More memory optimizations

package-lock.json

@@ -1,19 +1,19 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "1.5.0",
+	"version": "1.6.6",
 	"lockfileVersion": 3,
 	"requires": true,
 	"packages": {
 		"": {
 			"name": "@ztimson/ai-utils",
-			"version": "1.5.0",
+			"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",
@@ -1525,9 +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"

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "1.6.1",
+	"version": "1.7.2",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -29,7 +29,7 @@
 		"@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",

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/index.ts

@@ -1,11 +1,14 @@
 export * from './ai';
 export * from './antrhopic';
 export * from './audio';
-export * from './helpers';
 export * from './llm';
-export * from './memory';
+export * from './memory/graph';
+export * from './memory/kd-tree';
+export * from './memory/memory';
+export * from './memory/memory-state';
 export * from './open-ai';
 export * from './provider';
 export * from './token-pool'
 export * from './tools';
 export * from './vision';
+export * from './utils';

src/llm.ts

@@ -1,17 +1,19 @@
 import {clean, makeUnique, snakeCase} from '@ztimson/utils';
 import {AbortablePromise, Ai} from './ai.ts';
 import {Anthropic} from './antrhopic.ts';
+import {MemoryCache} from './memory/memory-state.ts';
+import {Memory, MemoryManager, MemoryOptions} from './memory/memory.ts';
 import {OpenAi} from './open-ai.ts';
 import {LLMProvider} from './provider.ts';
 import {AiTool, AiToolArg} from './tools.ts';
 import {fileURLToPath} from 'url';
 import {spawn} from 'node:child_process';
-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';
+import {stripHeader} from './utils.ts';
 
 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
@@ -19,6 +21,13 @@ const PDF_OCR_PAGE_THRESHOLD = 12; // above this many pages, OCR scanned pages i
 export type AnthropicConfig = {proto: 'anthropic', token: string | string[]};
 export type OpenAiConfig = {proto: 'openai', host?: string, token: string | string[]};
 
+export type AgentRef = {
+	name: string;
+	description?: string;
+	delegate?: boolean;
+	fn: () => Agent | null | Promise<Agent | null>;
+}
+
 export type Agent = {
 	name: string;
 	description?: string;
@@ -29,7 +38,7 @@ export type Agent = {
 	skills?: Skill[] | null;
 	tools?: AiTool[] | null;
 	mcp?: McpServer[] | null;
-	agents?: string[] | null;
+	agents?: AgentRef[] | null;
 }
 
 export type LLMFile = {
@@ -106,8 +115,8 @@ export type LLMRequest = {
 	skills?: Skill[];
 	/** MCP servers to connect and expose as tools */
 	mcp?: McpServer[];
-	/** Subagents exposed as delegatable/wrapped tools */
+	/** Subagents exposed as delegatable/wrapped tools, resolved lazily via their `fn` */
-	agents?: Agent[];
+	agents?: AgentRef[];
 	/** Attach files to request */
 	files?: LLMFile[];
 	/** @internal recursion guard for nested agent delegation */
@@ -243,10 +252,8 @@ class LLM {
 			} else if(isText) {
 				text = (await this.loadBuffer(file, true)).toString('utf-8');
 			} else {
-				text = `Unsupported file type: ${ext || mime}`;
+				text = typeof file.content === 'string' ? file.content : `[Binary file, unable to extract: ${name}]`;
 			}
-
-			// Cache result, skip re-extraction on future turns of the same conversation
 			file.content = text;
 			file.extracted = true;
 			delete file.path;
@@ -267,22 +274,21 @@ class LLM {
 		};
 	}
 
-	private setupAgent(agents: Agent[] = [], allAgents: Agent[], history: LLMMessage[], aborts: (() => void)[], depth = 0, delegateState: {resp: string | null}): AiTool[] {
+	private setupAgent(stubs: AgentRef[] = [], history: LLMMessage[], aborts: ((keep?: boolean) => void)[], depth = 0, delegateState: {resp: string | null}): AiTool[] {
-		return agents.map(a => {
+		return stubs.map(stub => {
-			const toolName = `${a.delegate ? '' : 'sub'}agent_${snakeCase(a.name)}`;
+			const toolName = `${stub.delegate ? '' : 'sub'}agent_${snakeCase(stub.name)}`;
 			return {
 				name: toolName,
-				description: `${a.delegate ? 'Delegate to ' : ''}Subagent: ${a.description || a.name}`,
+				description: `${stub.delegate ? 'Delegate to ' : ''}Subagent: ${stub.description || stub.name}`,
 				args: clean<any>({
-					context: !a.delegate ? {type: 'string', description: 'Summary of related messages, samples, files, etc...', required: true} : undefined,
+					context: !stub.delegate ? {type: 'string', description: 'Summary of related messages, samples, files, etc...', required: true} : undefined,
 					instructions: {type: 'string', description: 'Detailed instructions for subagent to complete', required: true},
 				}),
 				fn: async (args: any, stream: any, ai: any, id?: string) => {
 					if(depth >= MAX_AGENT_DEPTH) return 'Max agent delegation depth exceeded';
 
-					const nested = (a.agents || [])
+					const a = await stub.fn();
-						.map(name => allAgents.find(x => x.name === name))
+					if(!a) return `Agent "${stub.name}" could not be resolved`;
-						.filter((x): x is Agent => !!x && x.name !== a.name);
 
 					const q = a.delegate ? '' : `${args.instructions}${args.context ? `\n\n<context>${args.context}</context>` : ''}`;
 
@@ -299,7 +305,7 @@ ${a.system}`,
 						mcp: a.mcp || undefined,
 						skills: a.skills || undefined,
 						tools: a.tools || undefined,
-						agents: nested,
+						agents: a.agents || [],
 						_agentDepth: depth + 1,
 					} as any);
 					aborts.push(request.abort);
@@ -399,11 +405,13 @@ ${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;
@@ -413,7 +421,24 @@ ${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;
@@ -434,7 +459,7 @@ ${a.system}`,
 			// Agents
 			const agents = options.agents || this.ai.options?.llm?.agents;
 			const delegateState: {resp: string | null} = {resp: null};
-			if(agents?.length) tools.push(...this.setupAgent(agents, agents, history, nestedAborts, options._agentDepth || 0, delegateState));
+			if(agents?.length) tools.push(...this.setupAgent(agents, history, nestedAborts, options._agentDepth || 0, delegateState));
 
 			// Memory
 			const mem = MemoryManager.normalize(options.memory);
@@ -442,8 +467,8 @@ ${a.system}`,
 				const mems = mem.memory instanceof MemoryCache ? mem.memory.memories : mem.memory;
 				if(mems.length) {
 					if(mem.inject) {
-						const pool = 15; // candidates considered, cheap since only refs are listed
+						const pool = 15;
-						const budget = mem.maxTokens ?? 2000; // actual content injected
+						const budget = mem.maxTokens ?? 2000;
 						const relevant = await this.memoryManager.recollect(message, mem.memory, pool);
 
 						let used = 0;
@@ -478,21 +503,22 @@ Description: ${r.description}
 Linked: ${makeUnique([...r.links, ...r.backlinks]).join(', ')}
 <!-- Truncated -->`).join('\n\n') : ''}`.trim())
 					}
-					if(mem.tool) tools.push(this.memoryManager.tools.read(mem.memory));
+					if(mem.tool) tools.push(...this.memoryManager.tools.read(mem.memory));
 				}
 			}
 
-			if(aborted) throw Object.assign(new Error('Aborted'), {name: 'AbortError'});
+			if(aborted) abortNow();
 
-			// Files
-			const files = options.files || [];
 			const lastMsg = history[history.length - 1];
-			const originalContent = lastMsg?.content;
+			if(files.length && lastMsg?.role === 'user') lastMsg.files = files;
-			if(files.length && lastMsg?.role === 'user') {
+			const restores: {msg: LLMMessage, content: any}[] = [];
-				lastMsg.files = files;
+			for(const msg of history) {
-				const {text, images} = await this.resolveFiles(files);
+				if(msg.role !== 'user' || !msg.files?.length) continue;
-				const merged = text ? `${originalContent}\n\n${text}` : originalContent;
+				const {text, images} = await this.resolveFiles(msg.files);
-				lastMsg.content = images.length
+				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;
 			}
@@ -500,13 +526,20 @@ Linked: ${makeUnique([...r.links, ...r.backlinks]).join(', ')}
 			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;
+			}
 
-			if(files.length && lastMsg?.role === 'user') lastMsg.content = originalContent;
+			// Strip the file injection shim
+			restores.forEach(({msg, content}) => msg.content = content);
 
 			// Capture meta (duration / tps)
 			for(const h of history) {
@@ -563,24 +596,6 @@ Linked: ${makeUnique([...r.links, ...r.backlinks]).join(', ')}
 		return h;
 	}
 
-	/**
-	 * Compare the difference between embeddings (calculates the angle between two vectors)
-	 * @param {number[]} v1 First embedding / vector comparison
-	 * @param {number[]} v2 Second embedding / vector for comparison
-	 * @returns {number} Similarity values 0-1: 0 = unique, 1 = identical
-	 */
-	cosineSimilarity(v1: number[], v2: number[]): number {
-		if (v1.length !== v2.length) throw new Error('Vectors must be same length');
-		let dotProduct = 0, normA = 0, normB = 0;
-		for (let i = 0; i < v1.length; i++) {
-			dotProduct += v1[i] * v2[i];
-			normA += v1[i] * v1[i];
-			normB += v2[i] * v2[i];
-		}
-		const denominator = Math.sqrt(normA) * Math.sqrt(normB);
-		return denominator === 0 ? 0 : dotProduct / denominator;
-	}
-
 	/**
 	 * Chunk text into parts for AI digestion
 	 * @param {object | string} target Item that will be chunked (objects get converted)

src/memory.ts

@@ -1,535 +0,0 @@
-import {MemoryNode, rebuildGraph} from './helpers.ts';
-import {LLMRequest, LLMMessage} from './llm.ts';
-import {AiTool} from './tools.ts';
-import {KDPoint, KDTree} from './kd-tree.ts';
-
-const FACTS_HEADING = '## Facts';
-
-const GENERIC_TEMPLATE = `# {{Title}}
-
-## Summary
-
-## Details
-
-## Related`;
-
-export type Memory = {
-	name: string;
-	description: string;
-	content: string;
-	embedding: number[];
-	links: string[];
-	backlinks: string[];
-}
-
-type MemoryRef = {
-	name: string;
-	description: string;
-}
-
-type FactBucket = {
-	subject: string;
-	facts: string[];
-}
-
-function dedupeFacts(facts: string[]): string[] {
-	const seen = new Map<string, string>();
-	for (const f of facts) {
-		const clean = f.trim();
-		if (clean) seen.set(clean.toLowerCase(), clean);
-	}
-	return [...seen.values()];
-}
-
-function cosineDistance(a: number[], b: number[]): number {
-	let dot = 0, normA = 0, normB = 0;
-	for (let i = 0; i < a.length; i++) {
-		dot   += a[i] * b[i];
-		normA += a[i] * a[i];
-		normB += b[i] * b[i];
-	}
-	const denom = Math.sqrt(normA) * Math.sqrt(normB);
-	return denom === 0 ? 1 : 1 - dot / denom;
-}
-
-function 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)}))
-		.sort((a, b) => a.distance - b.distance)
-		.slice(0, limit)
-		.map(s => s.ref);
-}
-
-export function stripHeader(content: string): string {
-	return content.replace(/^---[\s\S]*?\n---\n?/, '').trimStart();
-}
-
-export class MemoryCache {
-	private tree!: KDTree<MemoryRef>;
-	public memories: Memory[];
-	public nodes: MemoryNode[] = [];
-
-	get length() { return this.memories.length; }
-
-	constructor(memories: Memory[]) {
-		this.memories = memories;
-		this.rebuild();
-	}
-
-	private buildTree(): KDTree<MemoryRef> {
-		const embedded = this.memories.filter(m => m.embedding?.length);
-		if (!embedded.length) return new KDTree<MemoryRef>(0);
-
-		const dims = embedded[0].embedding.length;
-		const points: KDPoint<MemoryRef>[] = embedded.map(m => ({
-			vector: m.embedding,
-			payload: {name: m.name, description: m.description},
-		}));
-
-		return new KDTree<MemoryRef>(dims, 'cosine', points);
-	}
-
-	search(query: number[], limit: number): MemoryRef[] {
-		if (!this.tree || this.tree.dims === 0) return [];
-		return this.tree.knn(query, limit).map(r => r.point.payload);
-	}
-
-	add(memory: Memory): void {
-		this.memories.push(memory);
-		this.rebuild();
-	}
-
-	update(memory: Memory): void {
-		const idx = this.memories.findIndex(m => m.name === memory.name);
-		if (idx !== -1) this.memories[idx] = memory;
-		else this.memories.push(memory);
-		this.rebuild();
-	}
-
-	remove(name: string): void {
-		const idx = this.memories.findIndex(m => m.name === name);
-		if (idx !== -1) {
-			this.memories.splice(idx, 1);
-			this.rebuild();
-		}
-	}
-
-	rebuild(): void {
-		this.nodes = rebuildGraph(this.memories);
-		this.tree = this.buildTree();
-	}
-}
-
-class MemoryAccessor {
-	readonly list: Memory[];
-	private readonly cache: MemoryCache | null;
-
-	constructor(memories: Memory[] | MemoryCache) {
-		this.cache = memories instanceof MemoryCache ? memories : null;
-		this.list = this.cache ? this.cache.memories : <Memory[]>memories;
-	}
-
-	find(name: string): Memory | undefined {
-		return this.list.find(m => m.name === name);
-	}
-
-	commit(): MemoryNode[] {
-		if (this.cache) {
-			this.cache.rebuild();
-			return this.cache.nodes;
-		}
-		return rebuildGraph(this.list);
-	}
-
-	ghosts(): string[] {
-		const nodes = this.cache ? this.cache.nodes : rebuildGraph(this.list);
-		return nodes.filter(n => n.missing).map(n => n.name);
-	}
-
-	search(vector: number[], limit: number): MemoryRef[] {
-		return this.cache ? this.cache.search(vector, limit) : cosineSearch(vector, this.list, limit);
-	}
-
-	forget(name: string): boolean {
-		const idx = this.list.findIndex(m => m.name === name);
-		if (idx === -1) return false;
-		this.list.splice(idx, 1);
-		this.commit();
-		return true;
-	}
-
-	async backfillEmbeddings(llm: any): Promise<number> {
-		const missing = this.list.filter(m => !m.embedding?.length);
-		if (!missing.length) return 0;
-		await Promise.all(missing.map(async node => {
-			const [e] = await llm.embedding(node.content);
-			if (e) node.embedding = e.embedding;
-		}));
-		this.commit();
-		return missing.length;
-	}
-}
-
-export type MemoryOptions = {
-	/** Memory object */
-	memory: Memory[] | MemoryCache;
-	/** Inject N memories into the system prompt */
-	inject?: boolean;
-	/** expose recall tool to LLM */
-	tool?: boolean;
-	/** Update memory on compression */
-	update?: boolean;
-	/** Max context size of memories to inject to each call (removed immediately after use) */
-	maxTokens?: number;
-}
-
-export class MemoryManager {
-	private recentlyTouched = new Map<string, number>();
-
-	private queues = new Map<string, {
-		dirty: boolean,
-		request: {abort?: () => void} | null,
-		task: Promise<void>,
-	}>();
-
-	tools = {
-		forget: (memories: Memory[] | MemoryCache): AiTool => ({
-			name: 'memory_forget',
-			description: 'Permanently delete a memory document and clean up all references to it',
-			args: {
-				name: {type: 'string', description: 'Exact memory name to forget', required: true}
-			},
-			fn: (args: any) => {
-				const result = this.forget(args.name, memories);
-				return result ? `Forgotten: ${args.name}` : `Not found: ${args.name}`;
-			},
-		}),
-
-		read: (memories: Memory[] | MemoryCache): AiTool => ({
-			name: 'memory_recall',
-			description: 'Read the full content of a memory document',
-			args: {
-				name: {type: 'string', description: 'Exact memory name', required: true},
-			},
-			fn: (args: any) => {
-				const mem = this.access(memories).find(args.name);
-				if (!mem) return 'Document not found';
-				this.touch(mem.name);
-				return mem.content;
-			},
-		}),
-
-		search: (memories: Memory[] | MemoryCache): AiTool => ({
-			name: 'memory_search',
-			description: 'Use embeddings to find the MOST relevant memories, even if NOT relevant',
-			args: {
-				query: {type: 'string', description: 'What to look for in the memories', required: true},
-				limit: {type: 'number', description: 'Number of memories to return', default: 1},
-			},
-			fn: async ({query, limit}) => {
-				const mem = await this.recollect(query, memories, limit)
-				return mem.map(m => `Memory: ${m.name}
-Description: ${m.description}
-Links: ${[...m.links, ...m.backlinks].join(', ')}
-\`\`\`
-${m.content}
-\`\`\``).join('\n\n');
-			},
-		}),
-	};
-
-	constructor(private llm: any) {}
-
-	static normalize(m?: Memory[] | MemoryCache | MemoryOptions) {
-		if (!m) return null;
-		const raw = m instanceof MemoryCache || Array.isArray(m);
-		return raw ? {memory: <Memory[] | MemoryCache>m, inject: true, tool: true, update: true} : {inject: true, tool: true, update: true, ...m};
-	}
-
-	private access(memories: Memory[] | MemoryCache): MemoryAccessor {
-		return new MemoryAccessor(memories);
-	}
-
-	private appendFacts(node: Memory, facts: string[]): void {
-		this.ensureDoc(node);
-		const body = stripHeader(node.content);
-		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`
-			: `${body.slice(0, idx + FACTS_HEADING.length)}\n${bullets}${body.slice(idx + FACTS_HEADING.length)}`;
-		node.content = this.touchHeader(node, newBody);
-	}
-
-	private ensureDoc(node: Memory): void {
-		if (node.content) return;
-		const title = node.name.split('/').pop() ?? node.name;
-		node.content = this.touchHeader(node, `# ${title}\n`);
-	}
-
-	private async factAgent(conversation: string, store: MemoryAccessor, options: LLMRequest, weekKey: string): Promise<FactBucket[]> {
-		const ghosts = store.ghosts();
-
-		const response = await this.llm.ask(conversation, {
-			model: options.model,
-			temperature: 0.2,
-			system: `You are a fact extractor to build obsidian knowledge vaults.
-Analyze this conversation and extract facts worth remembering long-term.
-
-Rules:
-- Always extract facts that the user explicitly told you to remember
-- ONLY extract current facts the USER explicitly stated about themselves, their work, projects or decisions that were MADE during this conversation
-- DO NOT extract greetings, pleasantries, or generic exchanges
-- DO NOT extract deltas or changes in facts; ONLY the end fact
-- DO NOT extract anything the AI/assistant itself said
-- If nothing worth remembering was said, return an empty buckets array
-
-When extracting facts, you MUST also decide the exact destination path:
-- Reuse node names (including ghost) as much as possible IF the facts belongs there
-- All information primarily about the user should go under "People/User"
-- When required, create a new path following collection/subject format (e.g., People/Sarah, Projects/Oxide) — you are not limited to any fixed list of collections, use whatever fits
-- For journal entries, use "Journal"
-
-Available nodes:
-- Journal
-${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 => `- ${g}: (Ghost)`).join('\n')}` : ''}`,
-			schema: {
-				buckets: {type: 'array', description: 'Groups of facts to remember, each assigned to a different node. Return an empty array if there is nothing worth storing in an obsidian vault', items: {
-					type: 'object', items: {
-							subject: {type: 'string', description: 'Exact existing node name OR new path (e.g. "People/Sarah", "Projects/Oxide"), or "Journal"', required: true},
-							facts: {
-								type: 'array',
-								description: 'Facts to store at this destination',
-								items: {type: 'string', description: 'A single fact'},
-							},
-						},
-					},
-				},
-			},
-		});
-
-		const buckets = new Map<string, string[]>();
-		for(const bucket of response.buckets ?? []) {
-			const subject = bucket.subject.trim().toLowerCase() === 'journal'
-				? `Journal/${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}));
-	}
-
-	private getWeekMonday(date: Date = new Date()): string {
-		const d = new Date(Date.UTC(date.getFullYear(), date.getMonth(), date.getDate()));
-		const day = d.getUTCDay();
-		const diff = day === 0 ? -6 : 1 - day;
-		d.setUTCDate(d.getUTCDate() + diff);
-		return d.toISOString().slice(0, 10);
-	}
-
-	private listNodes(memories: Memory[]): MemoryRef[] {
-		return memories.map(m => ({name: m.name, description: m.description}));
-	}
-
-	private reconcile(node: Memory, memories: Memory[] | MemoryCache, options: LLMRequest): Promise<void> {
-		const key = node.name;
-		const existing = this.queues.get(key);
-		if (existing) {
-			existing.dirty = true;
-			existing.request?.abort?.();
-			return existing.task;
-		}
-
-		const entry = {dirty: false, request: null, task: Promise.resolve()};
-		this.queues.set(key, entry);
-		const store = this.access(memories);
-		entry.task = (async () => {
-			do {
-				entry.dirty = false;
-				await this.docAgent(node, store.list, options, entry);
-			} while (entry.dirty);
-		})().finally(() => {
-			this.queues.delete(key);
-			store.commit();
-		});
-		return entry.task;
-	}
-
-	private async docAgent(node: Memory, memories: Memory[], options: LLMRequest, entry: {request: {abort?: () => void} | null}): Promise<void> {
-		const currentBody = stripHeader(node.content);
-		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-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.
-
-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"):
-\`\`\`markdown
-${GENERIC_TEMPLATE}
-\`\`\`
-
-Formatting rules:
-- Use Obsidian-style markdown: # headings, **bold** for emphasis, bullet & numbered lists for grouped 1D data, tables for 2D data
-- Link related concepts with [[WikiLink]] notation using full paths like [[People/Sarah]] or [[Projects/Website]]
-- Create links for specific entities (person, place, project, program) and abstract concepts, but skip generics (car, red, dog)
-- Keep the document concise, factual, and human-readable
-- Resolve contradictions: newer facts always win — delete the outdated statement entirely, never keep both
-- Do not add frontmatter blocks, filler, preamble, or AI commentary
-
-Other nodes in the vault (link to these instead of duplicating their content):
-${this.listNodes(memories).filter(n => n.name !== node.name).map(n => n.name).join(', ') || 'none'}
-
-Current document:
-\`\`\`markdown
-${currentBody}
-\`\`\``,
-				});
-				entry.request = request;
-				update = await request;
-			}
-		} catch (err: any) {
-			if (err?.name === 'AbortError') return;
-			throw err;
-		} finally {
-			entry.request = null;
-		}
-
-		if (!update?.content) return;
-		node.description = node.name !== 'People/User' ? update.description : 'All information about the current user';
-		node.content = this.touchHeader(node, update.content);
-		const [e] = await this.llm.embedding(node.content);
-		if (e) node.embedding = e.embedding;
-	}
-
-	private parseFrontmatter(content: string): {fm: Map<string, string>, body: string} {
-		const match = content.match(/^---\n([\s\S]*?)\n---\n?([\s\S]*)$/);
-		if (!match) return {fm: new Map(), body: content};
-		const fm = new Map<string, string>();
-		for (const line of match[1].split('\n')) {
-			const i = line.indexOf(':');
-			if (i === -1) continue;
-			fm.set(line.slice(0, i).trim(), line.slice(i + 1).trim());
-		}
-		return {fm, body: match[2]};
-	}
-
-	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('modified', new Date().toISOString());
-		return this.writeFrontmatter(fm, body);
-	}
-
-	private writeFrontmatter(fm: Map<string, string>, body: string): string {
-		const lines = [...fm.entries()].map(([k, v]) => `${k}: ${v}`);
-		return `---\n${lines.join('\n')}\n---\n\n${body.trimStart()}`;
-	}
-
-	decay() {
-		for (const [name, ttl] of this.recentlyTouched) {
-			if (ttl <= 1) this.recentlyTouched.delete(name);
-			else this.recentlyTouched.set(name, ttl - 1);
-		}
-	}
-
-	touch(name: string, ttl = 2) {
-		this.recentlyTouched.set(name, ttl);
-	}
-
-	forget(name: string, memories: Memory[] | MemoryCache): boolean {
-		return this.access(memories).forget(name);
-	}
-
-	async recollect(query: string, memories: Memory[] | MemoryCache, limit = 5, graphDepth = 1): Promise<Memory[]> {
-		const store = this.access(memories);
-		if (!store.list.length) return [];
-
-		await store.backfillEmbeddings(this.llm);
-
-		const [e] = await this.llm.embedding(query);
-		if (!e) return [];
-
-		const vectorResults = store.search(e.embedding, limit);
-		const found = new Set<string>(vectorResults.map(r => r.name));
-
-		if (graphDepth > 0) {
-			let frontier = [...found];
-			for (let depth = 0; depth < graphDepth && frontier.length; depth++) {
-				const next: string[] = [];
-				for (const name of frontier) {
-					const node = store.find(name);
-					if (!node) continue;
-					for (const link of node.links) {
-						if (!found.has(link) && store.find(link)) {
-							found.add(link);
-							next.push(link);
-						}
-					}
-				}
-				frontier = next;
-			}
-		}
-
-		const vectorOrder = vectorResults.map(r => r.name);
-		const graphExpansions = [...found].filter(n => !vectorOrder.includes(n));
-		return [...vectorOrder, ...graphExpansions].map(n => store.find(n)!).filter(Boolean);
-	}
-
-	async memorize(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest): Promise<Memory[]> {
-		const conversation = history
-			.filter(h => h.role === 'user' || h.role === 'assistant')
-			.map(h => `[${h.role}]: ${h.content}`).join('\n\n').trim();
-		if (!conversation) return [];
-
-		const uid = `${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 buckets = await this.factAgent(conversation, store, options, this.getWeekMonday());
-		const touched: Memory[] = [];
-
-		for (const {subject, facts} of buckets) {
-			let node = store.find(subject);
-			if (!node) {
-				node = {name: subject, description: '', content: '', embedding: [], links: [], backlinks: []};
-				store.list.push(node);
-			}
-			this.appendFacts(node, facts);
-			const [e] = await this.llm.embedding(node.content);
-			if (e) node.embedding = e.embedding;
-			this.touch(node.name);
-			touched.push(node);
-		}
-
-		if (touched.length) {
-			store.commit();
-			(pending as any).content = `Saved to ${touched.map(n => `[[${n.name}]]`).join(', ')}`;
-			await Promise.all(touched.map(node => this.reconcile(node, memories, options).catch(() => {})));
-		} else {
-			(pending as any).content = 'Nothing worth remembering.';
-		}
-
-		(touched as any).uid = uid;
-		return touched;
-	}
-
-	async reconcileVault(memories: Memory[] | MemoryCache, options: LLMRequest, scope: 'touched' | 'all' = 'touched'): Promise<void> {
-		const store = this.access(memories);
-		const targets = scope === 'all' ? store.list : store.list.filter(m => m.content.includes(FACTS_HEADING));
-		await Promise.all(targets.map(node => this.reconcile(node, memories, options)));
-		store.commit();
-	}
-}

src/memory/graph.ts

@@ -1,4 +1,5 @@
-import {Memory, MemoryCache} from './memory.ts';
+import {MemoryCache} from './memory-state.ts';
+import type {Memory} from './memory.ts';
 
 export type MemoryNode = {
 	name: string;
@@ -13,6 +14,48 @@ export function extractLinks(content: string): string[] {
 	return [...new Set([...matches].map(m => m[1].trim()))];
 }
 
+export function patchGraph(mems: Memory[], nodes: MemoryNode[], changed: Memory[]): MemoryNode[] {
+	const nameSet = new Set(mems.map(m => m.name));
+	const byName = new Map(nodes.map(n => [n.name, n]));
+
+	const ensureNode = (name: string): MemoryNode => {
+		let n = byName.get(name);
+		if (!n) {
+			n = {name, missing: !nameSet.has(name), links: [], backlinks: []};
+			byName.set(name, n);
+		}
+		return n;
+	};
+
+	for (const m of changed) {
+		const node = ensureNode(m.name);
+		node.missing = false; // real memory, promotes any pre-existing ghost entry
+		const oldLinks = m.links ?? [];
+		const newLinks = extractLinks(m.content).filter(l => l !== m.name);
+
+		for (const target of oldLinks.filter(l => !newLinks.includes(l))) {
+			const t = byName.get(target);
+			if (!t) continue;
+			t.backlinks = t.backlinks.filter(n => n !== m.name);
+			if (t.missing && !t.backlinks.length) byName.delete(target); // fully dereferenced ghost
+		}
+		for (const target of newLinks.filter(l => !oldLinks.includes(l))) {
+			const t = ensureNode(target);
+			if (!t.backlinks.includes(m.name)) t.backlinks.push(m.name);
+		}
+
+		m.links = newLinks;
+		node.links = newLinks;
+	}
+
+	for (const m of mems) {
+		const n = byName.get(m.name);
+		if (n) m.backlinks = n.backlinks;
+	}
+
+	return [...byName.values()];
+}
+
 export function rebuildGraph(memories: Memory[] | MemoryCache): MemoryNode[] {
 	const mems = memories instanceof MemoryCache ? memories.memories : memories;
 	const nameSet = new Set(mems.map(m => m.name));

src/memory/kd-tree.ts

@@ -1,3 +1,5 @@
+import {cosineDistance, euclideanDistance} from '../utils.ts';
+
 export type DistanceMetric = "euclidean" | "cosine";
 
 export interface KDPoint<T = unknown> {
@@ -15,28 +17,7 @@ interface KDNode<T> {
 	axis: number;
 	left: KDNode<T> | null;
 	right: KDNode<T> | null;
-}
+	deleted?: boolean;
-
-// ─── Distance helpers ─────────────────────────────────────────────────────────
-
-function euclidean(a: number[], b: number[]): number {
-	let sum = 0;
-	for (let i = 0; i < a.length; i++) {
-		const d = a[i] - b[i];
-		sum += d * d;
-	}
-	return Math.sqrt(sum);
-}
-
-function cosine(a: number[], b: number[]): number {
-	let dot = 0, normA = 0, normB = 0;
-	for (let i = 0; i < a.length; i++) {
-		dot   += a[i] * b[i];
-		normA += a[i] * a[i];
-		normB += b[i] * b[i];
-	}
-	const denom = Math.sqrt(normA) * Math.sqrt(normB);
-	return denom === 0 ? 1 : 1 - dot / denom; // distance = 1 - similarity
 }
 
 /**
@@ -95,6 +76,7 @@ class BoundedMaxHeap<T> {
  *
  * Supports:
  *  - Insertion of labeled points
+ *  - Lazy (tombstone) removal, physically purged on rebalance()
  *  - k-nearest-neighbor (KNN) search
  *  - Radius search (all points within a given distance)
  *  - Euclidean and cosine distance metrics
@@ -103,6 +85,7 @@ 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;
@@ -120,7 +103,7 @@ export class KDTree<T = unknown> {
 		points?: KDPoint<T>[]
 	) {
 		this.dims = dims;
-		this.distanceFn = metric === "cosine" ? cosine : euclidean;
+		this.distanceFn = metric === "cosine" ? cosineDistance : euclideanDistance;
 
 		if (points && points.length > 0) {
 			this.validateAll(points);
@@ -129,9 +112,15 @@ export class KDTree<T = unknown> {
 		}
 	}
 
-	/** Total number of points stored in the tree. */
+	/** Total number of live points stored in the tree (excludes tombstoned). */
 	get size(): number { return this._size; }
 
+	/** Fraction of physical nodes that are tombstoned (pending removal on next rebalance). */
+	get tombstoneRatio(): number {
+		const total = this._size + this._tombstones;
+		return total ? this._tombstones / total : 0;
+	}
+
 	// ── Insertion ──────────────────────────────────────────────────────────────
 
 	/**
@@ -144,10 +133,36 @@ export class KDTree<T = unknown> {
 		this._size++;
 	}
 
+	// ── Removal ────────────────────────────────────────────────────────────────
+
+	/**
+	 * Lazily remove all live points whose payload matches `predicate`.
+	 * O(n) traversal, but avoids a full tree rebuild. Call `rebalance()`
+	 * periodically (e.g. once tombstoneRatio crosses ~0.25) to reclaim space
+	 * and restore optimal query depth.
+	 * @returns number of points removed
+	 */
+	remove(predicate: (payload: T) => boolean): number {
+		let removed = 0;
+		const visit = (node: KDNode<T> | null): void => {
+			if (!node) return;
+			if (!node.deleted && predicate(node.point.payload)) {
+				node.deleted = true;
+				removed++;
+			}
+			visit(node.left);
+			visit(node.right);
+		};
+		visit(this.root);
+		this._size -= removed;
+		this._tombstones += removed;
+		return removed;
+	}
+
 	// ── KNN search ─────────────────────────────────────────────────────────────
 
 	/**
-	 * Find the k nearest neighbors to `query`.
+	 * Find the k nearest live neighbors to `query`.
 	 * Returns results sorted by distance ascending.
 	 */
 	knn(query: number[], k: number): KNNResult<T>[] {
@@ -171,7 +186,7 @@ export class KDTree<T = unknown> {
 	// ── Radius search ──────────────────────────────────────────────────────────
 
 	/**
-	 * Return all points whose distance to `query` is ≤ `radius`,
+	 * Return all live points whose distance to `query` is ≤ `radius`,
 	 * sorted by distance ascending.
 	 */
 	radiusSearch(query: number[], radius: number): KNNResult<T>[] {
@@ -186,7 +201,7 @@ export class KDTree<T = unknown> {
 
 	// ── Conversion ─────────────────────────────────────────────────────────────
 
-	/** Collect all points in the tree (order not guaranteed). */
+	/** Collect all live points in the tree (order not guaranteed). */
 	toArray(): KDPoint<T>[] {
 		const out: KDPoint<T>[] = [];
 		this.collect(this.root, out);
@@ -194,12 +209,14 @@ export class KDTree<T = unknown> {
 	}
 
 	/**
-	 * Rebuild the tree from its current points as a balanced tree.
+	 * Rebuild the tree from its current live points as a balanced tree.
-	 * Useful after many individual insertions to restore O(log n) query time.
+	 * Physically purges tombstones and restores O(log n) query time.
 	 */
 	rebalance(): void {
 		const points = this.toArray();
 		this.root = points.length ? this.buildBalanced(points, 0) : null;
+		this._size = points.length;
+		this._tombstones = 0;
 	}
 
 	// ── Private: build ─────────────────────────────────────────────────────────
@@ -251,8 +268,10 @@ export class KDTree<T = unknown> {
 	): void {
 		if (node === null) return;
 
-		const dist = this.distanceFn(query, node.point.vector);
+		if (!node.deleted) {
-		heap.push({ point: node.point, distance: dist });
+			const dist = this.distanceFn(query, node.point.vector);
+			heap.push({ point: node.point, distance: dist });
+		}
 
 		const axis = node.axis;
 		const diff = query[axis] - node.point.vector[axis];
@@ -261,11 +280,8 @@ export class KDTree<T = unknown> {
 			: [node.right, node.left];
 
 		this.searchKNN(near, query, k, heap, depth + 1);
-
-		// Only explore the far side if it could contain a closer point.
-		// For cosine distance we can't prune by axis gap alone, so always explore.
 		const shouldExplore =
-			this.distanceFn === cosine
+			this.distanceFn === cosineDistance
 				? true
 				: Math.abs(diff) < heap.worstDistance;
 
@@ -285,9 +301,11 @@ export class KDTree<T = unknown> {
 	): void {
 		if (node === null) return;
 
-		const dist = this.distanceFn(query, node.point.vector);
+		if (!node.deleted) {
-		if (dist <= radius) {
+			const dist = this.distanceFn(query, node.point.vector);
-			results.push({ point: node.point, distance: dist });
+			if (dist <= radius) {
+				results.push({ point: node.point, distance: dist });
+			}
 		}
 
 		const axis = node.axis;
@@ -299,7 +317,7 @@ export class KDTree<T = unknown> {
 		this.searchRadius(near, query, radius, results, depth + 1);
 
 		const shouldExplore =
-			this.distanceFn === cosine ? true : Math.abs(diff) <= radius;
+			this.distanceFn === cosineDistance ? true : Math.abs(diff) <= radius;
 
 		if (shouldExplore) {
 			this.searchRadius(far, query, radius, results, depth + 1);
@@ -310,7 +328,7 @@ export class KDTree<T = unknown> {
 
 	private collect(node: KDNode<T> | null, out: KDPoint<T>[]): void {
 		if (node === null) return;
-		out.push(node.point);
+		if (!node.deleted) out.push(node.point);
 		this.collect(node.left, out);
 		this.collect(node.right, out);
 	}

src/memory/memory-state.ts

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

src/memory/memory.ts

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

src/open-ai.ts

@@ -36,21 +36,49 @@ export class OpenAi extends LLMProvider {
 	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)}}],
-				}, {
-					role: 'tool',
-					tool_call_id: h.id,
-					content: h.error || h.content || '',
-				});
-			} else {
 				wire.push({role: h.role, content: this.toWireContent(h.content)});
+				continue;
 			}
+
+			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;
 	}
 
@@ -60,13 +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.maxTokens || this.ai.options.llm?.maxTokens || 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: {
@@ -74,8 +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])
+								: []
 						}
 					}
 				}))
@@ -83,60 +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};
 						});
@@ -144,12 +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) {
@@ -164,7 +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);

src/utils.ts

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