -801,9 +942,9 @@
 			}
 		},
 		"node_modules/@rolldown/binding-darwin-arm64": {
-			"version": "1.1.5",
+			"version": "1.2.4",
-			"resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-arm64/-/binding-darwin-arm64-1.1.5.tgz",
+			"resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-arm64/-/binding-darwin-arm64-1.2.4.tgz",
-			"integrity": "sha512-51Bnx9pNiMRKSUNtBfySkNJ9vMU9Hh3I1ozDd6gyPPYzaXCfnptUcEZxXGYFn+ul2dtcMUiqGR1Yai2K10uoTw==",
+			"integrity": "sha512-Dc5mPD8F5F/FS8i01syd7FTF6yB2fVthH/TRkjwJkzUK6EpoxHtqvZQP5Zwq80/5z19TWYHIg1KOHboCgVx/aQ==",
 			"cpu": [
 				"arm64"
 			],
@@ -818,9 +959,9 @@
 			}
 		},
 		"node_modules/@rolldown/binding-darwin-x64": {
-			"version": "1.1.5",
+			"version": "1.2.4",
-			"resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-x64/-/binding-darwin-x64-1.1.5.tgz",
+			"resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-x64/-/binding-darwin-x64-1.2.4.tgz",
-			"integrity": "sha512-Tm+gbfC0aHu1tBA/JvKQh32S0K6YgCHkiAF4/W6xX0K0RmNuc94VeK419dJoE65R5aRxmo+noZQSWrAMF6yb6g==",
+			"integrity": "sha512-fpDm4oBo6SqLvWUYCmFhdde3U9KH2fRNNMeAnAPAIwxRL345xutL0EtEUcuoxsoazdJGv/MuDBQHlCDrtbvqOg==",
 			"cpu": [
 				"x64"
 			],
@@ -835,9 +976,9 @@
 			}
 		},
 		"node_modules/@rolldown/binding-freebsd-x64": {
-			"version": "1.1.5",
+			"version": "1.2.4",
-			"resolved": "https://registry.npmjs.org/@rolldown/binding-freebsd-x64/-/binding-freebsd-x64-1.1.5.tgz",
+			"resolved": "https://registry.npmjs.org/@rolldown/binding-freebsd-x64/-/binding-freebsd-x64-1.2.4.tgz",
-			"integrity": "sha512-JMzDKCCXq93YccG5gz3hvOs1oXRKAf0XYpfOS88e+wZrC8Iugj6j68867vrYZkvpDDpKn/KoKORThmchMpF6TA==",
+			"integrity": "sha512-rSJoreDE/HoIzoaib6MTp5jQtCTdMHKIvItAKT/ImS6Y6Ww76oUaeMyp4Vc/fAgd/ehji068IxetHXAnqUwN9A==",
 			"cpu": [
 				"x64"
 			],
@@ -852,9 +993,9 @@
 			}
 		},
 		"node_modules/@rolldown/binding-linux-arm-gnueabihf": {
-			"version": "1.1.5",
+			"version": "1.2.4",
-			"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm-gnueabihf/-/binding-linux-arm-gnueabihf-1.1.5.tgz",
+			"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm-gnueabihf/-/binding-linux-arm-gnueabihf-1.2.4.tgz",
-			"integrity": "sha512-uML21j2K5TfPGutKxub+M+nLjZIrWjXQ5Grx4lCe/nimTj9B4L63zHpjXLl4y0L3mcm2htEQIb06oCG/szerNw==",
+			"integrity": "sha512-/jm8OGHgn7oGaJu3i/qZI9spUGcJ+y/lk43ttQ/iO1tOd9NissG6o97bighBCiL+BKRngmcDuR6ikfwYdJmVuQ==",
 			"cpu": [
 				"arm"
 			],
@@ -869,9 +1010,9 @@
 			}
 		},
 		"node_modules/@rolldown/binding-linux-arm64-gnu": {
-			"version": "1.1.5",
+			"version": "1.2.4",
-			"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-gnu/-/binding-linux-arm64-gnu-1.1.5.tgz",
+			"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-gnu/-/binding-linux-arm64-gnu-1.2.4.tgz",
-			"integrity": "sha512-navSiuTMogvnQoZoM/v+l3ZWo50/NTwSHSzheABx/RCnmUPaKwq9qSo4Br2OYRs21+Fz8uFqITZM3H4opOB0/Q==",
+			"integrity": "sha512-tIP06BeD9EqvECBrPZ+sqdPlYrT+aYaAiu1wYziVx5elRK/ftm33JxVDy2bXGbr6J0CrtirCkR87/X5a2euEng==",
 			"cpu": [
 				"arm64"
 			],
@@ -889,9 +1030,9 @@
 			}
 		},
 		"node_modules/@rolldown/binding-linux-arm64-musl": {
-			"version": "1.1.5",
+			"version": "1.2.4",
-			"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-musl/-/binding-linux-arm64-musl-1.1.5.tgz",
+			"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-musl/-/binding-linux-arm64-musl-1.2.4.tgz",
-			"integrity": "sha512-lAryqH7IteztmCXQXk0etKj4wBQ7Gx5S6LjKhsgp9zb8I5bsuvU/2llH1hDQcjsFeqIsovMVN339/8pUDDBXxA==",
+			"integrity": "sha512-Ql1Q0EQqVThvn9VAVlwNzsUvbSFtCMGjLpRRi4pk5i7NZZ4n5ISiLMjHYtus4VQ2PvkSw24zyaCVsiS+sXPj1w==",
 			"cpu": [
 				"arm64"
 			],
@@ -909,9 +1050,9 @@
 			}
 		},
 		"node_modules/@rolldown/binding-linux-ppc64-gnu": {
-			"version": "1.1.5",
+			"version": "1.2.4",
-			"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-ppc64-gnu/-/binding-linux-ppc64-gnu-1.1.5.tgz",
+			"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-ppc64-gnu/-/binding-linux-ppc64-gnu-1.2.4.tgz",
-			"integrity": "sha512-fsK/sNBnxzBlL4O1JNrZakVQxPspqpED5dLtNsZS9oOKmtSpdNIzxH2kkol5HYTWJN47sE20ztMJPxfZ89qGOg==",
+			"integrity": "sha512-GjbjXD4XXfN19D0LZNbmiCBUoDiRACsYHr0yaIbbn8aFsXjHZifcYqu/W5Er5X2X990WjHXFrxarn5chzItorQ==",
 			"cpu": [
 				"ppc64"
 			],
@@ -929,9 +1070,9 @@
 			}
 		},
 		"node_modules/@rolldown/binding-linux-s390x-gnu": {
-			"version": "1.1.5",
+			"version": "1.2.4",
-			"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-s390x-gnu/-/binding-linux-s390x-gnu-1.1.5.tgz",
+			"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-s390x-gnu/-/binding-linux-s390x-gnu-1.2.4.tgz",
-			"integrity": "sha512-gLYb4BIadlfTOYT5gO503n8zQjXflgzpD0FcyKh0Mzx3rqCZKnHoJWV9xe1KXUJ5lx2JfcSHr/mhzS0PC/McAA==",
+			"integrity": "sha512-p5WR0NOwaRmJ/B1b6IjEFLLivwEsf3PrdBIhRbhTCQisbo2SvHHpG4ELB/+FgQNnB88LTOF86upmJmbvZdQ2lw==",
 			"cpu": [
 				"s390x"
 			],
@@ -949,9 +1090,9 @@
 			}
 		},
 		"node_modules/@rolldown/binding-linux-x64-gnu": {
-			"version": "1.1.5",
+			"version": "1.2.4",
-			"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-gnu/-/binding-linux-x64-gnu-1.1.5.tgz",
+			"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-gnu/-/binding-linux-x64-gnu-1.2.4.tgz",
-			"integrity": "sha512-FjcpEKUyJygHgs1o50VYNvkt5+7Le/VEdYt0AkRpkL33MnyQfwr8l5mXwMmfmTbyMPr5vJLC+8/Gd9gXnwU1QQ==",
+			"integrity": "sha512-4/GyVjmhR+Tc6HLJvwc1sOhPqAZtySiSMesOZyX6JQ5XBxoTDEMKQzvo07NIK6nTon/SivlZqvhzvuVBNQhObQ==",
 			"cpu": [
 				"x64"
 			],
@@ -969,9 +1110,9 @@
 			}
 		},
 		"node_modules/@rolldown/binding-linux-x64-musl": {
-			"version": "1.1.5",
+			"version": "1.2.4",
-			"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-musl/-/binding-linux-x64-musl-1.1.5.tgz",
+			"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-musl/-/binding-linux-x64-musl-1.2.4.tgz",
-			"integrity": "sha512-Me+PfPI2TMeOQk0gYWfLQZtTktrmzbr8cDboqX83XKc7UrgAi55gF+2dUkWdxd19n55Essp2yeca+O9N5rBxHg==",
+			"integrity": "sha512-l9eeLsCNvPpmSXUej0etw/J1eqV0Jj1D5G/xG6YTijmE6dkv6E2QezgWbTfQk63v952DPqrjOCoiqxq7Bw0YUQ==",
 			"cpu": [
 				"x64"
 			],
@@ -989,9 +1130,9 @@
 			}
 		},
 		"node_modules/@rolldown/binding-openharmony-arm64": {
-			"version": "1.1.5",
+			"version": "1.2.4",
-			"resolved": "https://registry.npmjs.org/@rolldown/binding-openharmony-arm64/-/binding-openharmony-arm64-1.1.5.tgz",
+			"resolved": "https://registry.npmjs.org/@rolldown/binding-openharmony-arm64/-/binding-openharmony-arm64-1.2.4.tgz",
-			"integrity": "sha512-yc5WrLzXks6zCQfn9Oxr8pORKyl/pF+QjHmW/Qx3qu0oyrrNC+y2JLTU1E2rcWYAmzlnqngWXHQjy51VzW70Vw==",
+			"integrity": "sha512-e0F355MSTMm3+UOqtV3L24gFUp2N5m1f8L/7d56deik6va+AXdrt9F8LbzGpeWGWRbZEDq4m8NVnJDeBtf9DZg==",
 			"cpu": [
 				"arm64"
 			],
@@ -1005,63 +1146,10 @@
 				"node": "^20.19.0 || >=22.12.0"
 			}
 		},
-		"node_modules/@rolldown/binding-wasm32-wasi": {
-			"version": "1.1.5",
-			"resolved": "https://registry.npmjs.org/@rolldown/binding-wasm32-wasi/-/binding-wasm32-wasi-1.1.5.tgz",
-			"integrity": "sha512-VbQGPX2b4r48TAMIM2cjgluIM1HYutm4pcTEJsle7iEP7sB1dFqtPLBVbdLAZCxy1txCcPxf4QFf4v8uvltPqA==",
-			"cpu": [
-				"wasm32"
-			],
-			"dev": true,
-			"license": "MIT",
-			"optional": true,
-			"dependencies": {
-				"@emnapi/core": "1.11.1",
-				"@emnapi/runtime": "1.11.1",
-				"@napi-rs/wasm-runtime": "^1.1.6"
-			},
-			"engines": {
-				"node": "^20.19.0 || >=22.12.0"
-			}
-		},
-		"node_modules/@rolldown/binding-wasm32-wasi/node_modules/@emnapi/core": {
-			"version": "1.11.1",
-			"resolved": "https://registry.npmjs.org/@emnapi/core/-/core-1.11.1.tgz",
-			"integrity": "sha512-RSvbQmHzdKzNsLYa/wHrbc3KN4sYLKAdPZxqiM2HATqv/SBk2/ENSHpvXGaLOMcsAyz0poEGqkmmKYG3OWiJEQ==",
-			"dev": true,
-			"license": "MIT",
-			"optional": true,
-			"dependencies": {
-				"@emnapi/wasi-threads": "1.2.2",
-				"tslib": "^2.4.0"
-			}
-		},
-		"node_modules/@rolldown/binding-wasm32-wasi/node_modules/@emnapi/runtime": {
-			"version": "1.11.1",
-			"resolved": "https://registry.npmjs.org/@emnapi/runtime/-/runtime-1.11.1.tgz",
-			"integrity": "sha512-vgj7R3y3Wgx24IQaGPA/R6YFXLHVMOZ0uVEyIQPaWs+rd1AzfEMXlAC22FYwO1XkKR6NPsq7mUandH8oIRdZFw==",
-			"dev": true,
-			"license": "MIT",
-			"optional": true,
-			"dependencies": {
-				"tslib": "^2.4.0"
-			}
-		},
-		"node_modules/@rolldown/binding-wasm32-wasi/node_modules/@emnapi/wasi-threads": {
-			"version": "1.2.2",
-			"resolved": "https://registry.npmjs.org/@emnapi/wasi-threads/-/wasi-threads-1.2.2.tgz",
-			"integrity": "sha512-c95qOXkHdydNKhscBTebqEC1CVAZpyqOfVfBzQ1qgzyl3gfeldUjIggDbIZgDKsHLgnsM+igH7TJ/eAasaVuMA==",
-			"dev": true,
-			"license": "MIT",
-			"optional": true,
-			"dependencies": {
-				"tslib": "^2.4.0"
-			}
-		},
 		"node_modules/@rolldown/binding-win32-arm64-msvc": {
-			"version": "1.1.5",
+			"version": "1.2.4",
-			"resolved": "https://registry.npmjs.org/@rolldown/binding-win32-arm64-msvc/-/binding-win32-arm64-msvc-1.1.5.tgz",
+			"resolved": "https://registry.npmjs.org/@rolldown/binding-win32-arm64-msvc/-/binding-win32-arm64-msvc-1.2.4.tgz",
-			"integrity": "sha512-gHv82k63z4qpV5+Q1y/12KrK0ltWBukVDI8nZcbT7Tt/ZlOIVwppazneq0F93oDxTo3IgAMEDIoQh3E2n6mVsw==",
+			"integrity": "sha512-AWLi0uBRYh6QlE7OKhiz+phZC0qwtij2QZmhmOdsLdFn64m7oMpooE9ICE3lhm9xMb4SpDo2WbHcxX1iFLFtqw==",
 			"cpu": [
 				"arm64"
 			],
@@ -1076,9 +1164,9 @@
 			}
 		},
 		"node_modules/@rolldown/binding-win32-x64-msvc": {
-			"version": "1.1.5",
+			"version": "1.2.4",
-			"resolved": "https://registry.npmjs.org/@rolldown/binding-win32-x64-msvc/-/binding-win32-x64-msvc-1.1.5.tgz",
+			"resolved": "https://registry.npmjs.org/@rolldown/binding-win32-x64-msvc/-/binding-win32-x64-msvc-1.2.4.tgz",
-			"integrity": "sha512-tTZuDBPw85tEN5PQi1pnEBzDy0Z49HtScLAbD5t6hyeU92A95pRWaSMw1GZZi/RwgSgUIl0xrSlXIT/9QzvYSA==",
+			"integrity": "sha512-UwSDJOg3dqCAejWdxclJjCsh3Qq4vLYMDxmyHqo1btz3stK2VqgwNd3mm5tuIwzSlGIQ/1H9Hr+Zn09mrezNqQ==",
 			"cpu": [
 				"x64"
 			],
@@ -1317,17 +1405,6 @@
 				"@tensorflow/tfjs-core": "4.22.0"
 			}
 		},
-		"node_modules/@tybys/wasm-util": {
-			"version": "0.10.3",
-			"resolved": "https://registry.npmjs.org/@tybys/wasm-util/-/wasm-util-0.10.3.tgz",
-			"integrity": "sha512-F3fo1MYrRJYL3zER0OUOmkutjr1Vp23m7OsSgp7nq4SP6OqX6C/56XFIPAl5bt3zaBRjmW7SGz3u/6LwFpYcOg==",
-			"dev": true,
-			"license": "MIT",
-			"optional": true,
-			"dependencies": {
-				"tslib": "^2.4.0"
-			}
-		},
 		"node_modules/@types/estree": {
 			"version": "1.0.9",
 			"resolved": "https://registry.npmjs.org/@types/estree/-/estree-1.0.9.tgz",
@@ -1448,9 +1525,9 @@
 			"license": "MIT"
 		},
 		"node_modules/@ztimson/utils": {
-			"version": "0.29.7",
+			"version": "0.30.8",
-			"resolved": "https://registry.npmjs.org/@ztimson/utils/-/utils-0.29.7.tgz",
+			"resolved": "https://registry.npmjs.org/@ztimson/utils/-/utils-0.30.8.tgz",
-			"integrity": "sha512-cjQ9+RjC5X7gKNA/hJHDf7OtyYCa+5E0PDc76lIaATwNAxXCSx2IO9r2wHiHtZGV5bldjnFmw7aOV8Jmq7SgKQ==",
+			"integrity": "sha512-+vBjcinqckqMHkP95xWiQeQz2E7Q1oS0b+Odjp+F9rvQ4z0US4JdodjqhEaqh+RAO/yP77x4Eu0a04yBB4HNdw==",
 			"license": "MIT",
 			"dependencies": {
 				"var-persist": "^1.0.1"
@@ -1544,9 +1621,9 @@
 			"license": "MIT"
 		},
 		"node_modules/brace-expansion": {
-			"version": "2.1.3",
+			"version": "2.1.4",
-			"resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-2.1.3.tgz",
+			"resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-2.1.4.tgz",
-			"integrity": "sha512-DRdx5neNsG/QXbniLFWi2YmC/68oeOOmKz6zOjVk6ZS1ZLXgLIKqVEc6hWsmkjBbgii0SwaBTcJ5XKj5gzY/4A==",
+			"integrity": "sha512-hGfVzPxthbf3+2yjg/RBs60cB0FhqBS/zvdV/4wn4/BmN0bNMMHPc4V/BbFieqf1TKAGGAHnY4eSjajCl0f2Xg==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
@@ -3011,9 +3088,9 @@
 			"license": "MIT"
 		},
 		"node_modules/nanoid": {
-			"version": "3.3.16",
+			"version": "3.3.18",
-			"resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.16.tgz",
+			"resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.18.tgz",
-			"integrity": "sha512-bzlKTyNJ7+LdGIIwy8ijFpIqEQIvafahV7eYykJ8Cvh42EdJeODoJ6gUJXpQJvej1BddH8OqTXZNE/KfbWAu8Q==",
+			"integrity": "sha512-DTg4MJbGMWkfi6VZFdNt2/caMbQy4Ou+Op/hJQvGEWcnVfoA1QA+xzRKAzw9jD6+GVOOeYr/mIcuDSdug6F6+w==",
 			"dev": true,
 			"funding": [
 				{
@@ -3233,6 +3310,38 @@
 			"dev": true,
 			"license": "MIT"
 		},
+		"node_modules/pdf-parse": {
+			"version": "2.4.5",
+			"resolved": "https://registry.npmjs.org/pdf-parse/-/pdf-parse-2.4.5.tgz",
+			"integrity": "sha512-mHU89HGh7v+4u2ubfnevJ03lmPgQ5WU4CxAVmTSh/sxVTEDYd1er/dKS/A6vg77NX47KTEoihq8jZBLr8Cxuwg==",
+			"license": "Apache-2.0",
+			"dependencies": {
+				"@napi-rs/canvas": "0.1.80",
+				"pdfjs-dist": "5.4.296"
+			},
+			"bin": {
+				"pdf-parse": "bin/cli.mjs"
+			},
+			"engines": {
+				"node": ">=20.16.0 <21 || >=22.3.0"
+			},
+			"funding": {
+				"type": "github",
+				"url": "https://github.com/sponsors/mehmet-kozan"
+			}
+		},
+		"node_modules/pdfjs-dist": {
+			"version": "5.4.296",
+			"resolved": "https://registry.npmjs.org/pdfjs-dist/-/pdfjs-dist-5.4.296.tgz",
+			"integrity": "sha512-DlOzet0HO7OEnmUmB6wWGJrrdvbyJKftI1bhMitK7O2N8W2gc757yyYBbINy9IDafXAV9wmKr9t7xsTaNKRG5Q==",
+			"license": "Apache-2.0",
+			"engines": {
+				"node": ">=20.16.0 || >=22.3.0"
+			},
+			"optionalDependencies": {
+				"@napi-rs/canvas": "^0.1.80"
+			}
+		},
 		"node_modules/picocolors": {
 			"version": "1.1.1",
 			"resolved": "https://registry.npmjs.org/picocolors/-/picocolors-1.1.1.tgz",
@@ -3272,9 +3381,9 @@
 			"license": "MIT"
 		},
 		"node_modules/postcss": {
-			"version": "8.5.25",
+			"version": "8.5.26",
-			"resolved": "https://registry.npmjs.org/postcss/-/postcss-8.5.25.tgz",
+			"resolved": "https://registry.npmjs.org/postcss/-/postcss-8.5.26.tgz",
-			"integrity": "sha512-DTPx3RWSSnWyzLxQnlH0rJP+EW5ekl16ZU4/psbIhA0e53kJfdgaN5vKM+xP7yJtXVu+nfdVFmlgFDEKAe4Pyw==",
+			"integrity": "sha512-u82N74LFzG8ca+dD8puPnplTXoGH4fTPpVGuIbt36G3qvNlkvfD0lEAZSxaly3KX8TS/L1A1gsCEmvKmBcVbkQ==",
 			"dev": true,
 			"funding": [
 				{
@@ -3292,7 +3401,7 @@
 			],
 			"license": "MIT",
 			"dependencies": {
-				"nanoid": "^3.3.16",
+				"nanoid": "^3.3.17",
 				"picocolors": "^1.1.1",
 				"source-map-js": "^1.2.1"
 			},
@@ -3434,13 +3543,13 @@
 			"license": "BSD-3-Clause"
 		},
 		"node_modules/rolldown": {
-			"version": "1.1.5",
+			"version": "1.2.4",
-			"resolved": "https://registry.npmjs.org/rolldown/-/rolldown-1.1.5.tgz",
+			"resolved": "https://registry.npmjs.org/rolldown/-/rolldown-1.2.4.tgz",
-			"integrity": "sha512-t9z29cJjXf/vxQ8dyhCSpt6H6aSwHTk8cT5I3iy6SMXuFpk5mB6PL6XfC8PCwrPTx93udwKUm9HRteAlTGBLiA==",
+			"integrity": "sha512-rSr7irW0K7QRWzjdJXqZowkcRdDtjRduh43rBltnVKd0VFq839l1lJoDvGJb6gl7+4rTTCrPWu+YfujUL8Ug7w==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
-				"@oxc-project/types": "=0.139.0",
+				"@oxc-project/types": "=0.144.0",
 				"@rolldown/pluginutils": "^1.0.0"
 			},
 			"bin": {
@@ -3450,21 +3559,20 @@
 				"node": "^20.19.0 || >=22.12.0"
 			},
 			"optionalDependencies": {
-				"@rolldown/binding-android-arm64": "1.1.5",
+				"@rolldown/binding-android-arm64": "1.2.4",
-				"@rolldown/binding-darwin-arm64": "1.1.5",
+				"@rolldown/binding-darwin-arm64": "1.2.4",
-				"@rolldown/binding-darwin-x64": "1.1.5",
+				"@rolldown/binding-darwin-x64": "1.2.4",
-				"@rolldown/binding-freebsd-x64": "1.1.5",
+				"@rolldown/binding-freebsd-x64": "1.2.4",
-				"@rolldown/binding-linux-arm-gnueabihf": "1.1.5",
+				"@rolldown/binding-linux-arm-gnueabihf": "1.2.4",
-				"@rolldown/binding-linux-arm64-gnu": "1.1.5",
+				"@rolldown/binding-linux-arm64-gnu": "1.2.4",
-				"@rolldown/binding-linux-arm64-musl": "1.1.5",
+				"@rolldown/binding-linux-arm64-musl": "1.2.4",
-				"@rolldown/binding-linux-ppc64-gnu": "1.1.5",
+				"@rolldown/binding-linux-ppc64-gnu": "1.2.4",
-				"@rolldown/binding-linux-s390x-gnu": "1.1.5",
+				"@rolldown/binding-linux-s390x-gnu": "1.2.4",
-				"@rolldown/binding-linux-x64-gnu": "1.1.5",
+				"@rolldown/binding-linux-x64-gnu": "1.2.4",
-				"@rolldown/binding-linux-x64-musl": "1.1.5",
+				"@rolldown/binding-linux-x64-musl": "1.2.4",
-				"@rolldown/binding-openharmony-arm64": "1.1.5",
+				"@rolldown/binding-openharmony-arm64": "1.2.4",
-				"@rolldown/binding-wasm32-wasi": "1.1.5",
+				"@rolldown/binding-win32-arm64-msvc": "1.2.4",
-				"@rolldown/binding-win32-arm64-msvc": "1.1.5",
+				"@rolldown/binding-win32-x64-msvc": "1.2.4"
-				"@rolldown/binding-win32-x64-msvc": "1.1.5"
 			}
 		},
 		"node_modules/safe-buffer": {
@@ -4021,16 +4129,16 @@
 			}
 		},
 		"node_modules/vite": {
-			"version": "8.1.5",
+			"version": "8.2.1",
-			"resolved": "https://registry.npmjs.org/vite/-/vite-8.1.5.tgz",
+			"resolved": "https://registry.npmjs.org/vite/-/vite-8.2.1.tgz",
-			"integrity": "sha512-7ULLwsCdYx/nRyrpiEwvqb5TFHrMVZyBt+rg/OAXT7rgj/z+DtTDyKFeLAdDkubDVDKD8jOsndmy7m55XcfUsw==",
+			"integrity": "sha512-EU/eS7BH3XROHh2YnBefjM6DBKA6ZeMZEYQbj7NLWg5wHYlhB8B/Mayd5XsgWq+NFYccDOTemRpdETWR6Ka/lw==",
 			"dev": true,
 			"license": "MIT",
 			"dependencies": {
-				"lightningcss": "^1.32.0",
+				"lightningcss": "^1.33.0",
 				"picomatch": "^4.0.5",
-				"postcss": "^8.5.17",
+				"postcss": "^8.5.25",
-				"rolldown": "~1.1.5",
+				"rolldown": "~1.2.1",
 				"tinyglobby": "^0.2.17"
 			},
 			"bin": {
@@ -4047,7 +4155,7 @@
 			},
 			"peerDependencies": {
 				"@types/node": "^20.19.0 || >=22.12.0",
-				"@vitejs/devtools": "^0.3.0",
+				"@vitejs/devtools": "^0.4.0",
 				"esbuild": "^0.27.0 || ^0.28.0",
 				"jiti": ">=1.21.0",
 				"less": "^4.0.0",
@@ -4132,9 +4240,9 @@
 			"license": "MIT"
 		},
 		"node_modules/wasm-feature-detect": {
-			"version": "1.8.0",
+			"version": "1.9.0",
-			"resolved": "https://registry.npmjs.org/wasm-feature-detect/-/wasm-feature-detect-1.8.0.tgz",
+			"resolved": "https://registry.npmjs.org/wasm-feature-detect/-/wasm-feature-detect-1.9.0.tgz",
-			"integrity": "sha512-zksaLKM2fVlnB5jQQDqKXXwYHLQUVH9es+5TOOHwGOVJOCeRBCiPjwSg+3tN2AdTCzjgli4jijCH290kXb/zWQ==",
+			"integrity": "sha512-zonE+xlIIYtxPy++L24ow0hAD8CICb4+FgPyROd3buyXIqsJvUEDkBgfCCoXOd1Hu3DUr0GOfnPIdcGV+YpNaA==",
 			"license": "Apache-2.0"
 		},
 		"node_modules/webidl-conversions": {

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "1.3.5",
+	"version": "1.7.2",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -26,12 +26,13 @@
 	},
 	"dependencies": {
 		"@anthropic-ai/sdk": "^0.102.0",
-		"@tensorflow/tfjs": "^4.22.0",
 		"@huggingface/transformers": "^4.2.0",
+		"@tensorflow/tfjs": "^4.22.0",
 		"@ztimson/node-utils": "^1.0.7",
-		"@ztimson/utils": "^0.29.4",
+		"@ztimson/utils": "^0.30.8",
 		"cheerio": "^1.2.0",
 		"openai": "^6.42.0",
+		"pdf-parse": "^2.4.5",
 		"tesseract.js": "^7.0.0"
 	},
 	"devDependencies": {

src/ai.ts

@@ -4,7 +4,7 @@ import { Audio } from './audio.ts';
 import {Vision} from './vision.ts';
 
 export type AbortablePromise<T> = Promise<T> & {
-	abort: () => any
+	abort: (keep?: boolean) => any
 };
 
 export type AiOptions = {

src/antrhopic.ts

@@ -1,65 +1,63 @@
 import {Anthropic as anthropic} from '@anthropic-ai/sdk';
-import {findByProp, objectMap, JSONSanitize, JSONAttemptParse} from '@ztimson/utils';
+import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, makeArray} from '@ztimson/utils';
 import {AbortablePromise, Ai} from './ai.ts';
 import {LLMMessage, LLMRequest} from './llm.ts';
 import {LLMProvider} from './provider.ts';
+import {TokenPool} from './token-pool.ts';
 import {convertSchema} from './tools.ts';
 
 export class Anthropic extends LLMProvider {
-	client!: anthropic;
+	private clients = new Map<string, anthropic>();
+	tokenPool!: TokenPool;
 
-	constructor(public readonly ai: Ai, public readonly apiToken: string, public model: string) {
+	constructor(public readonly ai: Ai, public readonly apiToken: string | string[], public model: string) {
 		super();
-		this.client = new anthropic({apiKey: apiToken});
+		this.tokenPool = new TokenPool(...makeArray(apiToken).filter(Boolean));
 	}
 
-	private toStandard(history: any[]): LLMMessage[] {
+	private getClient(token: string): anthropic {
-		const timestamp = Date.now();
+		let client = this.clients.get(token);
-		const messages: LLMMessage[] = [];
+		if(!client) {
-		for(let h of history) {
+			client = new anthropic({apiKey: token});
-			if(typeof h.content == 'string') {
+			this.clients.set(token, client);
-				messages.push(<any>{timestamp, ...h});
-			} else {
-				const textContent = h.content?.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
-				if(textContent) messages.push({role: h.role, content: textContent, timestamp: timestamp, duration: h.duration, tps: h.tps});
-				h.content.forEach((c: any) => {
-					if(c.type == 'tool_use') {
-						messages.push({role: 'tool', id: c.id, name: c.name, args: c.input, timestamp: h.timestamp, content: undefined, duration: h.duration, tps: h.tps});
-					} else if(c.type == 'tool_result') {
-						const m: any = messages.findLast(m => (<any>m).id == c.tool_use_id);
-						if(m) m[c.is_error ? 'error' : 'content'] = c.content;
-					}
-				});
-			}
 		}
-		return messages;
+		return client;
 	}
 
-	private fromStandard(history: LLMMessage[]): any[] {
+	private toWireContent(content: any): any {
-		for(let i = 0; i < history.length; i++) {
+		if(!Array.isArray(content)) return content;
-			if(history[i].role == 'tool') {
+		return content.map(c => c.type === 'image'
-				const h: any = history[i];
+			? {type: 'image', source: {type: 'base64', media_type: c.mime, data: c.data}}
-				history.splice(i, 1,
+			: {type: 'text', text: c.text});
+	}
+
+	/** Convert standard history -> Anthropic wire format */
+	private toWire(history: LLMMessage[]): any[] {
+		const wire: any[] = [];
+		for(const h of history) {
+			if(h.role === 'tool') {
+				wire.push(
 					{role: 'assistant', content: [{type: 'tool_use', id: h.id, name: h.name, input: h.args}]},
-					{role: 'user', content: [{type: 'tool_result', tool_use_id: h.id, is_error: !!h.error, content:  h.error || h.content}]}
+					{role: 'user', content: [{type: 'tool_result', tool_use_id: h.id, is_error: !!h.error, content: h.error || h.content || ''}]}
-				)
+				);
-				i++;
+			} else {
+				wire.push({role: h.role, content: this.toWireContent(h.content)});
 			}
 		}
-		return history;
+		return wire;
 	}
 
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<string | any> {
 		const controller = new AbortController();
-		return Object.assign(new Promise<any>(async (res) => {
+		return Object.assign(new Promise<any>(async (res, rej) => {
-			let history = this.fromStandard([
+			if(!options.history) options.history = [];
-				...(options.history || []).filter(h => h.role !== 'system'),
+			const history = options.history;
-				{role: 'user', content: message, timestamp: Date.now()}
+			if(message) history.push({role: 'user', content: message, timestamp: Date.now()});
-			]);
+
 			const tools = options.tools || this.ai.options.llm?.tools || [];
 			const requestParams: any = {
 				model: options.model || this.model,
-				max_tokens: options.max_tokens || this.ai.options.llm?.max_tokens || 4096,
+				max_tokens: options.maxTokens || this.ai.options.llm?.maxTokens || 4096,
 				system: options.system || this.ai.options.llm?.system || '',
 				temperature: options.temperature || this.ai.options.llm?.temperature || undefined,
 				tools: tools.map(t => ({
@@ -69,105 +67,97 @@ export class Anthropic extends LLMProvider {
 						type: 'object',
 						properties: t.args ? objectMap(t.args, (key, value) => ({...value, required: undefined})) : {},
 						required: t.args ? Object.entries(t.args).filter(t => t[1].required).map(t => t[0]) : []
-					},
+					}
-					fn: undefined
 				})),
-				messages: history,
 				stream: !!options.stream,
 			};
 
-			// Add structured output support
 			if(options.schema) {
-				requestParams.output_config = {
+				requestParams.output_config = {format: {type: 'json_schema', schema: convertSchema(options.schema)}};
-					format: {
-						type: 'json_schema',
-						schema: convertSchema(options.schema)
-					}
-				};
 			}
 
-			let resp: any, terminal = false, duration = 0, tps = 0;
+			try {
-			do {
+				let terminal = false;
-				requestParams.messages = history.map(({timestamp, ...m}) => m);
+				do {
-				const callStart = Date.now();
+					requestParams.messages = this.toWire(history.filter(h => h.role !== 'system'));
-				resp = await this.client.messages.create(requestParams).catch(err => {
-					err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
-					throw err;
-				});
 
-				let usage: any;
+					const callStart = Date.now();
-				if(options.stream) {
+					const resp: any = await this.tokenPool.run(token => this.getClient(token).messages.create(requestParams)).catch(err => {
-					resp.content = [];
+						err.message += `\n\nMessages:\n${JSON.stringify(requestParams.messages, null, 2)}`;
-					for await (const chunk of resp) {
+						throw err;
-						if(controller.signal.aborted) break;
+					});
-						if(chunk.type === 'content_block_start') {
+
-							if(chunk.content_block.type === 'text') {
+					let usage: any, content: any[] = [];
-								resp.content.push({type: 'text', text: ''});
+					if(options.stream) {
-							} else if(chunk.content_block.type === 'tool_use') {
+						for await (const chunk of resp) {
-								resp.content.push({type: 'tool_use', id: chunk.content_block.id, name: chunk.content_block.name, input: <any>''});
+							if(controller.signal.aborted) break;
+							if(chunk.type === 'content_block_start') {
+								if(chunk.content_block.type === 'text') content.push({type: 'text', text: ''});
+								else if(chunk.content_block.type === 'tool_use') content.push({type: 'tool_use', id: chunk.content_block.id, name: chunk.content_block.name, input: ''});
+							} else if(chunk.type === 'content_block_delta') {
+								if(chunk.delta.type === 'text_delta') {
+									content.at(-1).text += chunk.delta.text;
+									options.stream({text: chunk.delta.text});
+								} else if(chunk.delta.type === 'input_json_delta') {
+									content.at(-1).input += chunk.delta.partial_json;
+								}
+							} else if(chunk.type === 'content_block_stop') {
+								const last = content.at(-1);
+								if(last?.type === 'tool_use') last.input = last.input ? JSONAttemptParse(last.input, {}) : {};
+							} else if(chunk.type === 'message_delta') {
+								if(chunk.usage) usage = chunk.usage;
+							} else if(chunk.type === 'message_stop') {
+								break;
 							}
-						} else if(chunk.type === 'content_block_delta') {
-							if(chunk.delta.type === 'text_delta') {
-								const text = chunk.delta.text;
-								resp.content.at(-1).text += text;
-								options.stream({text});
-							} else if(chunk.delta.type === 'input_json_delta') {
-								resp.content.at(-1).input += chunk.delta.partial_json;
-							}
-						} else if(chunk.type === 'content_block_stop') {
-							const last = resp.content.at(-1);
-							if(last?.input != null) last.input = last.input ? JSONAttemptParse(last.input, {}) : {};
-						} else if(chunk.type === 'message_delta') {
-							if(chunk.usage) usage = chunk.usage;
-						} else if(chunk.type === 'message_stop') {
-							break;
 						}
+					} else {
+						usage = resp.usage;
+						content = resp.content;
 					}
-				} else {
+					const duration = Date.now() - callStart;
-					usage = resp.usage;
+					const tps = usage?.output_tokens && duration > 0 ? usage.output_tokens / (duration / 1000) : 0;
-				}
-				duration = Date.now() - callStart;
-				tps = usage?.output_tokens && duration > 0 ? usage.output_tokens / (duration / 1000) : 0;
 
-				const toolCalls = resp.content.filter((c: any) => c.type === 'tool_use');
+					const toolCalls = content.filter((c: any) => c.type === 'tool_use');
-				if(toolCalls.length && !controller.signal.aborted) {
+					if(toolCalls.length && !controller.signal.aborted) {
-					history.push({role: 'assistant', content: resp.content, timestamp: Date.now(), duration, tps});
+						const text = content.filter((c: any) => c.type === 'text').map((c: any) => c.text).join('\n\n').trim();
-					const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
+						if(text) history.push({role: 'assistant', content: text, timestamp: Date.now(), duration, tps});
-						const tool = tools.find(findByProp('name', toolCall.name));
-						if(options.stream) options.stream({tool: toolCall.name});
-						if(!tool) return {tool_use_id: toolCall.id, is_error: true, content: 'Tool not found'};
-						try {
-							const toolStream = options.stream && ((chunk: any) => {
-								if(chunk.done) { terminal = true; return; }
-								options.stream!(chunk);
-							});
-							const result = await tool.fn(toolCall.input, toolStream, this.ai, toolCall.id);
-							return {type: 'tool_result', tool_use_id: toolCall.id, content: typeof result == 'object' ? JSONSanitize(result) : result};
-						} catch (err: any) {
-							return {type: 'tool_result', tool_use_id: toolCall.id, is_error: true, content: err?.message || err?.toString() || 'Unknown'};
-						}
-					}));
-					history.push({role: 'user', content: results, timestamp: Date.now()});
-					requestParams.messages = history;
-				}
-			} while (!terminal && !controller.signal.aborted && resp.content.some((c: any) => c.type === 'tool_use'));
 
-			if(!terminal) {
+						const entries = toolCalls.map((tc: any) => {
-				const textContent = resp.content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
+							const entry: any = {role: 'tool', id: tc.id, name: tc.name, args: tc.input, content: undefined, timestamp: Date.now()};
-				history.push({role: 'assistant', content: textContent.trim(), timestamp: Date.now(), duration, tps});
+							history.push(entry);
+							return {tc, entry};
+						});
+
+						await Promise.all(entries.map(async ({tc, entry}: any) => {
+							const tool = tools.find(findByProp('name', tc.name));
+							if(options.stream) options.stream({tool: tc.name});
+							if(!tool) { entry.error = 'Tool not found'; return; }
+							try {
+								const toolStream = options.stream && ((chunk: any) => {
+									if(chunk.done) { terminal = true; return; }
+									options.stream!(chunk);
+								});
+								const result = await tool.fn(entry.args, toolStream, this.ai, tc.id);
+								entry.content = typeof result === 'object' ? JSONSanitize(result) : result;
+							} catch(err: any) {
+								entry.error = err?.message || err?.toString() || 'Unknown';
+							}
+						}));
+					} else {
+						terminal = true;
+						const text = content.filter((c: any) => c.type === 'text').map((c: any) => c.text).join('\n\n').trim();
+						if(text) history.push({role: 'assistant', content: text, timestamp: Date.now(), duration, tps});
+					}
+				} while(!terminal && !controller.signal.aborted);
+
+				if(options.stream) options.stream({done: true});
+
+				const turnStart = history.map(h => h.role).lastIndexOf('user');
+				const finalContent = history.slice(turnStart + 1).reduce((str, h) => h.role === 'assistant' ? str + (h.content || '') : str, '').trim();
+				res(options.schema ? JSONAttemptParse(finalContent, finalContent) : finalContent);
+			} catch(err) {
+				rej(err);
 			}
-
-			history = this.toStandard(history);
-			if(options.history) options.history.splice(0, options.history.length, ...history);
-			if(options.stream) options.stream({done: true});
-
-			const turnStart = history.map(h => h.role).lastIndexOf('user');
-			const finalContent = history.slice(turnStart + 1).reduce((str, h) => {
-				if(h.role === 'assistant') return str + (h.content || '');
-				return str;
-			}, '').trim();
-
-			res(options.schema ? JSONAttemptParse(finalContent, finalContent) : finalContent);
 		}), {abort: () => controller.abort()});
 	}
 }

src/index.ts

@@ -2,8 +2,13 @@ export * from './ai';
 export * from './antrhopic';
 export * from './audio';
 export * from './llm';
-export * from './memory';
+export * from './memory/graph';
+export * from './memory/kd-tree';
+export * from './memory/memory';
+export * from './memory/memory-state';
 export * from './open-ai';
 export * from './provider';
+export * from './token-pool'
 export * from './tools';
 export * from './vision';
+export * from './utils';

src/llm.ts

@@ -1,16 +1,32 @@
-import {snakeCase} from '@ztimson/utils';
+import {clean, makeUnique, snakeCase} from '@ztimson/utils';
 import {AbortablePromise, Ai} from './ai.ts';
 import {Anthropic} from './antrhopic.ts';
+import {MemoryCache} from './memory/memory-state.ts';
+import {Memory, MemoryManager, MemoryOptions} from './memory/memory.ts';
 import {OpenAi} from './open-ai.ts';
 import {LLMProvider} from './provider.ts';
 import {AiTool, AiToolArg} from './tools.ts';
 import {fileURLToPath} from 'url';
-import {dirname, join} from 'path';
 import {spawn} from 'node:child_process';
-import {Memory, MemoryCache, MemoryManager, MemoryOptions} from './memory.ts';
+import {mkdtempSync} from 'node:fs';
+import fs from 'node:fs/promises';
+import {tmpdir} from 'node:os';
+import {dirname, join, basename, extname} from 'path';
+import { PDFParse } from 'pdf-parse';
+import {stripHeader} from './utils.ts';
 
-export type AnthropicConfig = {proto: 'anthropic', token: string};
+const MAX_AGENT_DEPTH = 5;
-export type OpenAiConfig = {proto: 'openai', host?: string, token: string};
+const PDF_OCR_PAGE_THRESHOLD = 12; // above this many pages, OCR scanned pages instead of feeding images to the model
+
+export type AnthropicConfig = {proto: 'anthropic', token: string | string[]};
+export type OpenAiConfig = {proto: 'openai', host?: string, token: string | string[]};
+
+export type AgentRef = {
+	name: string;
+	description?: string;
+	delegate?: boolean;
+	fn: () => Agent | null | Promise<Agent | null>;
+}
 
 export type Agent = {
 	name: string;
@@ -22,16 +38,35 @@ export type Agent = {
 	skills?: Skill[] | null;
 	tools?: AiTool[] | null;
 	mcp?: McpServer[] | null;
-	agents?: string[] | null;
+	agents?: AgentRef[] | null;
 }
 
+export type LLMFile = {
+	/** Path to file on disk */
+	path?: string;
+	/** File content: raw text, base64-encoded binary, or a Buffer */
+	content?: string | Buffer;
+	/** Original filename, used to infer type from extension */
+	name?: string;
+	/** Mime type override, inferred from extension if omitted */
+	mime?: string;
+	/** @internal set once extraction has run, skips re-processing next turn */
+	extracted?: boolean;
+};
+
 export type LLMMessage = {
 	/** Message originator */
 	role: 'assistant' | 'system' | 'user';
 	/** Message content */
 	content: string | any;
+	/** Files attached to request */
+	files?: LLMFile[];
 	/** Timestamp */
 	timestamp?: number;
+	/** Response duration in ms */
+	duration?: number;
+	/** Tokens per second */
+	tps?: number;
 } | {
 	/** Tool call */
 	role: 'tool';
@@ -61,7 +96,7 @@ export type LLMRequest = {
 	/** Message history */
 	history?: LLMMessage[];
 	/** Max tokens for request */
-	max_tokens?: number;
+	maxTokens?: number;
 	/** 0 = Rigid Logic, 1 = Balanced, 2 = Hyper Creative **/
 	temperature?: number;
 	/** Available tools */
@@ -80,8 +115,10 @@ export type LLMRequest = {
 	skills?: Skill[];
 	/** MCP servers to connect and expose as tools */
 	mcp?: McpServer[];
-	/** Subagents exposed as delegatable/wrapped tools */
+	/** Subagents exposed as delegatable/wrapped tools, resolved lazily via their `fn` */
-	agents?: Agent[];
+	agents?: AgentRef[];
+	/** Attach files to request */
+	files?: LLMFile[];
 	/** @internal recursion guard for nested agent delegation */
 	_agentDepth?: number;
 }
@@ -104,9 +141,12 @@ export type Skill = {
 	content: string;
 }
 
-const MAX_AGENT_DEPTH = 5;
-
 class LLM {
+	private static AUDIO_EXT = ['wav','mp3','m4a','flac','ogg','aac','wma'];
+	private static IMAGE_EXT = ['png','jpg','jpeg','bmp','gif','tiff','webp'];
+	private static TEXT_EXT = ['txt','md','csv','json','xml','html','js','ts','py','yaml','yml','log'];
+	private static PDF_EXT = ['pdf'];
+
 	private memoryManager!: MemoryManager;
 
 	defaultModel!: string;
@@ -122,51 +162,157 @@ class LLM {
 		this.memoryManager = new MemoryManager(this);
 	}
 
-	private setupAgent(agents: Agent[] = [], allAgents: Agent[], pending: Map<string, any>, aborts: (() => void)[], depth = 0): AiTool[] {
+	private async loadBuffer(file: LLMFile, asText: boolean): Promise<Buffer> {
-		return agents.map(a => {
+		if(file.path) return fs.readFile(file.path);
-			const toolName = `${a.delegate ? '' : 'sub'}agent_${snakeCase(a.name)}`;
+		if(Buffer.isBuffer(file.content)) return file.content;
+		if(typeof file.content === 'string') return Buffer.from(file.content, asText ? 'utf-8' : 'base64');
+		throw new Error('No path or content provided');
+	}
+
+	private async writeTemp(name: string, buffer: Buffer): Promise<string> {
+		const path = join(mkdtempSync(join(tmpdir(), 'ai-file-')), name);
+		await fs.writeFile(path, buffer);
+		return path;
+	}
+
+	/**
+	 * Extract text from a PDF. Pages with no text layer (scanned/image-only) are handled as either:
+	 * - Rendered to images and returned alongside the text so the (vision-capable) model can read them directly
+	 * - OCR'd via Tesseract when the doc is too large to reasonably pass as images
+	 */
+	private async resolvePdf(buffer: Buffer): Promise<{text: string, images: {mime: string, data: string}[]}> {
+		const parser = new PDFParse({data: buffer});
+		try {
+			const {text, pages} = await parser.getText();
+			const scanned = (pages || []).filter(p => !p.text?.trim());
+			if(!scanned.length) return {text: text.trim() || '[Empty PDF]', images: []};
+			const total = pages.length;
+			const pageNums = scanned.map(p => p.num);
+			const {pages: shots} = await parser.getScreenshot({partial: pageNums});
+			if(total <= PDF_OCR_PAGE_THRESHOLD) {
+				return {
+					text: text.trim(),
+					images: shots.map(s => ({mime: 'image/png', data: Buffer.from(s.data).toString('base64')}))
+				};
+			}
+			const ocrText = await Promise.all(shots.map(async (s, i) => {
+				const path = await this.writeTemp(`page-${pageNums[i]}.png`, Buffer.from(s.data));
+				try {
+					return await this.ai.vision.ocr(path) || '';
+				} finally {
+					fs.rm(dirname(path), {recursive: true, force: true}).catch(() => {});
+				}
+			}));
+			return {text: [text.trim(), ...ocrText].filter(Boolean).join('\n\n'), images: []};
+		} finally {
+			await parser.destroy();
+		}
+	}
+
+	private async resolveFile(file: LLMFile): Promise<{text?: string, images?: {mime: string, data: string}[]}> {
+		const name = file.name || (file.path ? basename(file.path) : 'file');
+
+		// Already resolved on a previous turn, reuse cached text
+		if(file.extracted) return {text: `<file name="${name}">\n${file.content}\n</file>`};
+
+		const ext = extname(name).slice(1).toLowerCase();
+		const mime = file.mime || '';
+		const isAudio = mime.startsWith('audio/') || LLM.AUDIO_EXT.includes(ext);
+		const isImage = mime.startsWith('image/') || LLM.IMAGE_EXT.includes(ext);
+		const isPdf = mime === 'application/pdf' || LLM.PDF_EXT.includes(ext);
+		const isText = mime.startsWith('text/') || LLM.TEXT_EXT.includes(ext);
+
+		let tmpDir: string | null = null;
+		try {
+			if(isImage) {
+				const data = (await this.loadBuffer(file, false)).toString('base64');
+				return {images: [{mime: mime || `image/${ext === 'jpg' ? 'jpeg' : ext}`, data}]};
+			}
+
+			if(isPdf) {
+				const {text, images} = await this.resolvePdf(await this.loadBuffer(file, false));
+				// Only cache/skip re-processing when we didn't need to hand off images (OCR'd or fully text-based)
+				if(!images.length) {
+					file.content = text;
+					file.extracted = true;
+					delete file.path;
+				}
+				return {text: `<file name="${name}">\n${text || '[Scanned PDF - see attached page images]'}\n</file>`, images};
+			}
+
+			let text: string;
+			if(isAudio) {
+				let path = file.path;
+				if(!path) {
+					const buffer = await this.loadBuffer(file, false);
+					path = await this.writeTemp(name, buffer);
+					tmpDir = dirname(path);
+				}
+				text = await this.ai.audio.asr(path) || '';
+			} else if(isText) {
+				text = (await this.loadBuffer(file, true)).toString('utf-8');
+			} else {
+				text = typeof file.content === 'string' ? file.content : `[Binary file, unable to extract: ${name}]`;
+			}
+			file.content = text;
+			file.extracted = true;
+			delete file.path;
+
+			return {text: `<file name="${name}">\n${text}\n</file>`};
+		} catch(err: any) {
+			return {text: `<file name="${name}">Failed to process: ${err.message}</file>`};
+		} finally {
+			if(tmpDir) fs.rm(tmpDir, {recursive: true, force: true}).catch(() => {});
+		}
+	}
+
+	private async resolveFiles(files: LLMFile[]): Promise<{text: string, images: {mime: string, data: string}[]}> {
+		const resolved = await Promise.all(files.map(f => this.resolveFile(f)));
+		return {
+			text: resolved.filter(r => r.text).map(r => r.text).join('\n\n'),
+			images: resolved.flatMap(r => r.images || [])
+		};
+	}
+
+	private setupAgent(stubs: AgentRef[] = [], history: LLMMessage[], aborts: ((keep?: boolean) => void)[], depth = 0, delegateState: {resp: string | null}): AiTool[] {
+		return stubs.map(stub => {
+			const toolName = `${stub.delegate ? '' : 'sub'}agent_${snakeCase(stub.name)}`;
 			return {
 				name: toolName,
-				description: `${a.delegate ? 'Delegate to ' : ''}Subagent: ${a.description || a.name}`,
+				description: `${stub.delegate ? 'Delegate to ' : ''}Subagent: ${stub.description || stub.name}`,
-				args: {
+				args: clean<any>({
-					context: {type: 'string', description: 'Summary of related messages, samples, files, etc...', required: true},
+					context: !stub.delegate ? {type: 'string', description: 'Summary of related messages, samples, files, etc...', required: true} : undefined,
 					instructions: {type: 'string', description: 'Detailed instructions for subagent to complete', required: true},
-				},
+				}),
 				fn: async (args: any, stream: any, ai: any, id?: string) => {
 					if(depth >= MAX_AGENT_DEPTH) return 'Max agent delegation depth exceeded';
-					const subHistory: LLMMessage[] = [];
 
-					// Opt-in only, self always excluded regardless of whitelist
+					const a = await stub.fn();
-					const nested = (a.agents || [])
+					if(!a) return `Agent "${stub.name}" could not be resolved`;
-						.map(name => allAgents.find(x => x.name === name))
-						.filter((x): x is Agent => !!x && x.name !== a.name);
 
-					const start = Date.now();
+					const q = a.delegate ? '' : `${args.instructions}${args.context ? `\n\n<context>${args.context}</context>` : ''}`;
-					const request = this.ask(`${args.instructions}${args.context ? `\n\n<context>${args.context}</context>` : ''}`, {
+
-						system: `You are a specialized subagent. ${a.delegate ? 'Your output streams directly to the user for the remainder of this turn.' : 'You are wrapped in a tool call that will be analysis by an LLM'}
+					const request = this.ask(q, {
-As a subagent, focus on executing your task completely using available tools and returning only the final result - no commentary, questions, or dialogue.
+						system: `You are a specialized subagent being called from an orchestrator
+${a.delegate ? 'Your output streams directly to the user for the remainder of this turn. You are mid conversation' : 'You are wrapped in a tool call that will be analysis by an LLM'}
+Dispense with greetings and focus on your instructions using available tools and returning only the final result unless specifically instructed to converse
 
 ${a.system}`,
 						model: a.model || undefined,
 						temperature: a.temperature,
 						stream: a.delegate ? stream : undefined,
-						history: subHistory,
+						history: a.delegate ? history : [],
 						mcp: a.mcp || undefined,
 						skills: a.skills || undefined,
 						tools: a.tools || undefined,
-						agents: nested,
+						agents: a.agents || [],
 						_agentDepth: depth + 1,
 					} as any);
 					aborts.push(request.abort);
 					const resp = await request;
-					const duration = Date.now() - start;
-					const assistantTurns = subHistory.filter((h: any) => h.role === 'assistant' && h.duration);
-					const genTime = assistantTurns.reduce((s, h: any) => s + h.duration, 0);
-					const genTokens = assistantTurns.reduce((s, h: any) => s + (h.tps || 0) * (h.duration / 1000), 0);
-					const tps = genTime > 0 ? genTokens / (genTime / 1000) : 0;
 
 					if(a.delegate) {
-						pending.set(<string>id, {resp, subHistory, duration, tps});
+						delegateState.resp = resp;
 						return '';
 					}
 					return resp;
@@ -208,7 +354,7 @@ ${a.system}`,
 
 		const list = allTools.map(t => `- ${t.name}: ${t.description}`).join('\n');
 		return {
-			prompt: `You have access to the following MCP tools:\n${list}`,
+			prompt: `## MCP\nYou have access to the following MCP tools:\n${list}`,
 			tools: allTools
 		};
 	}
@@ -217,7 +363,7 @@ ${a.system}`,
 		if(!skills?.length) return {prompt: '', tools: []};
 		const list = skills.map(s => `- ${s.name}: ${s.description}`).join('\n');
 		return {
-			prompt: `You have access to the following skill documents, use \`read_skill\` to access them:\n${list}`,
+			prompt: `## Skills\nYou have access to the following skill documents, whenever there is overlap between a question and a skill file, use \`skill_read\` to get instructions and background knowledge:\n${list}`,
 			tools: [{
 				name: 'skill_read',
 				description: 'Read the full content of a skill/knowledge document',
@@ -259,17 +405,40 @@ ${a.system}`,
 		if(!this.models[m]) throw new Error(`Model does not exist: ${m}`);
 		let request: AbortablePromise<string> | null = null;
 		let aborted = false;
-		const nestedAborts: (() => void)[] = [];
+		let keepOnAbort = true;
-		const abort = () => {
+		const nestedAborts: ((keep?: boolean) => void)[] = [];
+		const abort = (keep = true) => {
 			aborted = true;
-			request?.abort?.();
+			keepOnAbort = keep;
-			nestedAborts.forEach(a => a());
+			request?.abort?.(keep);
+			nestedAborts.forEach(a => a(keep));
 		};
 
-		const promise = (async () => {
+		let promise: any;
+		const requestStart = Date.now();
+
+		promise = (async () => {
 			let tools: AiTool[] = options.tools || this.ai.options.llm?.tools || [];
 			const prompts: string[] = [];
 			let history = options.history || [];
+			const historyStart = history.length;
+			const files = options.files || [];
+			if(message || files.length) history.push({role: 'user', content: message || '', timestamp: Date.now()});
+
+			// Accumulate streamed text so it can be committed to history if aborted mid-generation
+			let partialText = '';
+			const onStream = options.stream;
+			const stream = (chunk: {text?: string, tool?: string, done?: true}) => {
+				if(chunk.text) partialText += chunk.text;
+				return onStream?.(chunk);
+			};
+
+			/** Commit (keep) or discard this turn's progress on abort, then throw */
+			const abortNow = (): never => {
+				if(keepOnAbort) { if(partialText) history.push({role: 'assistant', content: partialText, timestamp: Date.now()}); }
+				else history.splice(historyStart, history.length - historyStart);
+				throw Object.assign(new Error('Aborted'), {name: 'AbortError'});
+			};
 
 			// MCP
 			const mcp = options.mcp || this.ai.options?.llm?.mcp;
@@ -289,8 +458,8 @@ ${a.system}`,
 
 			// Agents
 			const agents = options.agents || this.ai.options?.llm?.agents;
-			const pendingDelegates = new Map<string, any>();
+			const delegateState: {resp: string | null} = {resp: null};
-			if(agents?.length) tools.push(...this.setupAgent(agents, agents, pendingDelegates, nestedAborts, options._agentDepth || 0));
+			if(agents?.length) tools.push(...this.setupAgent(agents, history, nestedAborts, options._agentDepth || 0, delegateState));
 
 			// Memory
 			const mem = MemoryManager.normalize(options.memory);
@@ -298,8 +467,8 @@ ${a.system}`,
 				const mems = mem.memory instanceof MemoryCache ? mem.memory.memories : mem.memory;
 				if(mems.length) {
 					if(mem.inject) {
-						const pool = 15; // candidates considered, cheap since only refs are listed
+						const pool = 15;
-						const budget = mem.maxTokens ?? 2000; // actual content injected
+						const budget = mem.maxTokens ?? 2000;
 						const relevant = await this.memoryManager.recollect(message, mem.memory, pool);
 
 						let used = 0;
@@ -313,81 +482,92 @@ ${a.system}`,
 							} else listed.push(r);
 						}
 
-						prompts.unshift(`You have access to the following memory files:
+						prompts.unshift(`## Memory
-${mems.map(m => `- ${m.name}: ${m.description}`).join('\n')}
+You have a background memory process which has prefetched relevant information${mem.update ? ' and will create new memories from this conversation' : ''} for you
-${preloaded.length ? `
+Assume it is perfect and never mention this process to anyone ever
-Relevant memories have been preloaded:
+Always use your memories to craft a personalized response, they contain links / [[wiki links]] which you use navigate between them
-${preloaded.map(r => `
+${mem.tool ? `You can access memory files via the \`memory_search\` and \`memory_recall\` tools
-**${r.name}**
+When you need information about the user, \`memory_recall\` \`People/User\` before asking (fetch if not included bellow)
-${r.description}
+When you need information not provided, attempt 1-3 \`memory_search\` calls with distinct queries before asking` : ''}
-${r.content}
+
-`).join('\n---\n')}
+${preloaded.length ? `### Prefetched Memories (Most relevant first):
-` : ''}${listed.length ? `
+
-Also relevant but not preloaded (use \`memory_recall\`): ${listed.map(r => r.name).join(', ')}
+${preloaded.map(r => `Memory: ${r.name}
-` : ''}`.trim());
+Description: ${r.description}
+Linked: ${makeUnique([...r.links, ...r.backlinks]).join(', ')}
+\`\`\`
+${stripHeader(r.content)}
+\`\`\``).join('\n\n')}` : ''}
+${mem.tool && listed.length ? '\n' + listed.map(r => `Memory: ${r.name}
+Description: ${r.description}
+Linked: ${makeUnique([...r.links, ...r.backlinks]).join(', ')}
+<!-- Truncated -->`).join('\n\n') : ''}`.trim())
 					}
-					if(mem.tool) tools.push(this.memoryManager.tools.read(mem.memory));
+					if(mem.tool) tools.push(...this.memoryManager.tools.read(mem.memory));
 				}
 			}
 
-			if(aborted) throw Object.assign(new Error('Aborted'), {name: 'AbortError'});
+			if(aborted) abortNow();
+
+			const lastMsg = history[history.length - 1];
+			if(files.length && lastMsg?.role === 'user') lastMsg.files = files;
+			const restores: {msg: LLMMessage, content: any}[] = [];
+			for(const msg of history) {
+				if(msg.role !== 'user' || !msg.files?.length) continue;
+				const {text, images} = await this.resolveFiles(msg.files);
+				if(!text && !images.length) continue;
+				restores.push({msg, content: msg.content});
+				const merged = text ? [msg.content, text].filter(Boolean).join('\n\n') : msg.content;
+				msg.content = images.length
+					? [...images.map(i => ({type: 'image', mime: i.mime, data: i.data})), {type: 'text', text: merged}]
+					: merged;
+			}
 
-			// Time each tool call's real execution so its history entry gets its own duration/tps
 			const toolTimings = new Map<string, {duration: number, tps: number}>();
 			tools = this.wrapToolTiming(tools, toolTimings);
 
-			prompts.unshift(options.system || this.ai.options.llm?.system || '');
+			if(aborted) abortNow();
-			request = this.models[m].ask(message, {...options, tools, system: prompts.filter(Boolean).join('\n\n')});
-			let resp = await request;
 
-			// Providers stamp duration/tps on assistant entries themselves (from real API usage).
+			prompts.unshift(options.system || this.ai.options.llm?.system || '');
-			// Overwrite tool entries with actual tool-execution timing instead of the LLM call timing.
+			request = this.models[m].ask('', {...options, tools, stream, system: prompts.filter(Boolean).join('\n\n')});
+			let resp: string;
+			try {
+				resp = await request;
+			} catch(err: any) {
+				if(aborted) return abortNow();
+				throw err;
+			}
+
+			// Strip the file injection shim
+			restores.forEach(({msg, content}) => msg.content = content);
+
+			// Capture meta (duration / tps)
 			for(const h of history) {
 				if(h.role === 'tool' && toolTimings.has(h.id)) Object.assign(h, toolTimings.get(h.id));
 			}
 
-			// Spice delegated agents response into history
+			if(typeof resp === 'string' && !resp.trim() && delegateState.resp !== null) resp = delegateState.resp;
-			let lastDelegateResp: string | null = null;
-			if(pendingDelegates.size) {
-				for(let i = 0; i < history.length; i++) {
-					const h: any = history[i];
-					if(h.role !== 'tool' || !pendingDelegates.has(h.id)) continue;
-					const {resp: delegateResp, subHistory, duration, tps} = pendingDelegates.get(h.id)!;
-					pendingDelegates.delete(h.id);
-					const insert: LLMMessage[] = [...subHistory.filter(sh => sh.role === 'tool'), {role: 'assistant', content: delegateResp, timestamp: Date.now(), duration, tps}];
-					history.splice(i + 1, 0, ...insert);
-					lastDelegateResp = delegateResp;
-					i += insert.length;
-				}
-			}
 
-			// If the orchestrator added no commentary of its own, its answer IS the delegate's answer
-			if(typeof resp === 'string' && !resp.trim() && lastDelegateResp !== null) resp = lastDelegateResp;
-
-			// Trim memory injections from history
 			if(mem?.tool) history.splice(0, history.length, ...history.filter(h => h.role !== 'tool' || h.name !== 'memory_recall'));
-
-			// Auto-memorize before compressing
 			if(options.compress && this.estimateTokens(history) >= options.compress.max) {
 				if(mem?.update) await this.memoryManager.memorize(history, mem.memory, {model: options.memoryModel || this.defaultModel, ...options});
 				const compressed = await this.compressHistory(history, options.compress.max, options.compress.min, options);
 				if(options.history) options.history.splice(0, options.history.length, ...compressed);
 			}
 
+			const requestDuration = Date.now() - requestStart;
+			const totalTokens = history
+				.filter((h: any) => h.role === 'assistant' && h.duration && h.tps)
+				.reduce((sum: number, h: any) => sum + h.tps * (h.duration / 1000), 0);
+			const requestTps = requestDuration > 0 ? totalTokens / (requestDuration / 1000) : 0;
+			Object.assign(promise, {duration: requestDuration, tps: requestTps});
+
 			return resp;
 		})();
 
 		return Object.assign(promise, {abort});
 	}
 
-	/**
-	 * Digest full conversation history into memory documents.
-	 * Call on session end to persist the conversation.
-	 */
-	async updateMemory(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest = {}): Promise<Memory[]> {
-		return this.memoryManager.memorize(history, memories, {model: this.defaultModel, ...options});
-	}
-
 	/**
 	 * Compress chat history to reduce context size
 	 * @param {LLMMessage[]} history Chatlog that will be compressed
@@ -416,24 +596,6 @@ Also relevant but not preloaded (use \`memory_recall\`): ${listed.map(r => r.nam
 		return h;
 	}
 
-	/**
-	 * Compare the difference between embeddings (calculates the angle between two vectors)
-	 * @param {number[]} v1 First embedding / vector comparison
-	 * @param {number[]} v2 Second embedding / vector for comparison
-	 * @returns {number} Similarity values 0-1: 0 = unique, 1 = identical
-	 */
-	cosineSimilarity(v1: number[], v2: number[]): number {
-		if (v1.length !== v2.length) throw new Error('Vectors must be same length');
-		let dotProduct = 0, normA = 0, normB = 0;
-		for (let i = 0; i < v1.length; i++) {
-			dotProduct += v1[i] * v2[i];
-			normA += v1[i] * v1[i];
-			normB += v2[i] * v2[i];
-		}
-		const denominator = Math.sqrt(normA) * Math.sqrt(normB);
-		return denominator === 0 ? 0 : dotProduct / denominator;
-	}
-
 	/**
 	 * Chunk text into parts for AI digestion
 	 * @param {object | string} target Item that will be chunked (objects get converted)
@@ -567,6 +729,14 @@ Also relevant but not preloaded (use \`memory_recall\`): ${listed.map(r => r.nam
 		};
 	}
 
+	/**
+	 * Digest full conversation history into memory documents.
+	 * Call on session end to persist the conversation.
+	 */
+	async memorize(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest = {}): Promise<Memory[]> {
+		return this.memoryManager.memorize(history, memories, {model: this.defaultModel, ...options});
+	}
+
 	/**
 	 * Create a summary of some text
 	 * @param {string} text Text to summarize

src/memory.ts

@@ -1,575 +0,0 @@
-import {LLMRequest, LLMMessage} from './llm.ts';
-import {AiTool} from './tools.ts';
-import {KDPoint, KDTree} from './kd-tree.ts';
-
-export class MemoryCache {
-	private tree: KDTree<MemoryRef>;
-	public memories: Memory[];
-
-	get length() { return this.memories.length; }
-
-	constructor(memories: Memory[]) {
-		this.memories = memories;
-		this.tree = this.buildTree();
-	}
-
-	private buildTree(): KDTree<MemoryRef> {
-		const embedded = this.memories.filter(m => m.embedding?.length);
-		if (!embedded.length) return new KDTree<MemoryRef>(0);
-
-		const dims = embedded[0].embedding.length;
-		const points: KDPoint<MemoryRef>[] = embedded.map(m => ({
-			vector: m.embedding,
-			payload: {name: m.name, description: m.description},
-		}));
-
-		return new KDTree<MemoryRef>(dims, 'cosine', points);
-	}
-
-	search(query: number[], limit: number): MemoryRef[] {
-		const results = this.tree.knn(query, limit);
-		return results.map(r => r.point.payload);
-	}
-
-	add(memory: Memory): void {
-		this.memories.push(memory);
-		this.rebuild();
-	}
-
-	update(memory: Memory): void {
-		const idx = this.memories.findIndex(m => m.name === memory.name);
-		if (idx !== -1) {
-			this.memories[idx] = memory;
-		} else {
-			this.memories.push(memory);
-		}
-		this.rebuild();
-	}
-
-	remove(name: string): void {
-		const idx = this.memories.findIndex(m => m.name === name);
-		if (idx !== -1) {
-			this.memories.splice(idx, 1);
-			this.rebuild();
-		}
-	}
-
-	rebuild(): void {
-		this.tree = this.buildTree();
-	}
-}
-
-export type MemoryOptions = {
-	/** Memory object */
-	memory: Memory[] | MemoryCache;
-	/** Inject N memories into the system prompt */
-	inject?: boolean;
-	/** expose recall tool to LLM */
-	tool?: boolean;
-	/** Update memory on compression */
-	update?: boolean;
-	/** Max context size of memories to inject to each call (removed immediately after use) */
-	maxTokens?: number;
-}
-
-export type Memory = {
-	name: string;
-	description: string;
-	content: string;
-	embedding: number[];
-}
-
-export type MemoryRef = {
-	name: string;
-	description: string;
-}
-
-export type FactBucket = {
-	subject: string;
-	facts: string[];
-}
-
-export type MemoryNode = {
-	name: string;
-	missing: boolean;
-	links: string[];
-	backlinks: string[];
-}
-
-function extractLinks(content: string): string[] {
-	if(!content) return [];
-	const matches = content.matchAll(/\[\[([^\]]+)\]\]/g);
-	return [...new Set([...matches].map(m => m[1].trim()))];
-}
-
-export function extractMetadata(content: string): {links: string[], backlinks: string[]} {
-	const match = content.match(/^---\n([\s\S]*?)\n---/);
-	if (!match) return {links: [], backlinks: []};
-
-	const fm = match[1];
-	const getList = (key: string): string[] => {
-		const m = fm.match(new RegExp(`^${key}:\\s*\\[(.*)\\]$`, 'm'));
-		if (!m || !m[1].trim()) return [];
-		return m[1].split(',').map(s => s.trim().replace(/^"|"$/g, '')).filter(Boolean);
-	};
-
-	return {
-		links: getList('links'),
-		backlinks: getList('backlinks'),
-	};
-}
-
-function dedupeFacts(facts: string[]): string[] {
-	const seen = new Map<string, string>();
-	for (const f of facts) {
-		const clean = f.trim();
-		if (clean) seen.set(clean.toLowerCase(), clean);
-	}
-	return [...seen.values()];
-}
-
-function cosineDistance(a: number[], b: number[]): number {
-	let dot = 0, normA = 0, normB = 0;
-	for (let i = 0; i < a.length; i++) {
-		dot   += a[i] * b[i];
-		normA += a[i] * a[i];
-		normB += b[i] * b[i];
-	}
-	const denom = Math.sqrt(normA) * Math.sqrt(normB);
-	return denom === 0 ? 1 : 1 - dot / denom;
-}
-
-function getWeekMonday(date: Date = new Date()): string {
-	const d = new Date(Date.UTC(date.getFullYear(), date.getMonth(), date.getDate()));
-	const day = d.getUTCDay();
-	const diff = day === 0 ? -6 : 1 - day;
-	d.setUTCDate(d.getUTCDate() + diff);
-	return d.toISOString().slice(0, 10);
-}
-
-function getWeekSunday(monday: string): string {
-	const d = new Date(`${monday}T00:00:00Z`);
-	d.setUTCDate(d.getUTCDate() + 6);
-	return d.toISOString().slice(0, 10);
-}
-
-export class MemoryManager {
-	private recentlyTouched = new Map<string, number>();
-
-	private pendingMemorizations = new Map<string, {
-		memories: Memory[] | MemoryCache,
-		tempMemoryName: string,
-		timestamp: number,
-	}>();
-
-	private queues = new Map<string, {
-		pending: string[],
-		request: {abort?: () => void} | null,
-		task: Promise<void>,
-	}>();
-
-	tools = {
-		read: (memories: Memory[] | MemoryCache): AiTool => ({
-			name: 'memory_recall',
-			description: 'Read the full content of a memory document',
-			args: {
-				name: {type: 'string', description: 'Exact memory name', required: true},
-			},
-			fn: (args: any) => {
-				const mems = memories instanceof MemoryCache ? memories.memories : memories;
-				const mem = mems.find(m => m.name === args.name);
-				if (!mem) return 'Document not found';
-				this.touch(mem.name);
-				return mem.content;
-			},
-		}),
-
-		forget: (memories: Memory[] | MemoryCache): AiTool => ({
-			name: 'memory_forget',
-			description: 'Permanently delete a memory document and clean up all references to it',
-			args: {
-				name: {type: 'string', description: 'Exact memory name to forget', required: true}
-			},
-			fn: (args: any) => {
-				const result = this.forget(args.name, memories);
-				return result ? `Forgotten: ${args.name}` : `Not found: ${args.name}`;
-			},
-		}),
-	};
-
-	constructor(private llm: any) {}
-
-	static normalize(m?: Memory[] | MemoryCache | MemoryOptions) {
-		if(!m) return null;
-		const raw = m instanceof MemoryCache || Array.isArray(m);
-		return raw ? {memory: <Memory[] | MemoryCache>m, inject: true, tool: true, update: true} : {inject: true, tool: true, update: true, ...m};
-	}
-
-	private async createTempMemory(conversation: string): Promise<Memory> {
-		const timestamp = Date.now();
-		const content = `---
-name: _temp_${timestamp}
-description: Temporary memory - processing in background
-tags: [_temporary]
-links: []
-backlinks: []
-modified: ${new Date().toISOString()}
----
-
-# Recent Conversation (Processing)
-
-${conversation}`;
-		const [e] = await this.llm.embedding(content);
-		return {
-			name: `_temp_${timestamp}`,
-			description: 'Temporary memory - processing in background',
-			content,
-			embedding: e?.embedding || [],
-		};
-	}
-
-	private applyHeader(content: string, header: string): string {
-		return `${header}\n\n${this.stripHeader(content)}`;
-	}
-
-	private async backgroundMemorization(conversation: string, memories: Memory[] | MemoryCache, options: LLMRequest, tempName: string): Promise<void> {
-		const mem = memories instanceof MemoryCache ? memories.memories : memories;
-		const monday = getWeekMonday();
-		const sunday = getWeekSunday(monday);
-		const buckets = await this.factAgent(conversation, mem, options, monday);
-		if(!buckets.length) return;
-		const jobs = [...buckets].map(({subject, facts}) => {
-			let node = mem.find(m => m.name === subject);
-			if(!node) {
-				node = {name: subject, description: '', content: '', embedding: [],};
-				mem.push(node);
-			}
-			const week = subject.startsWith('Journal/') ? {monday, sunday} : undefined;
-			return this.enqueue(node, facts, mem, options, tempName, week);
-		});
-		await Promise.all(jobs);
-	}
-
-	private buildHeader(node: Memory, week?: {monday: string, sunday: string}, links: string[] = [], backlinks: string[] = []): string {
-		const tags = node.name.split('/')[0]?.toLowerCase();
-		const lines = [
-			'---',
-			`name: ${node.name}`,
-			`description: ${node.description || ''}`,
-			tags ? `tags: [${tags}]` : '',
-			links.length ? `links: [${links.map(l => `"${l}"`).join(', ')}]` : 'links: []',
-			backlinks.length ? `backlinks: [${backlinks.map(l => `"${l}"`).join(', ')}]` : 'backlinks: []',
-			week ? `week: ${week.monday} – ${week.sunday}` : '',
-			`modified: ${new Date().toISOString()}`,
-			'---',
-		].filter(Boolean);
-		return lines.join('\n');
-	}
-
-	private cosineSearch(query: number[], memories: Memory[], limit: number): MemoryRef[] {
-		const scored = memories
-			.filter(m => m.embedding?.length)
-			.map(m => ({
-				ref: {name: m.name, description: m.description},
-				distance: cosineDistance(query, m.embedding),
-			}))
-			.sort((a, b) => a.distance - b.distance)
-			.slice(0, limit);
-		return scored.map(s => s.ref);
-	}
-
-	/**
-	 * Coalescing queue: if a doc is already compiling, abort the in-flight run, merge its
-	 * facts with the new ones and restart. Never blocks a pending update, never drops facts.
-	 */
-	private enqueue(node: Memory, facts: string[], memories: Memory[] | MemoryCache, options: LLMRequest, tempName: string, week?: {monday: string, sunday: string}): Promise<void> {
-		const key = node.name;
-		const existing = this.queues.get(key);
-		if (existing) {
-			existing.pending.push(...facts);
-			existing.request?.abort?.();
-			return existing.task;
-		}
-
-		const entry: {pending: string[], request: {abort?: () => void} | null, task: Promise<void>} = {pending: [...facts], request: null, task: Promise.resolve()};
-		this.queues.set(key, entry);
-		const m = memories instanceof MemoryCache ? memories.memories : memories;
-		entry.task = (async () => {
-			while (entry.pending.length) {
-				const batch = dedupeFacts(entry.pending.splice(0, entry.pending.length));
-				const written = await this.docAgent(node, batch, m, options, tempName, week, entry);
-				if (!written) entry.pending.unshift(...batch);
-			}
-		})().finally(() => {
-			this.queues.delete(key);
-			if(!this.queues.size && memories instanceof MemoryCache) memories.rebuild();
-		});
-		return entry.task;
-	}
-
-	private listNodes(memories: Memory[]): MemoryRef[] {
-		return memories.map(m => ({name: m.name, description: m.description}));
-	}
-
-	decay() {
-		for(const [name, ttl] of this.recentlyTouched) {
-			if(ttl <= 1) this.recentlyTouched.delete(name);
-			else this.recentlyTouched.set(name, ttl - 1);
-		}
-	}
-
-	forget(name: string, memories: Memory[] | MemoryCache): boolean {
-		const mem = memories instanceof MemoryCache ? memories.memories : memories;
-		const idx = mem.findIndex(m => m.name === name);
-		if (idx === -1) return false;
-
-		for (const node of mem) {
-			const {links, backlinks} = extractMetadata(node.content);
-			const newBacklinks = backlinks.filter(b => b !== name);
-			const newLinks = links.filter(l => l !== name);
-
-			if (newBacklinks.length !== backlinks.length || newLinks.length !== links.length) {
-				node.content = this.updateFrontmatter(node.content, {
-					links: newLinks,
-					backlinks: newBacklinks,
-				});
-			}
-		}
-
-		mem.splice(idx, 1);
-
-		if (memories instanceof MemoryCache) memories.rebuild();
-		return true;
-	}
-
-	getTouched(): string[] {
-		return [...this.recentlyTouched.keys()];
-	}
-
-	async memorize(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest): Promise<Memory[]> {
-		const conversation = history
-			.filter(h => h.role === 'user' || h.role === 'assistant')
-			.map(h => `[${h.role}]: ${h.content}`).join('\n\n').trim();
-		if(!conversation) return [];
-
-		const trackingId = `${Date.now()}_${Math.random()}`;
-		const tempMemory = await this.createTempMemory(conversation);
-		const mem = memories instanceof MemoryCache ? memories.memories : memories;
-		mem.push(tempMemory);
-		if (memories instanceof MemoryCache) memories.rebuild();
-		this.pendingMemorizations.set(trackingId, {
-			memories,
-			tempMemoryName: tempMemory.name,
-			timestamp: Date.now(),
-		});
-
-		try {
-			await this.backgroundMemorization(conversation, memories, options, tempMemory.name);
-			const finalMem = memories instanceof MemoryCache ? memories.memories : memories;
-			return finalMem.filter(m => !m.name.startsWith('_temp_'));
-		} finally {
-			const pending = this.pendingMemorizations.get(trackingId);
-			if (pending) {
-				const cleanMem = pending.memories instanceof MemoryCache
-					? pending.memories.memories
-					: pending.memories;
-				const idx = cleanMem.findIndex(m => m.name === pending.tempMemoryName);
-				if (idx !== -1) cleanMem.splice(idx, 1);
-				if (pending.memories instanceof MemoryCache) pending.memories.rebuild();
-			}
-			this.pendingMemorizations.delete(trackingId);
-		}
-	}
-
-	async recollect(query: string, memories: Memory[] | MemoryCache, limit = 5, graphDepth = 1): Promise<Memory[]> {
-		const mem: Memory[] = memories instanceof MemoryCache ? memories.memories : memories;
-		if (!mem.length) return [];
-
-		const [e] = await this.llm.embedding(query);
-		if (!e) return [];
-
-		let vectorResults: MemoryRef[];
-		if (memories instanceof MemoryCache) vectorResults = memories.search(e.embedding, limit);
-		else vectorResults = this.cosineSearch(e.embedding, mem, limit);
-		const found = new Set<string>(vectorResults.map(r => r.name));
-
-		if (graphDepth > 0) {
-			const frontier = [...found];
-			for (let depth = 0; depth < graphDepth; depth++) {
-				const next: string[] = [];
-				for (const name of frontier) {
-					const node = mem.find(m => m.name === name);
-					if (!node) continue;
-					const {links} = extractMetadata(node.content);
-					for (const link of links) {
-						if (!found.has(link) && mem.find(m => m.name === link)) {
-							found.add(link);
-							next.push(link);
-						}
-					}
-				}
-				frontier.splice(0, frontier.length, ...next);
-				if (!frontier.length) break;
-			}
-		}
-
-		const vectorOrder = vectorResults.map(r => r.name);
-		const graphExpansions = [...found].filter(n => !vectorOrder.includes(n));
-		const ordered = [...vectorOrder, ...graphExpansions];
-		return ordered.map(n => mem.find(m => m.name === n)!).filter(Boolean);
-	}
-
-	touch(name: string, ttl = 2) {
-		this.recentlyTouched.set(name, ttl);
-	}
-
-	private updateFrontmatter(content: string, updates: {links?: string[], backlinks?: string[]}): string {
-		const match = content.match(/^---\n([\s\S]*?)\n---\n\n?([\s\S]*)$/);
-		if (!match) return content;
-
-		const [, fm, body] = match;
-		let newFm = fm;
-
-		if (updates.links !== undefined) {
-			const linksList = updates.links.length ? `[${updates.links.map(l => `"${l}"`).join(', ')}]` : '[]';
-			newFm = newFm.replace(/^links:.*$/m, `links: ${linksList}`);
-		}
-
-		if (updates.backlinks !== undefined) {
-			const backlinksList = updates.backlinks.length ? `[${updates.backlinks.map(l => `"${l}"`).join(', ')}]` : '[]';
-			newFm = newFm.replace(/^backlinks:.*$/m, `backlinks: ${backlinksList}`);
-		}
-
-		newFm = newFm.replace(/^modified:.*$/m, `modified: ${new Date().toISOString()}`);
-
-		return `---\n${newFm}\n---\n\n${body}`;
-	}
-
-	private stripHeader(content: string): string {
-		return content.replace(/^---[\s\S]*?\n---\n?/, '').trimStart();
-	}
-
-	private async docAgent(node: Memory, facts: string[], memories: Memory[], options: LLMRequest, tempName: string, week: {monday: string, sunday: string} | undefined, entry: {request: {abort?: () => void} | null}): Promise<boolean> {
-		const {links: oldLinks} = extractMetadata(node.content);
-		const currentBody = this.stripHeader(node.content);
-		let update;
-		try {
-			for(let i = 0; i < 3 && !update?.content; i++) {
-				const request = this.llm.ask(`New Facts:\n${facts.map(f => `- ${f}`).join('\n')}`, {
-					model: options.model,
-					temperature: 0.3,
-					schema: {
-						description: {type: 'string', description: 'One-line description of what this document covers, no formatting or emojis', required: true},
-						content: {type: 'string', description: 'Rewritten document in markdown, without the frontmatter block', required: true},
-					},
-					system: `You are a knowledge base editor. Rewrite the current document below so it incorporates the new facts.
-
-Formatting rules:
-- Use Obsidian-style markdown: # headings, **bold** to add emphasis, __italics__ for titles, terms, etc, bullet & numbered lists for grouped 1D data and tables for 2D data
-- Link related concepts with [[WikiLink]] notation using full paths like [[People/Sarah]] or [[Projects/Website]]
-- Create links for specific entities (person, place, project, program) and abstract concepts (quantum mechanics, entropy) but skip generics (car, red, dog)
-- Keep the document concise, factual, and human-readable
-- Resolve contradictions: the new facts always win — delete the outdated statement entirely, never keep both
-- Later facts in the list override earlier ones
-- Do not add frontmatter blocks, filler, preamble, or AI commentary
-${week ? '- This is a weekly journal entry.\n' : ''}
-All nodes:
-${this.listNodes(memories).map(n => n.name).join(', ') || 'none'}
-
-Current document:
-\`\`\`markdown
-${currentBody}
-\`\`\``}
-				);
-				entry.request = request;
-				update = await request;
-			}
-		} catch (err: any) {
-			if (err?.name === 'AbortError') return false;
-			throw err;
-		} finally {
-			entry.request = null;
-		}
-
-		if(!update?.content) return false;
-		const newLinks = extractLinks(update.content).filter(l => l !== node.name && l !== tempName);
-		const newLinkSet = new Set(newLinks);
-		const oldLinkSet = new Set(oldLinks);
-
-		for (const added of newLinkSet) {
-			if (!oldLinkSet.has(added)) {
-				const target = memories.find(m => m.name === added);
-				if (target) {
-					const {backlinks} = extractMetadata(target.content);
-					if (!backlinks.includes(node.name)) {
-						target.content = this.updateFrontmatter(target.content, {
-							backlinks: [...backlinks, node.name],
-						});
-					}
-				}
-			}
-		}
-		for (const removed of oldLinkSet) {
-			if (!newLinkSet.has(removed)) {
-				const target = memories.find(m => m.name === removed);
-				if (target) {
-					const {backlinks} = extractMetadata(target.content);
-					target.content = this.updateFrontmatter(target.content, {
-						backlinks: backlinks.filter(b => b !== node.name),
-					});
-				}
-			}
-		}
-
-		const {backlinks} = extractMetadata(node.content);
-		node.description = node.name !== 'Person/User' ? update.description : 'All information about the current user';
-		node.content = this.applyHeader(update.content, this.buildHeader(node, week, newLinks, backlinks));
-		const [e] = await this.llm.embedding(node.content);
-		if(e) node.embedding = e.embedding;
-		return true;
-	}
-
-	private async factAgent(conversation: string, memories: Memory[], options: LLMRequest, weekKey: string): Promise<FactBucket[]> {
-		const buckets = new Map<string, string[]>();
-		await this.llm.ask(conversation, {
-			model: options.model,
-			temperature: 0.2,
-			system: `You are a fact extractor. Analyze this conversation and extract facts worth remembering long-term.
-
-Rules:
-- ONLY extract current facts the USER explicitly stated about themselves, their work, or their projects
-- ONLY extract decisions that were MADE during this conversation
-- DO NOT extract anything the AI said, its capabilities, or meta-conversation about the AI
-- DO NOT extract greetings, pleasantries, or generic exchanges
-- DO NOT extract deltas or changes in facts; ONLY the end fact
-- If nothing worth remembering was said, do not call any tools
-
-When extracting facts, you MUST also decide the exact destination path:
-- Use an existing node name if the facts clearly belong there
-- All information primary about the user should go under "People/User"
-- When required, create a new path following collection/subject format (e.g., People/Sarah, Projects/Oxide)
-- For journal entries, use "Journal"
-
-Available nodes:
-- Journal
-${this.listNodes(memories).filter(n => !n.name.includes('_temp_') && !n.name.includes('Journal')).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}`,
-			tools: [{
-				name: 'facts_extract',
-				description: 'Submit facts with their destination',
-				args: {
-					destination: {type: 'string', description: 'Exact existing node name OR new path (e.g. "People/Sarah", "Projects/Oxide")', required: true},
-					facts: {type: 'string', description: 'Comma-separated facts', required: true},
-				},
-				fn: (args: any) => {
-					const subject = args.destination.trim().toLowerCase() === 'journal'
-						? `Journal/${weekKey}` : args.destination.trim();
-					const facts = buckets.get(subject) ?? [];
-					facts.push(...dedupeFacts(String(args.facts).split(',')));
-					buckets.set(subject, facts);
-					return 'Recorded';
-				},
-			}],
-		});
-		return buckets.entries().toArray().map(([subject, facts]) => ({subject, facts}));
-	}
-}

src/memory/graph.ts

@@ -0,0 +1,128 @@
+import {MemoryCache} from './memory-state.ts';
+import type {Memory} from './memory.ts';
+
+export type MemoryNode = {
+	name: string;
+	missing: boolean;
+	links: string[];
+	backlinks: string[];
+}
+
+export function extractLinks(content: string): string[] {
+	if (!content) return [];
+	const matches = content.matchAll(/\[\[([^\]|]+)(?:\|[^\]]*)?\]\]/g);
+	return [...new Set([...matches].map(m => m[1].trim()))];
+}
+
+export function patchGraph(mems: Memory[], nodes: MemoryNode[], changed: Memory[]): MemoryNode[] {
+	const nameSet = new Set(mems.map(m => m.name));
+	const byName = new Map(nodes.map(n => [n.name, n]));
+
+	const ensureNode = (name: string): MemoryNode => {
+		let n = byName.get(name);
+		if (!n) {
+			n = {name, missing: !nameSet.has(name), links: [], backlinks: []};
+			byName.set(name, n);
+		}
+		return n;
+	};
+
+	for (const m of changed) {
+		const node = ensureNode(m.name);
+		node.missing = false; // real memory, promotes any pre-existing ghost entry
+		const oldLinks = m.links ?? [];
+		const newLinks = extractLinks(m.content).filter(l => l !== m.name);
+
+		for (const target of oldLinks.filter(l => !newLinks.includes(l))) {
+			const t = byName.get(target);
+			if (!t) continue;
+			t.backlinks = t.backlinks.filter(n => n !== m.name);
+			if (t.missing && !t.backlinks.length) byName.delete(target); // fully dereferenced ghost
+		}
+		for (const target of newLinks.filter(l => !oldLinks.includes(l))) {
+			const t = ensureNode(target);
+			if (!t.backlinks.includes(m.name)) t.backlinks.push(m.name);
+		}
+
+		m.links = newLinks;
+		node.links = newLinks;
+	}
+
+	for (const m of mems) {
+		const n = byName.get(m.name);
+		if (n) m.backlinks = n.backlinks;
+	}
+
+	return [...byName.values()];
+}
+
+export function rebuildGraph(memories: Memory[] | MemoryCache): MemoryNode[] {
+	const mems = memories instanceof MemoryCache ? memories.memories : memories;
+	const nameSet = new Set(mems.map(m => m.name));
+
+	for (const m of mems) m.links = extractLinks(m.content).filter(l => l !== m.name);
+	for (const m of mems) m.backlinks = [];
+	for (const m of mems) {
+		for (const link of m.links) {
+			const target = mems.find(t => t.name === link);
+			if (target) target.backlinks.push(m.name);
+		}
+	}
+
+	const nodes: MemoryNode[] = mems.map(m => ({
+		name: m.name,
+		missing: false,
+		links: m.links,
+		backlinks: m.backlinks,
+	}));
+
+	const ghosts = new Set<string>();
+	for (const node of nodes) {
+		for (const link of node.links) {
+			if (!nameSet.has(link)) ghosts.add(link);
+		}
+	}
+
+	return [
+		...nodes,
+		...[...ghosts].map(name => ({
+			name,
+			missing: true,
+			links: [],
+			backlinks: nodes.filter(n => n.links.includes(name)).map(n => n.name),
+		})),
+	];
+}
+
+export function renderMemoryGraph(nodes: MemoryNode[]): string {
+	if (!nodes.length) return 'No memories yet.';
+
+	const groups = new Map<string, (MemoryNode & {label: string})[]>();
+	for (const node of nodes) {
+		const [prefix, ...rest] = node.name.split('/');
+		const group = rest.length ? prefix : 'Root';
+		const label = rest.length ? rest.join('/') : node.name;
+		if (!groups.has(group)) groups.set(group, []);
+		groups.get(group)!.push({...node, label});
+	}
+
+	const ghostCount = nodes.filter(n => n.missing).length;
+	const lines = [`Memory Graph (${nodes.length} nodes, ${ghostCount} ghost${ghostCount === 1 ? '' : 's'})`, ''];
+
+	for (const group of [...groups.keys()].sort()) {
+		const items = groups.get(group)!.sort((a, b) => a.label.localeCompare(b.label));
+		lines.push(`${group}/`);
+		items.forEach((n, i) => {
+			const last = i === items.length - 1;
+			const branch = last ? '└─' : '├─';
+			const pad = last ? '  ' : '│ ';
+			const tag = n.missing ? ' (ghost)' : '';
+			lines.push(`  ${branch} ${n.label}${tag}`);
+			if (n.links.length) lines.push(`  ${pad}   → ${n.links.join(', ')}`);
+			if (n.backlinks.length) lines.push(`  ${pad}   ← ${n.backlinks.join(', ')}`);
+		});
+		lines.push('');
+	}
+
+	return lines.join('\n').trimEnd();
+}

src/memory/kd-tree.ts

@@ -1,3 +1,5 @@
+import {cosineDistance, euclideanDistance} from '../utils.ts';
+
 export type DistanceMetric = "euclidean" | "cosine";
 
 export interface KDPoint<T = unknown> {
@@ -15,28 +17,7 @@ interface KDNode<T> {
 	axis: number;
 	left: KDNode<T> | null;
 	right: KDNode<T> | null;
-}
+	deleted?: boolean;
-
-// ─── Distance helpers ─────────────────────────────────────────────────────────
-
-function euclidean(a: number[], b: number[]): number {
-	let sum = 0;
-	for (let i = 0; i < a.length; i++) {
-		const d = a[i] - b[i];
-		sum += d * d;
-	}
-	return Math.sqrt(sum);
-}
-
-function cosine(a: number[], b: number[]): number {
-	let dot = 0, normA = 0, normB = 0;
-	for (let i = 0; i < a.length; i++) {
-		dot   += a[i] * b[i];
-		normA += a[i] * a[i];
-		normB += b[i] * b[i];
-	}
-	const denom = Math.sqrt(normA) * Math.sqrt(normB);
-	return denom === 0 ? 1 : 1 - dot / denom; // distance = 1 - similarity
 }
 
 /**
@@ -95,6 +76,7 @@ class BoundedMaxHeap<T> {
  *
  * Supports:
  *  - Insertion of labeled points
+ *  - Lazy (tombstone) removal, physically purged on rebalance()
  *  - k-nearest-neighbor (KNN) search
  *  - Radius search (all points within a given distance)
  *  - Euclidean and cosine distance metrics
@@ -103,9 +85,11 @@ class BoundedMaxHeap<T> {
 export class KDTree<T = unknown> {
 	private root: KDNode<T> | null = null;
 	private _size = 0;
-	private readonly dims: number;
+	private _tombstones = 0;
 	private readonly distanceFn: (a: number[], b: number[]) => number;
 
+	readonly dims: number;
+
 	/**
 	 * @param dims           Dimensionality of all vectors (must be consistent).
 	 * @param metric         Distance metric to use. Default: "euclidean".
@@ -119,7 +103,7 @@ export class KDTree<T = unknown> {
 		points?: KDPoint<T>[]
 	) {
 		this.dims = dims;
-		this.distanceFn = metric === "cosine" ? cosine : euclidean;
+		this.distanceFn = metric === "cosine" ? cosineDistance : euclideanDistance;
 
 		if (points && points.length > 0) {
 			this.validateAll(points);
@@ -128,9 +112,15 @@ export class KDTree<T = unknown> {
 		}
 	}
 
-	/** Total number of points stored in the tree. */
+	/** Total number of live points stored in the tree (excludes tombstoned). */
 	get size(): number { return this._size; }
 
+	/** Fraction of physical nodes that are tombstoned (pending removal on next rebalance). */
+	get tombstoneRatio(): number {
+		const total = this._size + this._tombstones;
+		return total ? this._tombstones / total : 0;
+	}
+
 	// ── Insertion ──────────────────────────────────────────────────────────────
 
 	/**
@@ -143,10 +133,36 @@ export class KDTree<T = unknown> {
 		this._size++;
 	}
 
+	// ── Removal ────────────────────────────────────────────────────────────────
+
+	/**
+	 * Lazily remove all live points whose payload matches `predicate`.
+	 * O(n) traversal, but avoids a full tree rebuild. Call `rebalance()`
+	 * periodically (e.g. once tombstoneRatio crosses ~0.25) to reclaim space
+	 * and restore optimal query depth.
+	 * @returns number of points removed
+	 */
+	remove(predicate: (payload: T) => boolean): number {
+		let removed = 0;
+		const visit = (node: KDNode<T> | null): void => {
+			if (!node) return;
+			if (!node.deleted && predicate(node.point.payload)) {
+				node.deleted = true;
+				removed++;
+			}
+			visit(node.left);
+			visit(node.right);
+		};
+		visit(this.root);
+		this._size -= removed;
+		this._tombstones += removed;
+		return removed;
+	}
+
 	// ── KNN search ─────────────────────────────────────────────────────────────
 
 	/**
-	 * Find the k nearest neighbors to `query`.
+	 * Find the k nearest live neighbors to `query`.
 	 * Returns results sorted by distance ascending.
 	 */
 	knn(query: number[], k: number): KNNResult<T>[] {
@@ -170,7 +186,7 @@ export class KDTree<T = unknown> {
 	// ── Radius search ──────────────────────────────────────────────────────────
 
 	/**
-	 * Return all points whose distance to `query` is ≤ `radius`,
+	 * Return all live points whose distance to `query` is ≤ `radius`,
 	 * sorted by distance ascending.
 	 */
 	radiusSearch(query: number[], radius: number): KNNResult<T>[] {
@@ -185,7 +201,7 @@ export class KDTree<T = unknown> {
 
 	// ── Conversion ─────────────────────────────────────────────────────────────
 
-	/** Collect all points in the tree (order not guaranteed). */
+	/** Collect all live points in the tree (order not guaranteed). */
 	toArray(): KDPoint<T>[] {
 		const out: KDPoint<T>[] = [];
 		this.collect(this.root, out);
@@ -193,12 +209,14 @@ export class KDTree<T = unknown> {
 	}
 
 	/**
-	 * Rebuild the tree from its current points as a balanced tree.
+	 * Rebuild the tree from its current live points as a balanced tree.
-	 * Useful after many individual insertions to restore O(log n) query time.
+	 * Physically purges tombstones and restores O(log n) query time.
 	 */
 	rebalance(): void {
 		const points = this.toArray();
 		this.root = points.length ? this.buildBalanced(points, 0) : null;
+		this._size = points.length;
+		this._tombstones = 0;
 	}
 
 	// ── Private: build ─────────────────────────────────────────────────────────
@@ -250,8 +268,10 @@ export class KDTree<T = unknown> {
 	): void {
 		if (node === null) return;
 
-		const dist = this.distanceFn(query, node.point.vector);
+		if (!node.deleted) {
-		heap.push({ point: node.point, distance: dist });
+			const dist = this.distanceFn(query, node.point.vector);
+			heap.push({ point: node.point, distance: dist });
+		}
 
 		const axis = node.axis;
 		const diff = query[axis] - node.point.vector[axis];
@@ -260,11 +280,8 @@ export class KDTree<T = unknown> {
 			: [node.right, node.left];
 
 		this.searchKNN(near, query, k, heap, depth + 1);
-
-		// Only explore the far side if it could contain a closer point.
-		// For cosine distance we can't prune by axis gap alone, so always explore.
 		const shouldExplore =
-			this.distanceFn === cosine
+			this.distanceFn === cosineDistance
 				? true
 				: Math.abs(diff) < heap.worstDistance;
 
@@ -284,9 +301,11 @@ export class KDTree<T = unknown> {
 	): void {
 		if (node === null) return;
 
-		const dist = this.distanceFn(query, node.point.vector);
+		if (!node.deleted) {
-		if (dist <= radius) {
+			const dist = this.distanceFn(query, node.point.vector);
-			results.push({ point: node.point, distance: dist });
+			if (dist <= radius) {
+				results.push({ point: node.point, distance: dist });
+			}
 		}
 
 		const axis = node.axis;
@@ -298,7 +317,7 @@ export class KDTree<T = unknown> {
 		this.searchRadius(near, query, radius, results, depth + 1);
 
 		const shouldExplore =
-			this.distanceFn === cosine ? true : Math.abs(diff) <= radius;
+			this.distanceFn === cosineDistance ? true : Math.abs(diff) <= radius;
 
 		if (shouldExplore) {
 			this.searchRadius(far, query, radius, results, depth + 1);
@@ -309,7 +328,7 @@ export class KDTree<T = unknown> {
 
 	private collect(node: KDNode<T> | null, out: KDPoint<T>[]): void {
 		if (node === null) return;
-		out.push(node.point);
+		if (!node.deleted) out.push(node.point);
 		this.collect(node.left, out);
 		this.collect(node.right, out);
 	}

src/memory/memory-state.ts

@@ -0,0 +1,181 @@
+import {MemoryNode, patchGraph, rebuildGraph} from './graph.ts';
+import {KDTree} from './kd-tree.ts';
+import type {Memory, MemoryRef, MemoryStore} from './memory.ts';
+import {cosineDistance, embedMemoryFields} from '../utils.ts';
+
+const TREE_TOMBSTONE_LIMIT = 0.25;
+
+export function memoryStore(memories: MemoryStore): {
+	list: Memory[];
+	cache: MemoryCache | null;
+	find: (name: string) => Memory | undefined;
+	ghosts: () => string[];
+	search: (vector: number[], limit: number) => MemoryRef[];
+	forget: (name: string) => boolean;
+	rebuild: (changed?: Memory[]) => MemoryNode[];
+	backfillEmbeddings: (llm: any) => Promise<number>;
+} {
+	if(memories instanceof MemoryCache) {
+		return {
+			list: memories.memories,
+			cache: memories,
+			find: name => memories.find(name),
+			ghosts: () => memories.ghosts(),
+			search: (vector, limit) => memories.search(vector, limit),
+			forget: name => memories.remove(name),
+			rebuild: changed => memories.rebuild(changed),
+			backfillEmbeddings: llm => memories.backfillEmbeddings(llm),
+		};
+	}
+
+	return {
+		list: memories,
+		cache: null,
+		find: name => memories.find(m => m.name === name),
+		ghosts: () => rebuildGraph(memories).filter(n => n.missing).map(n => n.name),
+		search: (vector, limit) => memories
+			.filter(m => m.embedding?.length)
+			.map(m => ({
+				name: m.name,
+				description: m.description,
+				distance: cosineDistance(vector, m.embedding),
+			}))
+			.sort((a, b) => a.distance - b.distance)
+			.slice(0, limit),
+		forget: name => {
+			const idx = memories.findIndex(m => m.name === name);
+			if(idx === -1) return false;
+
+			memories.splice(idx, 1);
+			return true;
+		},
+		rebuild: changed => rebuildGraph(memories),
+		backfillEmbeddings: async llm => {
+			const missing = memories.filter(m => !m.embedding?.length);
+			if(!missing.length) return 0;
+
+			await Promise.all(missing.map(async node => {
+				await embedMemoryFields(node, llm);
+			}));
+
+			return missing.length;
+		},
+	};
+}
+
+export class MemoryCache {
+	private tree!: KDTree<MemoryRef>;
+	private indexed = new Map<string, number[]>();
+	public memories: Memory[];
+	public nodes: MemoryNode[] = [];
+
+	get length() {
+		return this.memories.length;
+	}
+
+	constructor(memories: Memory[]) {
+		this.memories = memories;
+		this.tree = new KDTree<MemoryRef>(0);
+		this.rebuild();
+	}
+
+	find(name: string): Memory | undefined {
+		return this.memories.find(m => m.name === name);
+	}
+
+	private syncTree(): void {
+		const current = new Set(this.memories.map(m => m.name));
+
+		for(const [name, emb] of [...this.indexed]) {
+			const mem = this.memories.find(m => m.name === name);
+
+			if(!mem || !current.has(name) || mem.embedding !== emb) {
+				this.tree.remove(p => p.name === name);
+				this.indexed.delete(name);
+			}
+		}
+
+		for(const mem of this.memories) {
+			if(!mem.embedding?.length || this.indexed.has(mem.name)) continue;
+
+			if(this.tree.dims === 0) {
+				this.tree = new KDTree<MemoryRef>(mem.embedding.length, 'cosine');
+			}
+
+			if(mem.embedding.length !== this.tree.dims) continue;
+
+			this.tree.insert({
+				vector: mem.embedding,
+				payload: {
+					name: mem.name,
+					description: mem.description,
+				},
+			});
+
+			this.indexed.set(mem.name, mem.embedding);
+		}
+
+		if(this.tree.tombstoneRatio > TREE_TOMBSTONE_LIMIT) {
+			this.tree.rebalance();
+		}
+	}
+
+	search(query: number[], limit: number): MemoryRef[] {
+		if(!this.tree || this.tree.dims === 0) return [];
+
+		return this.tree.knn(query, limit).map(r => ({
+			...r.point.payload,
+			distance: r.distance,
+		}));
+	}
+
+	add(memory: Memory): void {
+		this.memories.push(memory);
+		this.rebuild([memory]);
+	}
+
+	update(memory: Memory): void {
+		const existing = this.find(memory.name);
+
+		if(existing) Object.assign(existing, memory);
+		else this.memories.push(memory);
+
+		this.rebuild([existing ?? memory]);
+	}
+
+	remove(name: string): boolean {
+		const idx = this.memories.findIndex(m => m.name === name);
+		if(idx === -1) return false;
+
+		this.memories.splice(idx, 1);
+		this.rebuild();
+		return true;
+	}
+
+	ghosts(): string[] {
+		return this.nodes.filter(n => n.missing).map(n => n.name);
+	}
+
+	rebuild(changed?: Memory[]): MemoryNode[] {
+		this.nodes = changed?.length && this.nodes.length
+			? patchGraph(this.memories, this.nodes, changed)
+			: rebuildGraph(this.memories);
+
+		this.syncTree();
+		return this.nodes;
+	}
+
+	commit(changed?: Memory[]): MemoryNode[] {
+		return this.rebuild(changed);
+	}
+
+	async backfillEmbeddings(llm: any): Promise<number> {
+		const missing = this.memories.filter(m => !m.embedding?.length);
+		if(!missing.length) return 0;
+
+		await Promise.all(missing.map(node => embedMemoryFields(node, llm)));
+		this.commit(missing);
+
+		return missing.length;
+	}
+}

src/memory/memory.ts

@@ -0,0 +1,348 @@
+import {AiTool} from '../tools.ts';
+import type {LLMMessage, LLMRequest} from '../llm.ts';
+import {MemoryCache, memoryStore} from './memory-state.ts';
+import {cosineDistance, embedMemoryFields, stripHeader, updateMemory} from '../utils.ts';
+
+const FACT_SIMILARITY_THRESHOLD = 0.62;
+const DUPLICATE_THRESHOLD = 0.68;
+const PROTECTED_MEMORIES = ['People/User'];
+const COLLECTION_WORDS = ['project', 'projects', 'people', 'person', 'managed', 'guides', 'guide', 'research', 'class', 'classes'];
+
+export type Memory = {
+	name: string;
+	description: string;
+	content: string;
+	embedding: number[];
+	titleEmbedding?: number[];
+	bodyEmbeddings?: number[][];
+	links: string[];
+	backlinks: string[];
+}
+
+export type MemoryRef = {
+	name: string;
+	description: string;
+	distance?: number;
+}
+
+export type MemoryOptions = {
+	memory: Memory[] | MemoryCache;
+	inject?: boolean;
+	tool?: boolean;
+	update?: boolean;
+	maxTokens?: number;
+}
+
+export type MemoryStore = Memory[] | MemoryCache;
+
+/** Create an empty memory shell. */
+function emptyNode(name: string, description = ''): Memory {
+	return {name, description, content: `# ${name.split('/').pop()}\n`, embedding: [], links: [], backlinks: []};
+}
+
+function renderNode(node: Memory): string {
+	return `### ${node.name}
+Description: ${node.description}
+Links: ${[...node.links, ...node.backlinks].join(', ') || 'none'}
+
+\`\`\`markdown
+${node.content}
+\`\`\``;
+}
+
+function factSimilarity(a: Memory, b: Memory): number {
+	return !a.bodyEmbeddings?.length || !b.bodyEmbeddings?.length ? 0 : Math.max(...a.bodyEmbeddings.flatMap(av => b.bodyEmbeddings!.map(bv => 1 - cosineDistance(av, bv))));
+}
+
+function words(text: string): string[] {
+	return [...new Set(text.toLowerCase().replace(/[[\]()/_-]/g, ' ').replace(/[^a-z0-9\s]/g, '').split(/\s+/).filter(w => w && !COLLECTION_WORDS.includes(w)))];
+}
+
+function jaccard(a: string[], b: string[]): number {
+	const bs = new Set(b), hit = a.filter(x => bs.has(x)).length, total = new Set([...a, ...b]).size;
+	return total ? hit / total : 0;
+}
+
+function duplicateScore(a: Memory, b: Memory): number {
+	const name = Math.max(
+		jaccard(words(a.name), words(b.name)),
+		jaccard(words(a.name.split('/').pop() || a.name), words(b.name.split('/').pop() || b.name)),
+	);
+	const desc = jaccard(words(a.description), words(b.description));
+	const body = factSimilarity(a, b);
+	const emb = a.embedding?.length && b.embedding?.length && a.embedding.length === b.embedding.length ? 1 - cosineDistance(a.embedding, b.embedding) : 0;
+	return Math.max(body, name * 0.9 + desc * 0.06 + emb * 0.04, emb * 0.55 + name * 0.35 + desc * 0.1);
+}
+
+function homeScore(node: Memory): number {
+	return (PROTECTED_MEMORIES.includes(node.name) ? 1e9 : 0)
+		+ (node.name.includes('/') ? 4 : 0)
+		+ (node.description && node.description !== 'Persistent memory document' ? 1 : 0)
+		+ Math.min(stripHeader(node.content).length / 1000, 5);
+}
+
+function pickMerge(a: Memory, b: Memory, touched: Set<string>): [drop: Memory, home: Memory] {
+	const as = homeScore(a), bs = homeScore(b);
+	if(touched.has(a.name) && !touched.has(b.name)) return as > bs + 2 ? [b, a] : [a, b];
+	if(touched.has(b.name) && !touched.has(a.name)) return bs > as + 2 ? [a, b] : [b, a];
+	return as <= bs ? [a, b] : [b, a];
+}
+
+/** Build memory tools and memory index text. */
+export function memoryTools(llm: any, memories: MemoryStore): {tools: AiTool[]; list: string} {
+	const store = memoryStore(memories);
+	const names = new Map<string, string>();
+
+	for(const node of store.list)
+		if(!names.has(node.name)) names.set(node.name, `${node.name} - ${node.description}`);
+
+	for(const name of store.ghosts())
+		if(!names.has(name)) names.set(name, `${name} - ghost node`);
+
+	return {
+		list: [...names.values()].join('\n'),
+		tools: [
+			{
+				name: 'memory_search',
+				description: 'Semantically search memories for most relevant',
+				args: {
+					query: {type: 'string', description: 'Search query', required: true},
+					limit: {type: 'number', description: 'Maximum results, default 5', default: 5},
+				},
+				fn: async ({query, limit = 5}) => {
+					if(!query?.trim()) return 'Search query is required.';
+					const [chunk] = await llm.embedding(query, {maxTokens: 8000, overlapTokens: 0});
+					if(!chunk?.embedding) return 'Failed to create embedding from query';
+					const results = store.search(chunk.embedding, limit).map(ref => store.find(ref.name)).filter((node): node is Memory => !!node);
+					return results.length ? results.map(renderNode).join('\n\n---\n\n') : 'No relevant memories found.';
+				},
+			},
+			{
+				name: 'memory_read',
+				description: 'Read an entire memory document by name',
+				args: {name: {type: 'string', description: 'Exact document name', required: true}},
+				fn: async ({name}) => {
+					const node = store.find(name);
+					return node ? renderNode(node) : store.ghosts().includes(name) ? `"${name}" is a ghost node with no document of its own.` : `Not found: "${name}".`;
+				},
+			},
+			{
+				name: 'memory_delete',
+				description: 'Delete a duplicate or merged memory',
+				args: {name: {type: 'string', description: 'Exact document name', required: true}},
+				fn: async ({name}) => {
+					store.forget(name);
+					return `Removed: ${name}`;
+				},
+			},
+			{
+				name: 'memory_write',
+				description: 'Create or replace a memory document.',
+				args: {
+					name: {type: 'string', description: 'Document name following the entity naming convention.', required: true},
+					description: {type: 'string', description: 'One factual sentence describing the entire document subject', required: true},
+					content: {type: 'string', description: 'Complete Markdown document body, including the # title', required: true},
+				},
+				fn: async (args: any) => {
+					const name = String(args.name || '').trim();
+					if(!name) return 'A document name is required.';
+					const description = String(args.description || '').trim();
+					if(!description) return 'A document description is required.';
+					const content = String(args.content || '').trim();
+					if(!content) return 'Document content is required.';
+
+					let node = store.find(name);
+					if(!node) {
+						node = emptyNode(name, description);
+						if(store.cache) store.cache.add(node);
+						else store.list.push(node);
+					}
+
+					node.description = name === 'People/User' ? 'All information about the current user' : description.replace(/\s+/g, ' ').trim();
+					node.content = updateMemory(node, content);
+					await embedMemoryFields(node, llm);
+					store.cache?.commit([node]);
+					return `Updated ${name}`;
+				},
+			},
+		],
+	};
+}
+
+export class MemoryManager {
+	private memorized = new WeakMap<LLMMessage[], LLMMessage>();
+
+	constructor(private llm: any) {}
+
+	static normalize(memory?: Memory[] | MemoryCache | MemoryOptions): MemoryOptions | null {
+		if(!memory) return null;
+		if(Array.isArray(memory) || memory instanceof MemoryCache) return {memory, inject: true, tool: false, update: false};
+		if(typeof memory === 'object' && 'memory' in memory) return {inject: true, tool: false, update: false, ...memory};
+		return null;
+	}
+
+	private memorySystem(list: string): string {
+		return `You maintain notes written in markdown used for memories from recent conversations using your tools.
+Only preserve durable information worth remembering established by the USER.
+Do not store assistant guesses, speculation, suggestions, commentary, temporary state, or details that are not worth remembering.
+
+## Rules
+- ALWAYS READ a target memory before changing it, \`memory_write\` does a full replace, it DOES NOT append!
+- Memories should contain the final state, not deltas
+- New conversational context is authoritative when it contracts existing information; reconcile it
+- Only remove information when stale, contradicted or duplicated; always preserve existing information, formatting and keep related information together
+- Only merge memories when two or more nodes are clearly about the same thing; only split a memory when it is clearly about two distinct subjects
+- Use [[WikiLinks]] liberally to record aliases and relationships between entities, even ones without pages yet (ghost nodes)
+- Use headings, subheadings, lists, tables and other markdown formatting to make documents clean
+- Maintain a \`## Todo List\` of checkboxes AS THE FIRST SUBHEADING when an entity has tasks
+	- Only create todo items for USER tasks, not AI work
+	- Only store each in one place, no duplicates
+	- Use  \`People/User\` for personal tasks or as a fallback
+
+## Naming
+- Every fact should be grouped with the owning entity
+- Always follow the naming convention \`Collection/(Pro)Noun\`
+- Facts about the user belong under People/User
+- Reuse existing memories when they are clearly the same entity including aliases and ghost references.
+- Only create deeper paths when there is a real parent/child entity relationship: \`School/Class/Chapter\`
+
+Valid Examples:
+- People/User
+- People/John Smith
+- Projects/Momentum
+- Projects/Momentum/Marketing
+- Research/Object Recognition
+- Guides/HAM Radio SOP
+
+## Workflow
+
+1. Create groups of durable information and todos based on the owning entity & naming rules above
+2. For each group:
+  1. Read the existing memory(s)
+  2. Merge the information & todos based on the rules above
+  3. Write the entire patched document
+
+Available memories:
+
+${list || 'No memory documents exist yet.'}`;
+	}
+
+	private touchedNames(history: LLMMessage[]): string[] {
+		return [...new Set(history
+			.filter((h: any) => h.role === 'tool' && h.name === 'memory_write' && !h.error)
+			.map((h: any) => String(h.args?.name || h.content?.match(/^Updated (.+)$/)?.[1] || '').trim())
+			.filter(Boolean))];
+	}
+
+	private async backfillEmbeddings(store: ReturnType<typeof memoryStore>): Promise<void> {
+		const missing = store.list.filter(m => !m.embedding?.length || !m.titleEmbedding?.length || !m.bodyEmbeddings?.length);
+		await Promise.all(missing.map(m => embedMemoryFields(m, this.llm)));
+		store.cache?.commit(missing);
+	}
+
+	private closestDuplicate(node: Memory, store: ReturnType<typeof memoryStore>): Memory | null {
+		return store.list
+			.filter(m => m.name !== node.name && !m.name.startsWith('Journal/') && !node.name.startsWith('Journal/'))
+			.map(m => ({node: m, score: duplicateScore(node, m)}))
+			.filter(x => x.score >= DUPLICATE_THRESHOLD || factSimilarity(node, x.node) >= FACT_SIMILARITY_THRESHOLD)
+			.sort((a, b) => b.score - a.score)[0]?.node || null;
+	}
+
+	private async rehomeDeleted(drop: Memory, home: Memory, memories: MemoryStore, options: LLMRequest): Promise<void> {
+		const store = memoryStore(memories);
+		const backup = structuredClone(drop);
+		store.forget(drop.name);
+
+		try {
+			const memory = memoryTools(this.llm, memories);
+			await this.llm.ask(`A duplicate memory document was removed automatically.
+
+Deleted document:
+${renderNode(backup)}
+
+Closest surviving home:
+${renderNode(home)}
+
+Reinsert every durable unique fact, useful relationship, alias, and user todo from the deleted document into the best remaining memory document.
+Usually this should be "${home.name}", but use another existing memory if it is a better home.
+Read before writing. Write full replacement documents only.
+Do NOT recreate "${backup.name}" unless the deletion was wrong and it is clearly a distinct persistent entity.`, {
+				model: options.memoryModel || options.model,
+				temperature: 0.2,
+				maxTokens: options.maxTokens,
+				tools: memory.tools,
+				history: [],
+				system: this.memorySystem(memory.list),
+			});
+		} catch(err) {
+			if(!store.find(backup.name)) store.cache ? store.cache.add(backup) : store.list.push(backup);
+			throw err;
+		} finally {
+			store.cache?.commit(store.list);
+		}
+	}
+
+	private async reconcileSimilar(history: LLMMessage[], memories: MemoryStore, options: LLMRequest): Promise<void> {
+		const store = memoryStore(memories);
+		const touched = new Set(this.touchedNames(history));
+		const targets = store.list.filter(m => touched.has(m.name) || [...touched].some(t => duplicateScore(m, store.find(t) || m) >= DUPLICATE_THRESHOLD));
+		const deleted = new Set<string>();
+		if(!targets.length) return;
+		await this.backfillEmbeddings(store);
+
+		for(const node of targets) {
+			if(!store.find(node.name) || deleted.has(node.name) || PROTECTED_MEMORIES.includes(node.name)) continue;
+			const closest = this.closestDuplicate(node, store);
+			if(!closest) continue;
+			const [drop, home] = pickMerge(node, closest, touched);
+			if(deleted.has(drop.name) || PROTECTED_MEMORIES.includes(drop.name)) continue;
+			deleted.add(drop.name);
+			await this.rehomeDeleted(drop, home, memories, options);
+			await this.backfillEmbeddings(store);
+		}
+	}
+
+	async recollect(query: string, memory: MemoryStore, limit = 15): Promise<Memory[]> {
+		const store = memoryStore(memory);
+		if(!store.list.length || !query?.trim()) return [];
+		const [chunk] = await this.llm.embedding(query, {maxTokens: 8000, overlapTokens: 0});
+		return !chunk?.embedding ? [] : store.search(chunk.embedding, limit).map(ref => store.find(ref.name)).filter((m: Memory | undefined): m is Memory => !!m);
+	}
+
+	get tools(): {read: (memory: MemoryStore) => AiTool[]} {
+		return {read: (memory: MemoryStore) => memoryTools(this.llm, memory).tools};
+	}
+
+	async memorize(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest = {},): Promise<Memory[]> {
+		const store = memoryStore(memories);
+		const previous = this.memorized.get(history);
+		let start = 0;
+
+		if(previous) {
+			const index = history.indexOf(previous);
+			if(index >= 0) start = index + 1;
+		}
+
+		const turns = history.slice(start).filter((h: any) => h.role === 'user' || h.role === 'assistant');
+		const conversation = turns.map((h: any) => `[${h.role}]: ${h.content}`).join('\n\n').trim();
+		if(!conversation) return store.list;
+
+		const memory = memoryTools(this.llm, memories);
+		const memoryHistory: LLMMessage[] = [];
+
+		await this.llm.ask(conversation, {
+			model: options.memoryModel || options.model,
+			temperature: 0.2,
+			maxTokens: options.maxTokens,
+			tools: memory.tools,
+			history: memoryHistory,
+			system: this.memorySystem(memory.list),
+		});
+
+		await this.reconcileSimilar(memoryHistory, memories, options);
+
+		const lastTurn = turns.at(-1);
+		if(lastTurn) this.memorized.set(history, lastTurn);
+		return store.list;
+	}
+}

src/open-ai.ts

@@ -1,91 +1,99 @@
 import {OpenAI as openAI} from 'openai';
-import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, clean} from '@ztimson/utils';
+import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, clean, makeArray} from '@ztimson/utils';
 import {AbortablePromise, Ai} from './ai.ts';
 import {LLMMessage, LLMRequest} from './llm.ts';
 import {LLMProvider} from './provider.ts';
+import {TokenPool} from './token-pool.ts';
 import {convertSchema} from './tools.ts';
 
 export class OpenAi extends LLMProvider {
-	client!: openAI;
+	tokenPool!: TokenPool;
+	private clients = new Map<string, openAI>();
 
-	constructor(public readonly ai: Ai, public readonly host: string | null, public readonly token: string, public model: string) {
+	constructor(public readonly ai: Ai, public readonly host: string | null, public readonly token: string | string[], public model: string) {
 		super();
-		this.client = new openAI(clean({
+		const tokens = makeArray(token).filter(Boolean);
-			baseURL: host,
+		this.tokenPool = new TokenPool(...(tokens.length ? tokens : [host ? 'ignored' : '']));
-			apiKey: token || (host ? 'ignored' : undefined)
-		}));
 	}
 
-	private toStandard(history: any[]): LLMMessage[] {
+	private getClient(token: string): openAI {
+		let client = this.clients.get(token);
+		if(!client) {
+			client = new openAI(clean({baseURL: this.host, apiKey: token || undefined}));
+			this.clients.set(token, client);
+		}
+		return client;
+	}
+
+	private toWireContent(content: any): any {
+		if(!Array.isArray(content)) return content;
+		return content.map(c => c.type === 'image'
+			? {type: 'image_url', image_url: {url: `data:${c.mime};base64,${c.data}`}}
+			: {type: 'text', text: c.text});
+	}
+
+	/** Convert standard history -> OpenAI wire format */
+	private toWire(history: LLMMessage[], system?: string): any[] {
+		const wire: any[] = [];
+		if(system) wire.push({role: 'system', content: system});
+
 		for(let i = 0; i < history.length; i++) {
 			const h = history[i];
-			if(h.role === 'assistant' && h.tool_calls) {
-				const items: any[] = [];
-				if(h.content) items.push({role: 'assistant', content: h.content, timestamp: h.timestamp, duration: h.duration, tps: h.tps});
-				items.push(...h.tool_calls.map((tc: any) => ({
-					role: 'tool',
-					id: tc.id,
-					name: tc.function.name,
-					args: JSONAttemptParse(tc.function.arguments, {}),
-					timestamp: h.timestamp,
-					duration: h.duration,
-					tps: h.tps
-				})));
-				history.splice(i, 1, ...items);
-				i += items.length - 1;
-			} else if(h.role === 'tool') {
-				const record = history.find(h2 => h.tool_call_id == h2.id);
-				if(record) {
-					if(h.content?.includes('"error":')) record.error = h.content;
-					else record.content = h.content || '';
-				}
-				history.splice(i, 1);
-				i--;
-			}
-			if(!history[i]?.timestamp) history[i].timestamp = Date.now();
-		}
-		return history;
-	}
 
-	private fromStandard(history: LLMMessage[]): any[] {
+			if(h.role !== 'tool') {
-		return history.reduce((result, h) => {
+				wire.push({role: h.role, content: this.toWireContent(h.content)});
-			if(h.role === 'tool') {
+				continue;
-				result.push({
-					role: 'assistant',
-					content: null,
-					tool_calls: [{ id: h.id, type: 'function', function: { name: h.name, arguments: JSON.stringify(h.args) } }],
-					refusal: null,
-					annotations: [],
-					timestamp: h.timestamp,
-				}, {
-					role: 'tool',
-					tool_call_id: h.id,
-					content: h.error || h.content,
-					timestamp: h.timestamp,
-				});
-			} else {
-				result.push(h);
 			}
-			return result;
+
-		}, [] as any[]);
+			const calls: any[] = [];
+			const results: any[] = [];
+
+			while(i < history.length && history[i].role === 'tool') {
+				const tool: any = history[i];
+
+				calls.push({
+					id: tool.id,
+					type: 'function',
+					function: {
+						name: tool.name,
+						arguments: JSON.stringify(tool.args || {})
+					}
+				});
+
+				results.push({
+					role: 'tool',
+					tool_call_id: tool.id,
+					content: tool.error || tool.content || ''
+				});
+
+				i++;
+			}
+
+			wire.push({
+				role: 'assistant',
+				content: null,
+				tool_calls: calls
+			});
+
+			wire.push(...results);
+			i--;
+		}
+
+		return wire;
 	}
 
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<string | any> {
 		const controller = new AbortController();
 		return Object.assign(new Promise<any>(async (res, rej) => {
-			const base = (options.history || []).filter(h => h.role !== 'system');
+			if(!options.history) options.history = [];
-			let history = this.fromStandard([
+			const history = options.history;
-				...(options.system ? [{role: <any>'system', content: options.system, timestamp: Date.now()}] : []),
+			if(message) history.push({role: 'user', content: message, timestamp: Date.now()});
-				...base,
-				{role: 'user', content: message, timestamp: Date.now()}
-			]);
 			const tools = options.tools || this.ai.options.llm?.tools || [];
 			const requestParams: any = {
 				model: options.model || this.model,
-				messages: history,
 				stream: !!options.stream,
-				max_completion_tokens: options.max_tokens || this.ai.options.llm?.max_tokens || undefined,
+				max_completion_tokens: options.maxTokens ?? this.ai.options.llm?.maxTokens,
-				temperature: options.temperature || this.ai.options.llm?.temperature || undefined,
+				temperature: options.temperature ?? this.ai.options.llm?.temperature,
 				tools: tools.map(t => ({
 					type: 'function',
 					function: {
@@ -93,8 +101,12 @@ export class OpenAi extends LLMProvider {
 						description: t.description,
 						parameters: {
 							type: 'object',
-							properties: t.args ? objectMap(t.args, (key, value) => ({...value, required: undefined})) : {},
+							properties: t.args
-							required: t.args ? Object.entries(t.args).filter(t => t[1].required).map(t => t[0]) : []
+								? objectMap(t.args, (key, value) => ({...value, required: undefined}))
+								: {},
+							required: t.args
+								? Object.entries(t.args).filter(t => t[1].required).map(t => t[0])
+								: []
 						}
 					}
 				}))
@@ -104,106 +116,138 @@ export class OpenAi extends LLMProvider {
 				const schema = convertSchema(options.schema);
 				requestParams.response_format = {
 					type: 'json_schema',
-					json_schema: {
+					json_schema: {name: 'response', strict: true, schema}
-						name: 'response',
-						strict: true,
-						schema
-					}
 				};
 			}
-
 			if(options.stream) requestParams.stream_options = {include_usage: true};
-			let resp: any, terminal = false, duration = 0, tps = 0;
-			do {
-				requestParams.messages = history.map(({timestamp, ...m}) => m);
-				const callStart = Date.now();
-				resp = await this.client.chat.completions.create(requestParams).catch(err => {
-					err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
-					throw err;
-				});
 
-				let usage: any;
+			try {
-				if(options.stream) {
+				let terminal = false;
-					resp.choices = [{message: {role: 'assistant', content: '', tool_calls: [], timestamp: Date.now()}}];
+				let iteration = 0;
-					for await (const chunk of resp) {
+
-						if(controller.signal.aborted) break;
+				do {
-						if(chunk.usage) usage = chunk.usage;
+					iteration++;
-						if(chunk.choices[0]?.delta?.content) {
+					requestParams.messages = this.toWire(history.filter(h => h.role !== 'system'), options.system);
-							resp.choices[0].message.content += chunk.choices[0].delta.content;
+
-							options.stream({text: chunk.choices[0].delta.content});
+					const callStart = Date.now();
-						}
+					const resp: any = await this.tokenPool.run(token =>
-						if(chunk.choices[0]?.delta?.tool_calls) {
+						this.getClient(token).chat.completions.create(requestParams)
-							for(const deltaTC of chunk.choices[0].delta.tool_calls) {
+					).catch(err => {
-								const existing = resp.choices[0].message.tool_calls.find(tc => tc.index === deltaTC.index);
+						err.message += `\n\nMessages:\n${JSON.stringify(requestParams.messages, null, 2)}`;
-								if(existing) {
+						throw err;
-									if(deltaTC.id) existing.id = deltaTC.id;
+					});
-									if(deltaTC.type) existing.type = deltaTC.type;
+
-									if(deltaTC.function) {
+					let usage: any;
-										if(!existing.function) existing.function = {};
+					let finishReason: string | undefined;
-										if(deltaTC.function.name) existing.function.name = deltaTC.function.name;
+					let msg: any = {content: '', tool_calls: []};
-										if(deltaTC.function.arguments) existing.function.arguments = (existing.function.arguments || '') + deltaTC.function.arguments;
+					let streamedChars = 0;
-									}
+
-								} else {
+					if(options.stream) {
-									resp.choices[0].message.tool_calls.push({
+						let streamCompleted = false;
-										index: deltaTC.index,
+						try {
-										id: deltaTC.id || '',
+							for await (const chunk of resp) {
-										type: deltaTC.type || 'function',
+								if(controller.signal.aborted) break;
-										function: {
+								if(chunk.usage) usage = chunk.usage;
-											name: deltaTC.function?.name || '',
+
-											arguments: deltaTC.function?.arguments || ''
+								const choice = chunk.choices?.[0];
+								if(choice?.finish_reason) finishReason = choice.finish_reason;
+
+								if(choice?.delta?.content) {
+									msg.content += choice.delta.content;
+									streamedChars += choice.delta.content.length;
+									options.stream({text: choice.delta.content});
+								}
+
+								if(choice?.delta?.tool_calls) {
+									for(const deltaTC of choice.delta.tool_calls) {
+										const index = deltaTC.index ?? msg.tool_calls.length;
+										let existing = msg.tool_calls.find((tc: any) => tc.index === index);
+
+										if(!existing) {
+											existing = {index, id: '', function: {name: '', arguments: ''}};
+											msg.tool_calls.push(existing);
 										}
-									});
+
+										if(deltaTC.id) existing.id = deltaTC.id;
+										if(deltaTC.function?.name) existing.function.name = deltaTC.function.name;
+										if(deltaTC.function?.arguments) existing.function.arguments += deltaTC.function.arguments;
+									}
 								}
 							}
+
+							streamCompleted = true;
+						} catch(err) {
+							if(!controller.signal.aborted) throw err;
 						}
+
+						if(streamCompleted && !finishReason) finishReason = msg.tool_calls.length ? 'tool_calls' : 'stop';
+					} else {
+						usage = resp.usage;
+						finishReason = resp.choices[0].finish_reason;
+						msg = resp.choices[0].message;
 					}
-				} else {
-					usage = resp.usage;
-				}
-				duration = Date.now() - callStart;
-				tps = usage?.completion_tokens && duration > 0 ? usage.completion_tokens / (duration / 1000) : 0;
 
-				if(resp.error) throw new Error(resp.error);
+					const duration = Date.now() - callStart;
-				const toolCalls = resp.choices[0].message.tool_calls || [];
+					const tps = usage?.completion_tokens && duration > 0 ? usage.completion_tokens / (duration / 1000) : 0;
-				if(toolCalls.length && !controller.signal.aborted) {
-					history.push({...resp.choices[0].message, duration, tps});
-					const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
-						const tool = tools?.find(findByProp('name', toolCall.function.name));
-						if(options.stream) options.stream({tool: toolCall.function.name});
-						if(!tool) return {role: 'tool', tool_call_id: toolCall.id, content: '{"error": "Tool not found"}', timestamp: Date.now()};
-						try {
-							const args = JSONAttemptParse(toolCall.function.arguments, {});
-							const toolStream = options.stream && ((chunk: any) => {
-								if(chunk.done) { terminal = true; return; }
-								options.stream!(chunk);
-							});
-							const result = await tool.fn(args, toolStream, this.ai, toolCall.id);
-							return {role: 'tool', tool_call_id: toolCall.id, content: typeof result == 'object' ? JSONSanitize(result) : result, timestamp: Date.now()};
-						} catch (err: any) {
-							return {role: 'tool', tool_call_id: toolCall.id, content: JSONSanitize({error: err?.message || err?.toString() || 'Unknown'}), timestamp: Date.now()};
-						}
-					}));
-					history.push(...results);
-					requestParams.messages = history;
-				}
-			} while (!terminal && !controller.signal.aborted && resp.choices?.[0]?.message?.tool_calls?.length);
 
-			if(!terminal) {
+					if(finishReason === 'length' && !controller.signal.aborted) {
-				const textContent = resp.choices[0].message.content || '';
+						if(msg.content?.trim()) history.push({role: 'assistant', content: msg.content.trim(), timestamp: Date.now(), duration, tps});
-				history.push({role: 'assistant', content: textContent.trim(), timestamp: Date.now(), duration, tps});
+						throw new Error(`[OpenAI] Response hit token limit before completing`);
+					}
+
+					if(!finishReason && !controller.signal.aborted) {
+						throw new Error('[OpenAI] Completion ended without a usable response');
+					}
+
+					const toolCalls = msg.tool_calls || [];
+
+					if(toolCalls.length && !controller.signal.aborted) {
+						if(msg.content?.trim()) history.push({role: 'assistant', content: msg.content.trim(), timestamp: Date.now(), duration, tps});
+
+						const entries = toolCalls.map((tc: any) => {
+							const entry: any = {
+								role: 'tool',
+								id: tc.id,
+								name: tc.function.name,
+								args: JSONAttemptParse(tc.function.arguments, {}),
+								content: undefined,
+								timestamp: Date.now()
+							};
+
+							history.push(entry);
+							return {tc, entry};
+						});
+
+						await Promise.all(entries.map(async ({tc, entry}: any) => {
+							const tool = tools.find(findByProp('name', tc.function.name));
+							if(options.stream) options.stream({tool: tc.function.name});
+							if(!tool) return entry.error = 'Tool not found';
+							try {
+								const toolStream = options.stream && ((chunk: any) => {
+									if(chunk.done) return;
+									options.stream!(chunk);
+								});
+
+								const result = await tool.fn(entry.args, toolStream, this.ai, tc.id);
+								entry.content = typeof result === 'object' ? JSONSanitize(result) : result;
+							} catch(err: any) {
+								entry.error = err?.message || err?.toString() || 'Unknown';
+							}
+						}));
+					} else {
+						terminal = true;
+						const text = (msg.content || '').trim();
+						if(text) history.push({role: 'assistant', content: text, timestamp: Date.now(), duration, tps});
+					}
+				} while(!terminal && !controller.signal.aborted);
+
+				if(options.stream) options.stream({done: true});
+				const turnStart = history.map(h => h.role).lastIndexOf('user');
+				const finalContent = history.slice(turnStart + 1).reduce((str, h) => h.role === 'assistant' ? str + (h.content || '') : str, '').trim();
+				res(options.schema ? JSONAttemptParse(finalContent, finalContent) : finalContent);
+			} catch(err) {
+				rej(err);
 			}
-
-			history = this.toStandard(history);
-			if(options.history) options.history.splice(0, options.history.length, ...history.filter(h => h.role !== 'system'));
-			if(options.stream) options.stream({done: true});
-
-			const turnStart = history.map(h => h.role).lastIndexOf('user');
-			const finalContent = history.slice(turnStart + 1).reduce((str, h) => {
-				if(h.role === 'assistant') return str + (h.content || '');
-				return str;
-			}, '').trim();
-
-			res(options.schema ? JSONAttemptParse(finalContent, finalContent) : finalContent);
 		}), {abort: () => controller.abort()});
 	}
 }

src/token-pool.ts

@@ -0,0 +1,65 @@
+const DEFAULT_COOLDOWN = 15 * 60 * 1000;
+
+type TokenState = {
+	token: string;
+	cooldownUntil: number; // 0 = available now
+	lastError?: {code: number, message: string};
+};
+
+export class TokenPoolExhaustedError extends Error {
+	constructor(public tokens: Record<string, {code: number, message: string}>) {
+		super(`All tokens exhausted:\n${Object.entries(tokens).map(([t, e]) => `${t}: [${e.code}] ${e.message}`).join('\n')}`);
+		this.name = 'TokenPoolExhaustedError';
+	}
+}
+
+export class TokenPool {
+	private states: TokenState[];
+
+	constructor(...tokens: string[]) {
+		this.states = tokens.map(token => ({token, cooldownUntil: 0}));
+	}
+
+	private preview(token: string): string {
+		return token.length <= 8 ? '****' : `${token.slice(0, 4)}...${token.slice(-4)}`;
+	}
+
+	/** Anthropic & OpenAI SDKs both attach `status` to thrown errors */
+	private statusCode(err: any): number {
+		return err?.status ?? err?.response?.status ?? err?.statusCode;
+	}
+
+	private retryAfter(err: any): number {
+		const headers = err?.headers || err?.response?.headers;
+		const raw = headers?.get?.('retry-after') ?? headers?.['retry-after'];
+		if(raw) {
+			const seconds = Number(raw);
+			if(!isNaN(seconds)) return Date.now() + seconds * 1000;
+			const date = new Date(raw).getTime();
+			if(!isNaN(date)) return date;
+		}
+		return Date.now() + DEFAULT_COOLDOWN;
+	}
+
+	async run<T>(fn: (token: string) => Promise<T>): Promise<T> {
+		const now = Date.now();
+		for(const state of this.states) {
+			if(state.cooldownUntil > now) continue;
+			try {
+				const result = await fn(state.token);
+				state.cooldownUntil = 0;
+				state.lastError = undefined;
+				return result;
+			} catch(err: any) {
+				const code = this.statusCode(err);
+				if(![401, 403, 429].includes(code)) throw err;
+				state.cooldownUntil = code === 429 ? this.retryAfter(err) : Date.now() + DEFAULT_COOLDOWN;
+				state.lastError = {code, message: err?.message || 'Unknown error'};
+			}
+		}
+
+		const failures: Record<string, {code: number, message: string}> = {};
+		this.states.forEach(s => { if(s.lastError) failures[this.preview(s.token)] = s.lastError; });
+		throw new TokenPoolExhaustedError(failures);
+	}
+}

src/utils.ts

@@ -0,0 +1,82 @@
+import {Memory} from './memory/memory.ts';
+
+export function cosineDistance(a: number[], b: number[]): number {
+	let dot = 0, normA = 0, normB = 0;
+	for(let i = 0; i < a.length; i++) {
+		dot += a[i] * b[i];
+		normA += a[i] * a[i];
+		normB += b[i] * b[i];
+	}
+	const denom = Math.sqrt(normA) * Math.sqrt(normB);
+	return denom === 0 ? 1 : 1 - dot / denom;
+}
+
+export async function embedMemoryFields(node: Memory, llm: any): Promise<void> {
+	const body = stripHeader(node.content);
+	const [titleE] = await llm.embedding(node.name.split('/').pop() || node.name);
+	const [descE] = await llm.embedding(node.description || '');
+	const bodyChunks = body ? await llm.embedding(body) : [];
+	if(titleE) node.titleEmbedding = titleE.embedding;
+	if(descE) node.embedding = descE.embedding;
+	node.bodyEmbeddings = bodyChunks.map((c: any) => c.embedding).filter(Boolean);
+}
+
+export function euclideanDistance(a: number[], b: number[]): number {
+	let sum = 0;
+	for(let i = 0; i < a.length; i++) {
+		const d = a[i] - b[i];
+		sum += d * d;
+	}
+	return Math.sqrt(sum);
+}
+
+export function getWeekStart(date: Date = new Date()): string {
+	const d = new Date(Date.UTC(date.getFullYear(), date.getMonth(), date.getDate()));
+	const day = d.getUTCDay();
+	const diff = day === 0 ? -6 : 1 - day;
+	d.setUTCDate(d.getUTCDate() + diff);
+	return d.toISOString().slice(0, 10);
+}
+
+export function journalDescription(journalName?: string): string {
+	const start = journalName?.split('/').pop() || getWeekStart();
+	const d = new Date(`${start}T00:00:00Z`);
+	d.setUTCDate(d.getUTCDate() + 6);
+	const end = d.toISOString().slice(0, 10);
+	return `Log from ${start} - ${end}`;
+}
+
+function parseFrontmatter(content: string): {fm: Map<string, string>, body: string} {
+	const match = content.match(/^---\n([\s\S]*?)\n---\n?([\s\S]*)$/);
+	if(!match) return {fm: new Map(), body: content};
+	const fm = new Map<string, string>();
+	for(const line of match[1].split('\n')) {
+		const i = line.indexOf(':');
+		if(i === -1) continue;
+		const key = line.slice(0, i).trim();
+		const raw = line.slice(i + 1).trim();
+		let value = raw;
+		try { value = JSON.parse(raw); } catch { }
+		fm.set(key, value);
+	}
+	return {fm, body: match[2]};
+}
+
+export function writeFrontmatter(fm: Map<string, string>, body: string): string {
+	const lines = [...fm.entries()].map(([k, v]) =>
+		`${k}: ${JSON.stringify(String(v).replace(/\s+/g, ' ').trim())}`);
+	return `---\n${lines.join('\n')}\n---\n\n${body.trimStart()}`;
+}
+
+export function stripHeader(content: string): string {
+	return content.replace(/^---[\s\S]*?\n---\n?/, '').trimStart();
+}
+
+export function updateMemory(node: Memory, body: string): string {
+	const {fm} = parseFrontmatter(node.content);
+	fm.set('name', node.name);
+	fm.set('description', (node.name.startsWith('Journal/') ? journalDescription(node.name) : node.description)
+		|| 'Persistent memory document');
+	fm.set('modified', new Date().toISOString());
+	return writeFrontmatter(fm, stripHeader(body));
+}