Fix open-ai early termination & memory improvements

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},

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": "1.4.5",
+	"version": "1.6.12",
 	"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

@@ -24,6 +24,13 @@ export class Anthropic extends LLMProvider {
 		return client;
 	}
 
+	private toWireContent(content: any): any {
+		if(!Array.isArray(content)) return content;
+		return content.map(c => c.type === 'image'
+			? {type: 'image', source: {type: 'base64', media_type: c.mime, data: c.data}}
+			: {type: 'text', text: c.text});
+	}
+
 	/** Convert standard history -> Anthropic wire format */
 	private toWire(history: LLMMessage[]): any[] {
 		const wire: any[] = [];
@@ -34,7 +41,7 @@ export class Anthropic extends LLMProvider {
 					{role: 'user', content: [{type: 'tool_result', tool_use_id: h.id, is_error: !!h.error, content: h.error || h.content || ''}]}
 				);
 			} else {
-				wire.push({role: h.role, content: h.content});
+				wire.push({role: h.role, content: this.toWireContent(h.content)});
 			}
 		}
 		return wire;
@@ -50,7 +57,7 @@ export class Anthropic extends LLMProvider {
 			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 => ({

src/helpers.ts

@@ -13,6 +13,56 @@ export function extractLinks(content: string): string[] {
 	return [...new Set([...matches].map(m => m[1].trim()))];
 }
 
+/**
+ * Incrementally patch the graph for a set of changed memories, instead of
+ * re-scanning every document. Only the changed memories' own content is
+ * re-parsed for links; affected targets have their backlinks patched.
+ * Does NOT handle node deletion — full rebuildGraph() is still required
+ * when a memory is removed, since that needs a backlink sweep across
+ * everyone who might reference it.
+ */
+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));

src/kd-tree.ts

@@ -15,6 +15,7 @@ interface KDNode<T> {
 	axis: number;
 	left: KDNode<T> | null;
 	right: KDNode<T> | null;
+	deleted?: boolean;
 }
 
 // ─── Distance helpers ─────────────────────────────────────────────────────────
@@ -95,6 +96,7 @@ 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
@@ -103,6 +105,7 @@ class BoundedMaxHeap<T> {
 export class KDTree<T = unknown> {
 	private root: KDNode<T> | null = null;
 	private _size = 0;
+	private _tombstones = 0;
 	private readonly distanceFn: (a: number[], b: number[]) => number;
 
 	readonly dims: number;
@@ -129,9 +132,15 @@ 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 ──────────────────────────────────────────────────────────────
 
 	/**
@@ -144,10 +153,36 @@ 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>[] {
@@ -171,7 +206,7 @@ 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>[] {
@@ -186,7 +221,7 @@ 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);
@@ -194,12 +229,14 @@ 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 ─────────────────────────────────────────────────────────
@@ -251,8 +288,10 @@ 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];
@@ -285,9 +324,11 @@ 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;
@@ -310,7 +351,7 @@ 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/llm.ts

@@ -1,15 +1,20 @@
-import {clean, snakeCase} from '@ztimson/utils';
+import {clean, makeUnique, snakeCase} from '@ztimson/utils';
 import {AbortablePromise, Ai} from './ai.ts';
 import {Anthropic} from './antrhopic.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 {Memory, MemoryCache, MemoryManager, MemoryOptions, stripHeader} 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';
 
 const MAX_AGENT_DEPTH = 5;
+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[]};
@@ -27,11 +32,26 @@ export type Agent = {
 	agents?: string[] | 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 */
@@ -67,7 +87,7 @@ 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 */
@@ -88,6 +108,8 @@ export type LLMRequest = {
 	mcp?: McpServer[];
 	/** Subagents exposed as delegatable/wrapped tools */
 	agents?: Agent[];
+	/** Attach files to request */
+	files?: LLMFile[];
 	/** @internal recursion guard for nested agent delegation */
 	_agentDepth?: number;
 }
@@ -111,6 +133,11 @@ export type Skill = {
 }
 
 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;
@@ -126,7 +153,119 @@ class LLM {
 		this.memoryManager = new MemoryManager(this);
 	}
 
-	private setupAgent(agents: Agent[] = [], allAgents: Agent[], history: LLMMessage[], aborts: (() => void)[], depth = 0, delegateState: {resp: string | null}): AiTool[] {
+	private async loadBuffer(file: LLMFile, asText: boolean): Promise<Buffer> {
+		if(file.path) return fs.readFile(file.path);
+		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: ${err.message}</file>`};
+		} finally {
+			if(tmpDir) fs.rm(tmpDir, {recursive: true, force: true}).catch(() => {});
+		}
+	}
+
+	private async resolveFiles(files: LLMFile[]): Promise<{text: string, images: {mime: string, data: string}[]}> {
+		const resolved = await Promise.all(files.map(f => this.resolveFile(f)));
+		return {
+			text: resolved.filter(r => r.text).map(r => r.text).join('\n\n'),
+			images: resolved.flatMap(r => r.images || [])
+		};
+	}
+
+	private setupAgent(agents: Agent[] = [], allAgents: Agent[], history: LLMMessage[], aborts: ((keep?: boolean) => void)[], depth = 0, delegateState: {resp: string | null}): AiTool[] {
 		return agents.map(a => {
 			const toolName = `${a.delegate ? '' : 'sub'}agent_${snakeCase(a.name)}`;
 			return {
@@ -207,7 +346,7 @@ ${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
 		};
 	}
@@ -216,7 +355,7 @@ ${a.system}`,
 		if(!skills?.length) return {prompt: '', tools: []};
 		const list = skills.map(s => `- ${s.name}: ${s.description}`).join('\n');
 		return {
-			prompt: `You have access to the following skill documents, whenever there is overlap between a question and a skill file, use \`skill_read\` to get instructions and background knowledge:\n${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',
@@ -258,11 +397,13 @@ ${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));
 		};
 
 		let promise: any;
@@ -272,7 +413,24 @@ ${a.system}`,
 			let tools: AiTool[] = options.tools || this.ai.options.llm?.tools || [];
 			const prompts: string[] = [];
 			let history = options.history || [];
-			if(message) history.push({role: 'user', content: message, timestamp: Date.now()});
+			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;
@@ -301,8 +459,8 @@ ${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;
@@ -316,45 +474,64 @@ ${a.system}`,
 							} else listed.push(r);
 						}
 
-						prompts.unshift(`You have a background memory process which has prefetched relevant information${mem.update ? ' and will create new memories from this conversation' : ''} for you
+						prompts.unshift(`## Memory
+You have a background memory process which has prefetched relevant information${mem.update ? ' and will create new memories from this conversation' : ''} for you
 Assume it is perfect and never mention this process to anyone ever
 Always use your memories to craft a personalized response, they contain links / [[wiki links]] which you use navigate between them
 ${mem.tool ? `You can access memory files via the \`memory_search\` and \`memory_recall\` tools
 When you need information about the user, \`memory_recall\` \`People/User\` before asking (fetch if not included bellow)
-When you need information not provided, attempt 1-3 \`memory_search\` calls with unique queries before asking` : ''}
+When you need information not provided, attempt 1-3 \`memory_search\` calls with distinct queries before asking` : ''}
 
-${preloaded.length ? `
+${preloaded.length ? `### Prefetched Memories (Most relevant first):
-Prefetched Memories (Most relevant first):
 
 ${preloaded.map(r => `Memory: ${r.name}
 Description: ${r.description}
-Linked: ${[r.links, ...r.backlinks].join(', ')}
+Linked: ${makeUnique([...r.links, ...r.backlinks]).join(', ')}
 \`\`\`
-${r.content}
+${stripHeader(r.content)}
 \`\`\``).join('\n\n')}` : ''}
-
+${mem.tool && listed.length ? '\n' + listed.map(r => `Memory: ${r.name}
-${mem.tool && listed.length ? listed.map(r => `Memory: ${r.name}
 Description: ${r.description}
-Linked: ${[r.links, ...r.backlinks].join(', ')}
+Linked: ${makeUnique([...r.links, ...r.backlinks]).join(', ')}
-<!-- Truncated -->`).join('\n\n') : ''}
+<!-- Truncated -->`).join('\n\n') : ''}`.trim())
-
-${mem.tool ? `Full memory list:
-${mems.map(m => `- ${m.name}: ${m.description}`).join('\n')}` : ''}`.trim())
 					}
 					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;
+			}
 
 			const toolTimings = new Map<string, {duration: number, tps: number}>();
 			tools = this.wrapToolTiming(tools, toolTimings);
 
-			if(aborted) throw Object.assign(new Error('Aborted'), {name: 'AbortError'});
+			if(aborted) abortNow();
 
 			prompts.unshift(options.system || this.ai.options.llm?.system || '');
-			request = this.models[m].ask('', {...options, tools, system: prompts.filter(Boolean).join('\n\n')});
+			request = this.models[m].ask('', {...options, tools, stream, system: prompts.filter(Boolean).join('\n\n')});
-			let resp = await request;
+			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) {

src/memory.ts

@@ -1,23 +1,21 @@
-import {MemoryNode, rebuildGraph} from './helpers.ts';
+import {MemoryNode, patchGraph, rebuildGraph} from './helpers.ts';
 import {LLMRequest, LLMMessage} from './llm.ts';
 import {AiTool} from './tools.ts';
-import {KDPoint, KDTree} from './kd-tree.ts';
+import {KDTree} from './kd-tree.ts';
 
-const FACTS_HEADING = '## Facts';
+const FACT_SIMILARITY_THRESHOLD = 0.62;
-
+const PENDING_HEADING = '## Pending';
-const GENERIC_TEMPLATE = `# {{Title}}
+const TODO_HEADING = '## Todo list';
-
+const TREE_TOMBSTONE_LIMIT = 0.25;
-## Summary
+const ALIAS_MATCH_THRESHOLD = 0.55;
-
-## Details
-
-## Related`;
 
 export type Memory = {
 	name: string;
 	description: string;
 	content: string;
 	embedding: number[];
+	titleEmbedding?: number[];
+	bodyEmbeddings?: number[][];
 	links: string[];
 	backlinks: string[];
 }
@@ -25,6 +23,7 @@ export type Memory = {
 type MemoryRef = {
 	name: string;
 	description: string;
+	distance?: number;
 }
 
 type FactBucket = {
@@ -32,19 +31,32 @@ type FactBucket = {
 	facts: string[];
 }
 
+type MemoryTask = {
+	/** Exact node name / new persistent entity path this task belongs to, or '' for a personal task with no entity (goes to the journal) */
+	subject: string;
+	task: string;
+	done: boolean;
+}
+
+type FactAgentResult = {
+	buckets: FactBucket[];
+	journal: string;
+	tasks: MemoryTask[];
+}
+
 function dedupeFacts(facts: string[]): string[] {
 	const seen = new Map<string, string>();
-	for (const f of facts) {
+	for(const f of facts) {
 		const clean = f.trim();
-		if (clean) seen.set(clean.toLowerCase(), clean);
+		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++) {
+	for(let i = 0; i < a.length; i++) {
-		dot   += a[i] * b[i];
+		dot += a[i] * b[i];
 		normA += a[i] * a[i];
 		normB += b[i] * b[i];
 	}
@@ -55,14 +67,34 @@ function cosineDistance(a: number[], b: number[]): number {
 function cosineSearch(query: number[], memories: Memory[], limit: number): MemoryRef[] {
 	return memories
 		.filter(m => m.embedding?.length)
-		.map(m => ({ref: {name: m.name, description: m.description}, distance: cosineDistance(query, m.embedding)}))
+		.map(m => ({name: m.name, description: m.description, distance: cosineDistance(query, m.embedding)}))
 		.sort((a, b) => a.distance - b.distance)
-		.slice(0, limit)
+		.slice(0, limit);
-		.map(s => s.ref);
+}
+
+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 stripHeader(content: string): string {
+	return content.replace(/^---[\s\S]*?\n---\n?/, '').trimStart();
+}
+
+/** True if a task has no persistent entity of its own and belongs in the journal instead. */
+function isPersonalTask(t: MemoryTask): boolean {
+	const s = (t.subject ?? '').trim().toLowerCase();
+	return !s || s === 'journal' || s.startsWith('journal/');
 }
 
 export class MemoryCache {
 	private tree!: KDTree<MemoryRef>;
+	private indexed = new Map<string, number[]>();
 	public memories: Memory[];
 	public nodes: MemoryNode[] = [];
 
@@ -70,50 +102,62 @@ export class MemoryCache {
 
 	constructor(memories: Memory[]) {
 		this.memories = memories;
+		this.tree = new KDTree<MemoryRef>(0);
 		this.rebuild();
 	}
 
-	private buildTree(): KDTree<MemoryRef> {
+	private syncTree(): void {
-		const embedded = this.memories.filter(m => m.embedding?.length);
+		const current = new Set(this.memories.map(m => m.name));
-		if (!embedded.length) return new KDTree<MemoryRef>(0);
 
-		const dims = embedded[0].embedding.length;
+		for(const [name, emb] of [...this.indexed]) {
-		const points: KDPoint<MemoryRef>[] = embedded.map(m => ({
+			const mem = this.memories.find(m => m.name === name);
-			vector: m.embedding,
+			if(!mem || !current.has(name) || mem.embedding !== emb) {
-			payload: {name: m.name, description: m.description},
+				this.tree.remove(p => p.name === name);
-		}));
+				this.indexed.delete(name);
+			}
+		}
 
-		return new KDTree<MemoryRef>(dims, 'cosine', points);
+		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; // guard against embedding model/dim drift
+			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 [];
+		if(!this.tree || this.tree.dims === 0) return [];
-		return this.tree.knn(query, limit).map(r => r.point.payload);
+		return this.tree.knn(query, limit).map(r => ({...r.point.payload, distance: r.distance}));
 	}
 
 	add(memory: Memory): void {
 		this.memories.push(memory);
-		this.rebuild();
+		this.rebuild([memory]);
 	}
 
 	update(memory: Memory): void {
-		const idx = this.memories.findIndex(m => m.name === memory.name);
+		const existing = this.memories.find(m => m.name === memory.name);
-		if (idx !== -1) this.memories[idx] = memory;
+		if(existing) Object.assign(existing, memory);
 		else this.memories.push(memory);
-		this.rebuild();
+		this.rebuild([existing ?? memory]);
 	}
 
 	remove(name: string): void {
 		const idx = this.memories.findIndex(m => m.name === name);
-		if (idx !== -1) {
+		if(idx !== -1) {
 			this.memories.splice(idx, 1);
 			this.rebuild();
 		}
 	}
 
-	rebuild(): void {
+	rebuild(changed?: Memory[]): void {
-		this.nodes = rebuildGraph(this.memories);
+		this.nodes = (changed?.length && this.nodes.length)
-		this.tree = this.buildTree();
+			? patchGraph(this.memories, this.nodes, changed)
+			: rebuildGraph(this.memories);
+		this.syncTree();
 	}
 }
 
@@ -130,9 +174,9 @@ class MemoryAccessor {
 		return this.list.find(m => m.name === name);
 	}
 
-	commit(): MemoryNode[] {
+	commit(changed?: Memory[]): MemoryNode[] {
-		if (this.cache) {
+		if(this.cache) {
-			this.cache.rebuild();
+			this.cache.rebuild(changed);
 			return this.cache.nodes;
 		}
 		return rebuildGraph(this.list);
@@ -149,7 +193,7 @@ class MemoryAccessor {
 
 	forget(name: string): boolean {
 		const idx = this.list.findIndex(m => m.name === name);
-		if (idx === -1) return false;
+		if(idx === -1) return false;
 		this.list.splice(idx, 1);
 		this.commit();
 		return true;
@@ -157,11 +201,8 @@ class MemoryAccessor {
 
 	async backfillEmbeddings(llm: any): Promise<number> {
 		const missing = this.list.filter(m => !m.embedding?.length);
-		if (!missing.length) return 0;
+		if(!missing.length) return 0;
-		await Promise.all(missing.map(async node => {
+		await Promise.all(missing.map(node => embedMemoryFields(node, llm)));
-			const [e] = await llm.embedding(node.content);
-			if (e) node.embedding = e.embedding;
-		}));
 		this.commit();
 		return missing.length;
 	}
@@ -181,13 +222,13 @@ export type MemoryOptions = {
 }
 
 export class MemoryManager {
-	private recentlyTouched = new Map<string, number>();
+	private mergeLock: Promise<any> = Promise.resolve();
-
 	private queues = new Map<string, {
 		dirty: boolean,
 		request: {abort?: () => void} | null,
 		task: Promise<void>,
 	}>();
+	private recentlyTouched = new Map<string, number>();
 
 	tools = {
 		forget: (memories: Memory[] | MemoryCache): AiTool => ({
@@ -206,11 +247,11 @@ export class MemoryManager {
 			name: 'memory_recall',
 			description: 'Read the full content of a memory document',
 			args: {
-				name: {type: 'string', description: 'Exact memory name', required: true},
+				name: {type: 'string', description: 'Exact memory name', required: true}
 			},
 			fn: (args: any) => {
-				const mem = this.access(memories).find(args.name);
+				const mem = new MemoryAccessor(memories).find(args.name);
-				if (!mem) return 'Document not found';
+				if(!mem) return 'Document not found';
 				this.touch(mem.name);
 				return mem.content;
 			},
@@ -224,7 +265,7 @@ export class MemoryManager {
 				limit: {type: 'number', description: 'Number of memories to return', default: 1},
 			},
 			fn: async ({query, limit}) => {
-				const mem = await this.recollect(query, memories, limit)
+				const mem = await this.recollect(query, memories, limit);
 				return mem.map(m => `Memory: ${m.name}
 Description: ${m.description}
 Links: ${[...m.links, ...m.backlinks].join(', ')}
@@ -238,68 +279,121 @@ ${m.content}
 	constructor(private llm: any) {}
 
 	static normalize(m?: Memory[] | MemoryCache | MemoryOptions) {
-		if (!m) return null;
+		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 access(memories: Memory[] | MemoryCache): MemoryAccessor {
+	private stage(node: Memory, block: string): void {
-		return new MemoryAccessor(memories);
+		if(!node.content) {
-	}
+			const title = node.name.split('/').pop() ?? node.name;
-
+			node.content = this.touchHeader(node, `# ${title}\n`);
-	private appendFacts(node: Memory, facts: string[]): void {
+		}
-		this.ensureDoc(node);
+		const body = stripHeader(node.content);
-		const body = this.stripHeader(node.content);
+		const idx = body.indexOf(PENDING_HEADING);
-		const bullets = facts.map(f => `- ${f}`).join('\n');
-		const idx = body.indexOf(FACTS_HEADING);
 		const newBody = idx === -1
-			? `${body.trimEnd()}\n\n${FACTS_HEADING}\n${bullets}\n`
+			? `${body.trimEnd()}\n\n${PENDING_HEADING}\n${block}\n`
-			: `${body.slice(0, idx + FACTS_HEADING.length)}\n${bullets}${body.slice(idx + FACTS_HEADING.length)}`;
+			: `${body.slice(0, idx + PENDING_HEADING.length)}\n${block}${body.slice(idx + PENDING_HEADING.length)}`;
 		node.content = this.touchHeader(node, newBody);
 	}
 
-	private ensureDoc(node: Memory): void {
+	private resolveSubject(subject: string, store: MemoryAccessor): string {
-		if (node.content) return;
+		function normalize(name: string): string {
-		const title = node.name.split('/').pop() ?? node.name;
+			return name.trim().toLowerCase().replace(/\s+/g, ' ');
-		node.content = this.touchHeader(node, `# ${title}\n`);
+		}
+
+		const trimmed = subject.trim();
+		const exact = store.find(trimmed);
+		if(exact) return exact.name;
+
+		const normalized = normalize(trimmed);
+		const caseInsensitive = store.list.find(m => normalize(m.name) === normalized);
+		if(caseInsensitive) return caseInsensitive.name;
+
+		const root = trimmed.split('/')[0];
+		const leaf = trimmed.split('/').slice(1).join('/') || trimmed;
+		const candidates = store.list.filter(m => m.name.split('/')[0] === root && m.name !== trimmed);
+		if(!candidates.length) return trimmed;
+
+		const leaves = candidates.map(m => m.name.split('/').slice(1).join('/') || m.name);
+		const probe = leaves.length > 1 ? leaves : [...leaves, ''];
+		const {max, similarities} = this.llm.fuzzyMatch(leaf, ...probe);
+		if(max >= ALIAS_MATCH_THRESHOLD) return candidates[similarities.indexOf(max)].name;
+
+		return trimmed;
 	}
 
-	private async factAgent(conversation: string, store: MemoryAccessor, options: LLMRequest, weekKey: string): Promise<FactBucket[]> {
+	private async factAgent(conversation: string, store: MemoryAccessor, options: LLMRequest): Promise<FactAgentResult> {
 		const ghosts = store.ghosts();
 
 		const response = await this.llm.ask(conversation, {
 			model: options.model,
 			temperature: 0.2,
-			system: `You are a fact extractor to build obsidian knowledge vaults.
+			system: `Turn this conversation into a persistent memory file by extracting information into organized bullet points
-Analyze this conversation and extract facts worth remembering long-term.
 
-Rules:
+Think of this like an Obsidian vault with a clear division of responsibility:
-- Always extract facts that the user explicitly told you to remember
+- The JOURNAL is a timeline. It answers "what happened, and when" and is the only place with a sense of time.
-- ONLY extract current facts the USER explicitly stated about themselves, their work, projects or decisions that were MADE during this conversation
+- ENTITY DOSSIERS are a wiki. They answer "what is currently true about this subject", with no sense of time — only current state.
-- DO NOT extract greetings, pleasantries, or generic exchanges
+- Never blur the two: a one-off event, conversation, or debugging session is a journal entry, not an entity, even if it's detailed.
-- DO NOT extract deltas or changes in facts; ONLY the end fact
-- DO NOT extract anything the AI/assistant itself said
-- If nothing worth remembering was said, return an empty buckets array
 
-When extracting facts, you MUST also decide the exact destination path:
+1. Journal Log
-- Reuse node names (including ghost) as much as possible IF the facts belongs there
+- A chronological, skimmable log of what actually happened: real discussions, decisions made, progress on projects, problems worked through
-- All information primarily about the user should go under "People/User"
+- This is NOT a transcript, and it is NOT a step-by-step record, its a compressed log of notable events & developments
-- When required, create a new path following collection/subject format (e.g., People/Sarah, Projects/Oxide) — you are not limited to any fixed list of collections, use whatever fits
+- One line per development is usually enough: what was worked on and the outcome, not the blow-by-blow of how
-- For journal entries, use "Journal"
+- Skip small talk and trivial exchanges entirely. Skip anything that's purely a todo item (goes in Todo Tasks) or a durable fact about a subject (goes in Entity Dossiers)
+
+2. Todo Tasks
+- Extract concrete tasks the user says need to be done, should be done, or were completed
+- Return the task text and whether it is still todo or is done
+- A completed task should be marked done, not recreated as a new todo
+- Only extract actionable tasks, not general goals or observations
+- Assign each task a subject:
+    - If the task belongs to a persistent entity (a project, a class, etc.), use that entity's exact node name, or a new entity path if it doesn't exist yet
+    - If it's a personal/life task with no entity of its own (reach out to someone, reply to an email, pay a bill, etc.), leave subject as an empty string — it belongs in the journal, not a new document
+
+3. Entity Dossiers
+- Detailed dossiers with all information regarding a subject
+- Record the final/end state, not intermediate changes
+- Ignore assistant claims, guesses, greetings, or temporary details
+- NEVER create a document for something that's only meaningful as a point in time — a single conversation, a one-off decision, a debugging session, a date. That's a journal entry, not an entity
+- identify its HOME ENTITY:
+    - The HOME ENTITY name should always be a [abstract|pro]noun
+    - The grammatical subject/owner of the fact is the strongest clue
+    - Always preference an existing entity over creating a new one
+    - New child entities are appropriate only when they are themselves distinct persistent entities
+    - A document represents a persistent entity, not a topic, feature, bug, event, decision, setting, or conversation fragment
+    - Put project facts under the project they belong to, person facts under the person, etc
+
+Example Entity Naming Convention:
+- Projects/[Name]
+- People/[Name]
+- History/[Name]
+- Science/[Name]
+- [Subject]/[Name]
+- Class/[Name]/[Chapter]
+
+Use [[WikiLinks]] to express relationships between entities. NEVER create documents just to hold relationships
+Keep journal material in the journal; don't turn journal events into entities unless they represent something persistent
 
 Available nodes:
-- Journal
+${this.listNodes(store.list).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}
-${this.listNodes(store.list).filter(n => !n.name.includes('Journal')).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}
 ${ghosts.length ? `${ghosts.map(g => `- ${g}: (Ghost)`).join('\n')}` : ''}`,
 			schema: {
-				buckets: {type: 'array', description: 'Groups of facts to remember, each assigned to a different node. Return an empty array if there is nothing worth storing in an obsidian vault', items: {
+				journal: {type: 'string', description: 'Short day-to-day recap; empty if nothing happened.', required: false},
-					type: 'object', items: {
+				tasks: {
-							subject: {type: 'string', description: 'Exact existing node name OR new path (e.g. "People/Sarah", "Projects/Oxide"), or "Journal"', required: true},
+					type: 'array', description: 'Concrete tasks mentioned or completed in the conversation.', required: false, items: {
-							facts: {
+						type: 'object', items: {
-								type: 'array',
+							subject: {type: 'string', description: 'Exact node name / new persistent entity path this task belongs to, or an empty string if this is a personal task with no entity of its own (those go in the journal)', required: true},
-								description: 'Facts to store at this destination',
+							task: {type: 'string', description: 'Concise actionable task', required: true},
-								items: {type: 'string', description: 'A single fact'},
+							done: {type: 'boolean', description: 'Whether the task is completed', required: true},
-							},
+						},
+					}
+				},
+				buckets: {
+					type: 'array', description: 'Groups of facts to remember; empty array if nothing worth storing.', items: {
+						type: 'object', items: {
+							subject: {type: 'string', description: 'Exact node name or new persistent entity path', required: true},
+							facts: {type: 'array', description: 'Facts to store here', items: {type: 'string'}},
 						},
 					},
 				},
@@ -308,17 +402,20 @@ ${ghosts.length ? `${ghosts.map(g => `- ${g}: (Ghost)`).join('\n')}` : ''}`,
 
 		const buckets = new Map<string, string[]>();
 		for(const bucket of response.buckets ?? []) {
-			const subject = bucket.subject.trim().toLowerCase() === 'journal'
+			const subject = bucket.subject.trim();
-				? `Journal/${weekKey}` : bucket.subject.trim();
 			const facts = buckets.get(subject) ?? [];
 			facts.push(...dedupeFacts(bucket.facts));
 			buckets.set(subject, facts);
 		}
 
-		return buckets.entries().toArray().map(([subject, facts]) => ({subject, facts}));
+		return {
+			buckets: buckets.entries().toArray().map(([subject, facts]) => ({subject, facts})),
+			journal: (response.journal ?? '').trim(),
+			tasks: response.tasks ?? [],
+		};
 	}
 
-	private getWeekMonday(date: Date = new Date()): string {
+	private 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;
@@ -326,14 +423,92 @@ ${ghosts.length ? `${ghosts.map(g => `- ${g}: (Ghost)`).join('\n')}` : ''}`,
 		return d.toISOString().slice(0, 10);
 	}
 
+	private journalDescription(journalName?: string): string {
+		const start = journalName?.split('/').pop() || this.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} - ${end}`;
+	}
+
+	private getIncompleteTodos(content: string): string[] {
+		const body = stripHeader(content);
+		const match = body.match(/## Todo list\n([\s\S]*?)(?=\n## |$)/i);
+		if(!match) return [];
+		return match[1].split('\n')
+			.map(line => line.match(/^\s*-\s*\[([ xX])\]\s+(.+?)\s*$/))
+			.filter((m): m is RegExpMatchArray => !!m && m[1].toLowerCase() !== 'x')
+			.map(m => m[2].trim());
+	}
+
 	private listNodes(memories: Memory[]): MemoryRef[] {
 		return memories.map(m => ({name: m.name, description: m.description}));
 	}
 
+	private async mergeAgent(node: Memory, memories: Memory[] | MemoryCache, options: LLMRequest): Promise<Memory | null> {
+		function factSimilarity(a: Memory, b: Memory): number {
+			if(!a.bodyEmbeddings?.length || !b.bodyEmbeddings?.length) return 0;
+			let best = 0;
+			for(const av of a.bodyEmbeddings) {
+				for(const bv of b.bodyEmbeddings) best = Math.max(best, 1 - cosineDistance(av, bv));
+			}
+			return best;
+		}
+
+		if(!node.embedding?.length || node.name.startsWith('Journal/')) return null;
+		const store = new MemoryAccessor(memories);
+		const candidates = store.list
+			.filter(m => m.name !== node.name && !m.name.startsWith('Journal/'))
+			.filter(m => factSimilarity(node, m) >= FACT_SIMILARITY_THRESHOLD);
+
+		if(!candidates.length) return null;
+		const closest = candidates.sort((a, b) => factSimilarity(node, b) - factSimilarity(node, a))[0];
+		const result = await this.llm.ask('', {
+			model: options.model,
+			temperature: 0.3,
+			schema: {
+				aContent: {type: 'string', description: 'Updated document A body in markdown, without frontmatter.', required: true},
+				bContent: {type: 'string', description: 'Updated document B body in markdown, without frontmatter.', required: true},
+			},
+			system: `Maintain these two persistent knowledge-base documents like a wiki.
+
+Do NOT merge, rename, or delete either document. Both represent entities that should remain independently addressable.
+
+The documents were selected because their facts may overlap. Your job is to reconcile duplicated information and connect the documents:
+- Decide which document is the HOME for each duplicated fact.
+- Keep the authoritative copy in that home document.
+- In the other document, replace the information with a short preamble and [[WikiLink]] to the home entity explaining the relationship.
+- If the documents are distinct entities but merely related, keep their distinct facts and add useful [[WikiLinks]] between them.
+- Do not delete useful entity-specific facts just because they are similar.
+- Do not invent relationships or facts.
+- Preserve useful history, technical specifics, structure, and existing [[WikiLinks]].
+- Most current truth wins when facts conflict.
+- Keep both documents concise and information-dense.
+- No frontmatter, preamble, filler, or AI commentary.
+
+Document A ("${node.name}"):
+\`\`\`markdown
+${stripHeader(node.content)}
+\`\`\`
+
+Document B ("${closest.name}"):
+\`\`\`markdown
+${stripHeader(closest.content)}
+\`\`\``,
+		});
+		const a = store.find(node.name);
+		const b = store.find(closest.name);
+		if(!a || !b || !result?.aContent || !result?.bContent) return null;
+		a.content = this.touchHeader(a, result.aContent);
+		b.content = this.touchHeader(b, result.bContent);
+		await Promise.all([embedMemoryFields(a, this.llm), embedMemoryFields(b, this.llm)]);
+		return a;
+	}
+
 	private reconcile(node: Memory, memories: Memory[] | MemoryCache, options: LLMRequest): Promise<void> {
 		const key = node.name;
 		const existing = this.queues.get(key);
-		if (existing) {
+		if(existing) {
 			existing.dirty = true;
 			existing.request?.abort?.();
 			return existing.task;
@@ -341,105 +516,122 @@ ${ghosts.length ? `${ghosts.map(g => `- ${g}: (Ghost)`).join('\n')}` : ''}`,
 
 		const entry = {dirty: false, request: null, task: Promise.resolve()};
 		this.queues.set(key, entry);
-		const store = this.access(memories);
+		const store = new MemoryAccessor(memories);
 		entry.task = (async () => {
-			do {
+			let current = node;
-				entry.dirty = false;
+			try {
-				await this.docAgent(node, store.list, options, entry);
+				do {
-			} while (entry.dirty);
+					entry.dirty = false;
-		})().finally(() => {
+					await this.docAgent(current, store.list, options, entry);
-			this.queues.delete(key);
+					this.mergeLock = this.mergeLock.then(() => this.mergeAgent(current, memories, options));
-			store.commit();
+					const result = await this.mergeLock;
-		});
+					if(result) current = result;
+				} while(entry.dirty);
+			} finally {
+				store.commit([node]);
+				this.queues.delete(key);
+			}
+		})();
 		return entry.task;
 	}
 
 	private async docAgent(node: Memory, memories: Memory[], options: LLMRequest, entry: {request: {abort?: () => void} | null}): Promise<void> {
-		const currentBody = this.stripHeader(node.content);
+		if(!memories.includes(node)) return;
-		let update;
+		const currentBody = stripHeader(node.content);
-		try {
+		const journal = node.name.startsWith('Journal/');
-			for (let i = 0; i < 2 && !update?.content; i++) {
+		const system = (journal
-				const request = this.llm.ask(currentBody, {
+			? `You maintain one persistent journal document
-					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 body in markdown, without the frontmatter block', required: true},
-					},
-					system: `You are a knowledge base editor maintaining one document in an Obsidian-style vault.
 
-If the document has a "${FACTS_HEADING}" section, integrate every bullet under it into the appropriate part of the document, then remove the "${FACTS_HEADING}" section entirely. If there is no such section, just tidy the document per the rules below.
+Rewrite the ENTIRE journal, folding "## Pending" into the existing content removing the heading
 
-Structure: follow this generic shape loosely, adapting section names/order to what the content actually needs (e.g. journal-style docs may want a timeline instead of "Details"):
+Journal design:
-\`\`\`markdown
+- Preserve the chronological daily log
-${GENERIC_TEMPLATE}
+- Maintain a single \`## Todo list\` section for this entity: reconcile tasks semantically (merge equivalent tasks, remove duplicates, preserve incomplete tasks, check off completed ones), and keep it distinct from the narrative/fact sections
-\`\`\`
+- Group information by day under a date heading
+- Keep journal entries high level and concise: what was worked on and the outcome, not a step-by-step record of how — that detail lives in conversation history, not here
+- Use [[WikiLinks]] for persistent entities; don't turn ordinary journal events into entities
+- No frontmatter, preamble, filler, or AI commentary`
+			: `You maintain one persistent knowledge-base entity document
 
-Formatting rules:
+Rewrite the ENTIRE document, folding "## Pending" into the existing content. Remove the Pending section when finished.
-- Use Obsidian-style markdown: # headings, **bold** for emphasis, bullet & numbered lists for grouped 1D data, 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, but skip generics (car, red, dog)
-- Keep the document concise, factual, and human-readable
-- Resolve contradictions: newer facts always win — delete the outdated statement entirely, never keep both
-- Do not add frontmatter blocks, filler, preamble, or AI commentary
 
-Other nodes in the vault (link to these instead of duplicating their content):
+Document design:
+- The document represents one persistent entity. Keep information about that entity together and organized into sections
+- Merge any pending information in, newest fact wins conflicts; remove redundant content
+- Maintain a single \`## Todo list\` section for this entity: reconcile tasks semantically (merge equivalent tasks, remove duplicates, preserve incomplete tasks, check off completed ones), and keep it distinct from the narrative/fact sections
+- Let the structure fit the entity; there is NO fixed template
+- Add headings only when they meaningfully organize recurring information; don't create headings for one-off facts
+- Keep the document concise and information-dense without removing useful technical specifics
+- Current truth wins when facts conflict. Preserve older conflict as context, only when it adds useful meaning
+- No frontmatter, preamble, filler, or AI commentary`) + `
+
+Available nodes to link to:
 ${this.listNodes(memories).filter(n => n.name !== node.name).map(n => n.name).join(', ') || 'none'}
 
 Current document:
 \`\`\`markdown
 ${currentBody}
-\`\`\``,
+\`\`\``;
+		let update;
+		try {
+			for(let i = 0; i < 2 && !update?.content; i++) {
+				const request = this.llm.ask(currentBody, {
+					model: options.model,
+					temperature: 0.3,
+					schema: {
+						description: {type: 'string', description: 'One factual sentence describing the document\'s ENTIRE SUBJECT MATTER — for use as a search/merge fingerprint', required: true},
+						content: {type: 'string', description: 'Rewritten document body in markdown, without the frontmatter block', required: true},
+					},
+					system,
 				});
 				entry.request = request;
 				update = await request;
 			}
-		} catch (err: any) {
+		} catch(err: any) {
-			if (err?.name === 'AbortError') return;
+			if(err?.name === 'AbortError') return;
 			throw err;
 		} finally {
 			entry.request = null;
 		}
 
-		if (!update?.content) return;
+		if(!update?.content) return;
-		node.description = node.name !== 'People/User' ? update.description : 'All information about the current user';
+		node.description = node.name.startsWith('Journal/') ? this.journalDescription(node.name) : node.name !== 'People/User' ? update.description.replaceAll(/[\n:]/g, '') : 'All information about the current user';
 		node.content = this.touchHeader(node, update.content);
-		const [e] = await this.llm.embedding(node.content);
+		await embedMemoryFields(node, this.llm);
-		if (e) node.embedding = e.embedding;
 	}
 
 	private 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};
+		if(!match) return {fm: new Map(), body: content};
 		const fm = new Map<string, string>();
-		for (const line of match[1].split('\n')) {
+		for(const line of match[1].split('\n')) {
 			const i = line.indexOf(':');
-			if (i === -1) continue;
+			if(i === -1) continue;
-			fm.set(line.slice(0, i).trim(), line.slice(i + 1).trim());
+			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]};
 	}
 
-	private stripHeader(content: string): string {
-		return content.replace(/^---[\s\S]*?\n---\n?/, '').trimStart();
-	}
-
 	private touchHeader(node: Memory, body: string): string {
 		const {fm} = this.parseFrontmatter(node.content);
 		fm.set('name', node.name);
-		fm.set('description', node.description || '');
+		fm.set('description', (node.name.startsWith('Journal/') ? this.journalDescription(node.name) : node.description) || 'Persistent memory document');
 		fm.set('modified', new Date().toISOString());
-		return this.writeFrontmatter(fm, body);
+		return this.writeFrontmatter(fm, stripHeader(body));
 	}
 
 	private writeFrontmatter(fm: Map<string, string>, body: string): string {
-		const lines = [...fm.entries()].map(([k, v]) => `${k}: ${v}`);
+		const lines = [...fm.entries()].map(([k, v]) => `${k}: ${JSON.stringify(String(v).replace(/\s+/g, ' ').trim())}`);
 		return `---\n${lines.join('\n')}\n---\n\n${body.trimStart()}`;
 	}
 
 	decay() {
-		for (const [name, ttl] of this.recentlyTouched) {
+		for(const [name, ttl] of this.recentlyTouched) {
-			if (ttl <= 1) this.recentlyTouched.delete(name);
+			if(ttl <= 1) this.recentlyTouched.delete(name);
 			else this.recentlyTouched.set(name, ttl - 1);
 		}
 	}
@@ -449,30 +641,43 @@ ${currentBody}
 	}
 
 	forget(name: string, memories: Memory[] | MemoryCache): boolean {
-		return this.access(memories).forget(name);
+		return new MemoryAccessor(memories).forget(name);
 	}
 
 	async recollect(query: string, memories: Memory[] | MemoryCache, limit = 5, graphDepth = 1): Promise<Memory[]> {
-		const store = this.access(memories);
+		function rank(query: number[], candidates: Memory[], limit: number): Memory[] {
-		if (!store.list.length) return [];
+			const scored = candidates.map(m => {
+				const titleSim = m.titleEmbedding?.length ? 1 - cosineDistance(query, m.titleEmbedding) : 0;
+				const descSim = m.embedding?.length ? 1 - cosineDistance(query, m.embedding) : 0;
+				const bodySim = m.bodyEmbeddings?.length
+					? Math.max(...m.bodyEmbeddings.map(b => 1 - cosineDistance(query, b)))
+					: 0;
+				return {memory: m, score: titleSim * 0.5 + descSim * 0.35 + bodySim * 0.15};
+			});
+			return scored.sort((a, b) => b.score - a.score).slice(0, limit).map(s => s.memory);
+		}
 
+		const store = new MemoryAccessor(memories);
+		if(!store.list.length) return [];
 		await store.backfillEmbeddings(this.llm);
 
 		const [e] = await this.llm.embedding(query);
-		if (!e) return [];
+		if(!e) return [];
 
-		const vectorResults = store.search(e.embedding, limit);
+		const pool = store.search(e.embedding, Math.max(limit * 3, limit));
-		const found = new Set<string>(vectorResults.map(r => r.name));
+		const poolMemories = pool.map(r => store.find(r.name)).filter((m): m is Memory => !!m);
+		const ranked = rank(e.embedding, poolMemories, limit);
+		const found = new Set<string>(ranked.map(m => m.name));
 
-		if (graphDepth > 0) {
+		if(graphDepth > 0) {
 			let frontier = [...found];
-			for (let depth = 0; depth < graphDepth && frontier.length; depth++) {
+			for(let depth = 0; depth < graphDepth && frontier.length; depth++) {
 				const next: string[] = [];
-				for (const name of frontier) {
+				for(const name of frontier) {
 					const node = store.find(name);
-					if (!node) continue;
+					if(!node) continue;
-					for (const link of node.links) {
+					for(const link of node.links) {
-						if (!found.has(link) && store.find(link)) {
+						if(!found.has(link) && store.find(link)) {
 							found.add(link);
 							next.push(link);
 						}
@@ -482,42 +687,96 @@ ${currentBody}
 			}
 		}
 
-		const vectorOrder = vectorResults.map(r => r.name);
+		const rankedOrder = ranked.map(m => m.name);
-		const graphExpansions = [...found].filter(n => !vectorOrder.includes(n));
+		const graphExpansions = [...found].filter(n => !rankedOrder.includes(n));
-		return [...vectorOrder, ...graphExpansions].map(n => store.find(n)!).filter(Boolean);
+		return [...rankedOrder, ...graphExpansions].map(n => store.find(n)!).filter(Boolean);
 	}
 
 	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 [];
+		if(!conversation) return [];
 
 		const uid = `${Date.now()}_${Math.random().toString(36).slice(2)}`;
 		const pending = {role: 'tool', name: 'memory_process', id: uid, content: conversation} as unknown as LLMMessage;
 		history.push(pending);
 
-		const store = this.access(memories);
+		const store = new MemoryAccessor(memories);
-		const buckets = await this.factAgent(conversation, store, options, this.getWeekMonday());
+		const {buckets, journal, tasks} = await this.factAgent(conversation, store, options);
 		const touched: Memory[] = [];
 
-		for (const {subject, facts} of buckets) {
+		const personalTasks = tasks.filter(isPersonalTask);
-			let node = store.find(subject);
+		const entityTasks = tasks.filter(t => !isPersonalTask(t));
-			if (!node) {
+
-				node = {name: subject, description: '', content: '', embedding: [], links: [], backlinks: []};
+		if(journal || personalTasks.length) {
+			const journalName = `Journal/${this.getWeekStart()}`;
+			let jnode = store.find(journalName);
+			const isNew = !jnode;
+			if(!jnode) {
+				jnode = {
+					name: journalName,
+					description: this.journalDescription(),
+					content: '',
+					embedding: [],
+					links: [],
+					backlinks: [],
+				};
+				store.list.push(jnode);
+			}
+
+			const blocks: string[] = [];
+			if(journal) blocks.push(`### ${new Date().toISOString().slice(0, 10)}\n${journal}`);
+			if(isNew) {
+				const previousDate = new Date(`${this.getWeekStart()}T00:00:00Z`);
+				previousDate.setUTCDate(previousDate.getUTCDate() - 7);
+				const previous = store.find(`Journal/${previousDate.toISOString().slice(0, 10)}`);
+				if(previous) {
+					const todos = this.getIncompleteTodos(previous.content);
+					if(todos.length) blocks.push(`${TODO_HEADING}\n${todos.map(task => `- [ ] ${task}`).join('\n')}`);
+				}
+			}
+			if(personalTasks.length) blocks.push(`${TODO_HEADING}\n${personalTasks.map(task => `- [${task.done ? 'x' : ' '}] ${task.task}`).join('\n')}`);
+			if(blocks.length) this.stage(jnode, blocks.join('\n\n'));
+			touched.push(jnode);
+		}
+
+		const entityStaging = new Map<string, {facts: string[], tasks: MemoryTask[]}>();
+		for(const {subject, facts} of buckets) {
+			const resolved = this.resolveSubject(subject, store);
+			const entry = entityStaging.get(resolved) ?? {facts: [], tasks: []};
+			entry.facts.push(...facts);
+			entityStaging.set(resolved, entry);
+		}
+		for(const task of entityTasks) {
+			const resolved = this.resolveSubject(task.subject, store);
+			const entry = entityStaging.get(resolved) ?? {facts: [], tasks: []};
+			entry.tasks.push(task);
+			entityStaging.set(resolved, entry);
+		}
+
+		for(const [resolved, {facts, tasks: subjectTasks}] of entityStaging) {
+			let node = store.find(resolved);
+			if(!node) {
+				node = {name: resolved, description: 'Persistent memory document', content: '', embedding: [], links: [], backlinks: []};
 				store.list.push(node);
 			}
-			this.appendFacts(node, facts);
+			const blocks: string[] = [];
-			const [e] = await this.llm.embedding(node.content);
+			if(facts.length) blocks.push(facts.map(f => `- ${f}`).join('\n'));
-			if (e) node.embedding = e.embedding;
+			if(subjectTasks.length) blocks.push(`${TODO_HEADING}\n${subjectTasks.map(t => `- [${t.done ? 'x' : ' '}] ${t.task}`).join('\n')}`);
-			this.touch(node.name);
+			if(blocks.length) this.stage(node, blocks.join('\n\n'));
 			touched.push(node);
 		}
 
-		if (touched.length) {
+		await Promise.all(touched.map(async node => {
-			store.commit();
+			await embedMemoryFields(node, this.llm);
+			this.touch(node.name);
+		}));
+
+		if(touched.length) {
+			store.commit(touched);
 			(pending as any).content = `Saved to ${touched.map(n => `[[${n.name}]]`).join(', ')}`;
-			await Promise.all(touched.map(node => this.reconcile(node, memories, options).catch(() => {})));
+			Promise.all(touched.map(node => this.reconcile(node, memories, options).catch(() => {})));
 		} else {
 			(pending as any).content = 'Nothing worth remembering.';
 		}
@@ -526,9 +785,9 @@ ${currentBody}
 		return touched;
 	}
 
-	async reconcileVault(memories: Memory[] | MemoryCache, options: LLMRequest, scope: 'touched' | 'all' = 'touched'): Promise<void> {
+	async reconcileAll(memories: Memory[] | MemoryCache, options: LLMRequest, scope: 'touched' | 'all' = 'touched'): Promise<void> {
-		const store = this.access(memories);
+		const store = new MemoryAccessor(memories);
-		const targets = scope === 'all' ? store.list : store.list.filter(m => m.content.includes(FACTS_HEADING));
+		const targets = scope === 'all' ? store.list : store.list.filter(m => m.content.includes(PENDING_HEADING));
 		await Promise.all(targets.map(node => this.reconcile(node, memories, options)));
 		store.commit();
 	}

src/open-ai.ts

@@ -25,25 +25,60 @@ export class OpenAi extends LLMProvider {
 		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(const h of history) {
+
-			if(h.role === 'tool') {
+		for(let i = 0; i < history.length; i++) {
-				wire.push({
+			const h = history[i];
-					role: 'assistant',
+
-					content: null,
+			if(h.role !== 'tool') {
-					tool_calls: [{id: h.id, type: 'function', function: {name: h.name, arguments: JSON.stringify(h.args)}}],
+				wire.push({role: h.role, content: this.toWireContent(h.content)});
-				}, {
+				continue;
-					role: 'tool',
-					tool_call_id: h.id,
-					content: h.error || h.content || '',
-				});
-			} else {
-				wire.push({role: h.role, content: h.content});
 			}
+
+			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;
 	}
 
@@ -53,13 +88,12 @@ export class OpenAi extends LLMProvider {
 			if(!options.history) options.history = [];
 			const history = options.history;
 			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,
 				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: {
@@ -67,8 +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])
+								: []
 						}
 					}
 				}))
@@ -76,60 +114,106 @@ export class OpenAi extends LLMProvider {
 
 			if(options.schema) {
 				const schema = convertSchema(options.schema);
-				requestParams.response_format = {type: 'json_schema', json_schema: {name: 'response', strict: true, schema}};
+				requestParams.response_format = {
+					type: 'json_schema',
+					json_schema: {name: 'response', strict: true, schema}
+				};
 			}
 			if(options.stream) requestParams.stream_options = {include_usage: true};
 
 			try {
 				let terminal = false;
+				let iteration = 0;
+
 				do {
+					iteration++;
 					requestParams.messages = this.toWire(history.filter(h => h.role !== 'system'), options.system);
 
 					const callStart = Date.now();
-					const resp: any = await this.tokenPool.run(token => this.getClient(token).chat.completions.create(requestParams)).catch(err => {
+					const resp: any = await this.tokenPool.run(token =>
+						this.getClient(token).chat.completions.create(requestParams)
+					).catch(err => {
 						err.message += `\n\nMessages:\n${JSON.stringify(requestParams.messages, null, 2)}`;
 						throw err;
 					});
 
-					let usage: any, msg: any = {content: '', tool_calls: []};
+					let usage: any;
+					let finishReason: string | undefined;
+					let msg: any = {content: '', tool_calls: []};
+					let streamedChars = 0;
+
 					if(options.stream) {
-						for await (const chunk of resp) {
+						let streamCompleted = false;
-							if(controller.signal.aborted) break;
+						try {
-							if(chunk.usage) usage = chunk.usage;
+							for await (const chunk of resp) {
-							if(chunk.choices[0]?.delta?.content) {
+								if(controller.signal.aborted) break;
-								msg.content += chunk.choices[0].delta.content;
+								if(chunk.usage) usage = chunk.usage;
-								options.stream({text: chunk.choices[0].delta.content});
+
-							}
+								const choice = chunk.choices?.[0];
-							if(chunk.choices[0]?.delta?.tool_calls) {
+								if(choice?.finish_reason) finishReason = choice.finish_reason;
-								for(const deltaTC of chunk.choices[0].delta.tool_calls) {
+
-									const existing = msg.tool_calls.find((tc: any) => tc.index === deltaTC.index);
+								if(choice?.delta?.content) {
-									if(existing) {
+									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;
-									} else {
-										msg.tool_calls.push({
-											index: deltaTC.index,
-											id: deltaTC.id || '',
-											function: {name: deltaTC.function?.name || '', 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;
 					}
+
 					const duration = Date.now() - callStart;
 					const tps = usage?.completion_tokens && duration > 0 ? usage.completion_tokens / (duration / 1000) : 0;
 
+					if(finishReason === 'length' && !controller.signal.aborted) {
+						if(msg.content?.trim()) history.push({role: 'assistant', content: msg.content.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()};
+							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};
 						});
@@ -137,12 +221,13 @@ export class OpenAi extends LLMProvider {
 						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) { entry.error = 'Tool not found'; return; }
+							if(!tool) return entry.error = 'Tool not found';
 							try {
 								const toolStream = options.stream && ((chunk: any) => {
-									if(chunk.done) { terminal = true; return; }
+									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) {
@@ -157,7 +242,6 @@ export class OpenAi extends LLMProvider {
 				} 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);