More memory optimizations

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "1.6.13",
+	"version": "1.7.2",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",

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 */
@@ -265,22 +274,21 @@ class LLM {
 		};
 	}
 
-	private setupAgent(agents: Agent[] = [], allAgents: Agent[], history: LLMMessage[], aborts: ((keep?: boolean) => 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>` : ''}`;
 
@@ -297,7 +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);
@@ -451,7 +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);
@@ -495,7 +503,7 @@ 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));
 				}
 			}
 
@@ -588,24 +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

@@ -1,794 +0,0 @@
-import {MemoryNode, patchGraph, rebuildGraph} from './helpers.ts';
-import {LLMRequest, LLMMessage} from './llm.ts';
-import {AiTool} from './tools.ts';
-import {KDTree} from './kd-tree.ts';
-
-const FACT_SIMILARITY_THRESHOLD = 0.62;
-const PENDING_HEADING = '## Pending';
-const TODO_HEADING = '## Todo list';
-const TREE_TOMBSTONE_LIMIT = 0.25;
-const ALIAS_MATCH_THRESHOLD = 0.55;
-
-export type Memory = {
-	name: string;
-	description: string;
-	content: string;
-	embedding: number[];
-	titleEmbedding?: number[];
-	bodyEmbeddings?: number[][];
-	links: string[];
-	backlinks: string[];
-}
-
-type MemoryRef = {
-	name: string;
-	description: string;
-	distance?: number;
-}
-
-type FactBucket = {
-	subject: string;
-	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) {
-		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 => ({name: m.name, description: m.description, distance: cosineDistance(query, m.embedding)}))
-		.sort((a, b) => a.distance - b.distance)
-		.slice(0, limit);
-}
-
-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[] = [];
-
-	get length() { return this.memories.length; }
-
-	constructor(memories: Memory[]) {
-		this.memories = memories;
-		this.tree = new KDTree<MemoryRef>(0);
-		this.rebuild();
-	}
-
-	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; // 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 [];
-		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.memories.find(m => m.name === memory.name);
-		if(existing) Object.assign(existing, memory);
-		else this.memories.push(memory);
-		this.rebuild([existing ?? memory]);
-	}
-
-	remove(name: string): void {
-		const idx = this.memories.findIndex(m => m.name === name);
-		if(idx !== -1) {
-			this.memories.splice(idx, 1);
-			this.rebuild();
-		}
-	}
-
-	rebuild(changed?: Memory[]): void {
-		this.nodes = (changed?.length && this.nodes.length)
-			? patchGraph(this.memories, this.nodes, changed)
-			: rebuildGraph(this.memories);
-		this.syncTree();
-	}
-}
-
-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(changed?: Memory[]): MemoryNode[] {
-		if(this.cache) {
-			this.cache.rebuild(changed);
-			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(node => embedMemoryFields(node, llm)));
-		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 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 => ({
-			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 = new MemoryAccessor(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 stage(node: Memory, block: string): void {
-		if(!node.content) {
-			const title = node.name.split('/').pop() ?? node.name;
-			node.content = this.touchHeader(node, `# ${title}\n`);
-		}
-		const body = stripHeader(node.content);
-		const idx = body.indexOf(PENDING_HEADING);
-		const newBody = idx === -1
-			? `${body.trimEnd()}\n\n${PENDING_HEADING}\n${block}\n`
-			: `${body.slice(0, idx + PENDING_HEADING.length)}\n${block}${body.slice(idx + PENDING_HEADING.length)}`;
-		node.content = this.touchHeader(node, newBody);
-	}
-
-	private resolveSubject(subject: string, store: MemoryAccessor): string {
-		function normalize(name: string): string {
-			return name.trim().toLowerCase().replace(/\s+/g, ' ');
-		}
-
-		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): Promise<FactAgentResult> {
-		const ghosts = store.ghosts();
-
-		const response = await this.llm.ask(conversation, {
-			model: options.model,
-			temperature: 0.2,
-			system: `Turn this conversation into a persistent memory file by extracting information into organized bullet points
-
-Think of this like an Obsidian vault with a clear division of responsibility:
-- The JOURNAL is a timeline. It answers "what happened, and when" and is the only place with a sense of time.
-- ENTITY DOSSIERS are a wiki. They answer "what is currently true about this subject", with no sense of time — only current state.
-- Never blur the two: a one-off event, conversation, or debugging session is a journal entry, not an entity, even if it's detailed.
-
-1. Journal Log
-- A chronological, skimmable log of what actually happened: real discussions, decisions made, progress on projects, problems worked through
-- This is NOT a transcript, and it is NOT a step-by-step record, its a compressed log of notable events & developments
-- One line per development is usually enough: what was worked on and the outcome, not the blow-by-blow of how
-- 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:
-${this.listNodes(store.list).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}
-${ghosts.length ? `${ghosts.map(g => `- ${g}: (Ghost)`).join('\n')}` : ''}`,
-			schema: {
-				journal: {type: 'string', description: 'Short day-to-day recap; empty if nothing happened.', required: false},
-				tasks: {
-					type: 'array', description: 'Concrete tasks mentioned or completed in the conversation.', required: false, items: {
-						type: 'object', items: {
-							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},
-							task: {type: 'string', description: 'Concise actionable task', required: true},
-							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'}},
-						},
-					},
-				},
-			},
-		});
-
-		const buckets = new Map<string, string[]>();
-		for(const bucket of response.buckets ?? []) {
-			const subject = bucket.subject.trim();
-			const facts = buckets.get(subject) ?? [];
-			facts.push(...dedupeFacts(bucket.facts));
-			buckets.set(subject, facts);
-		}
-
-		return {
-			buckets: buckets.entries().toArray().map(([subject, facts]) => ({subject, facts})),
-			journal: (response.journal ?? '').trim(),
-			tasks: response.tasks ?? [],
-		};
-	}
-
-	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;
-		d.setUTCDate(d.getUTCDate() + diff);
-		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) {
-			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 = new MemoryAccessor(memories);
-		entry.task = (async () => {
-			let current = node;
-			try {
-				do {
-					entry.dirty = false;
-					await this.docAgent(current, store.list, options, entry);
-					this.mergeLock = this.mergeLock.then(() => this.mergeAgent(current, memories, options));
-					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> {
-		if(!memories.includes(node)) return;
-		const currentBody = stripHeader(node.content);
-		const journal = node.name.startsWith('Journal/');
-		const system = (journal
-			? `You maintain one persistent journal document
-
-Rewrite the ENTIRE journal, folding "## Pending" into the existing content removing the heading
-
-Journal design:
-- Preserve the chronological daily log
-- 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
-
-Rewrite the ENTIRE document, folding "## Pending" into the existing content. Remove the Pending section when finished.
-
-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
-${currentBody}
-\`\`\``;
-		let update;
-		try {
-			for(let i = 0; i < 2 && !update?.content; i++) {
-				const request = this.llm.ask(currentBody, {
-					model: options.model,
-					temperature: 0.3,
-					schema: {
-						description: {type: 'string', description: 'One factual sentence describing the document\'s ENTIRE SUBJECT MATTER — for use as a search/merge fingerprint', required: true},
-						content: {type: 'string', description: 'Rewritten document body in markdown, without the frontmatter block', required: true},
-					},
-					system,
-				});
-				entry.request = request;
-				update = await request;
-			}
-		} catch(err: any) {
-			if(err?.name === 'AbortError') return;
-			throw err;
-		} finally {
-			entry.request = null;
-		}
-
-		if(!update?.content) return;
-		node.description = node.name.startsWith('Journal/') ? this.journalDescription(node.name) : node.name !== 'People/User' ? update.description.replaceAll(/[\n:]/g, '') : 'All information about the current user';
-		node.content = this.touchHeader(node, update.content);
-		await embedMemoryFields(node, this.llm);
-	}
-
-	private parseFrontmatter(content: string): {fm: Map<string, string>, body: string} {
-		const match = content.match(/^---\n([\s\S]*?)\n---\n?([\s\S]*)$/);
-		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;
-			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 touchHeader(node: Memory, body: string): string {
-		const {fm} = this.parseFrontmatter(node.content);
-		fm.set('name', node.name);
-		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, stripHeader(body));
-	}
-
-	private 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()}`;
-	}
-
-	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 new MemoryAccessor(memories).forget(name);
-	}
-
-	async recollect(query: string, memories: Memory[] | MemoryCache, limit = 5, graphDepth = 1): Promise<Memory[]> {
-		function rank(query: number[], candidates: Memory[], limit: number): Memory[] {
-			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 [];
-
-		const pool = store.search(e.embedding, Math.max(limit * 3, limit));
-		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) {
-			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 rankedOrder = ranked.map(m => m.name);
-		const graphExpansions = [...found].filter(n => !rankedOrder.includes(n));
-		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}]: ${h.content}`).join('\n\n').trim();
-		if(!conversation) return [];
-
-		const uid = `${Date.now()}_${Math.random().toString(36).slice(2)}`;
-		const pending = {role: 'tool', name: 'memory_process', id: uid, content: conversation} as unknown as LLMMessage;
-		history.push(pending);
-
-		const store = new MemoryAccessor(memories);
-		const {buckets, journal, tasks} = await this.factAgent(conversation, store, options);
-		const touched: Memory[] = [];
-
-		const personalTasks = tasks.filter(isPersonalTask);
-		const entityTasks = tasks.filter(t => !isPersonalTask(t));
-
-		if(journal || personalTasks.length) {
-			const journalName = `Journal/${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);
-			}
-			const blocks: string[] = [];
-			if(facts.length) blocks.push(facts.map(f => `- ${f}`).join('\n'));
-			if(subjectTasks.length) blocks.push(`${TODO_HEADING}\n${subjectTasks.map(t => `- [${t.done ? 'x' : ' '}] ${t.task}`).join('\n')}`);
-			if(blocks.length) this.stage(node, blocks.join('\n\n'));
-			touched.push(node);
-		}
-
-		await Promise.all(touched.map(async node => {
-			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(', ')}`;
-			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 reconcileAll(memories: Memory[] | MemoryCache, options: LLMRequest, scope: 'touched' | 'all' = 'touched'): Promise<void> {
-		const store = new MemoryAccessor(memories);
-		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/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;
@@ -13,14 +14,6 @@ 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]));

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> {
@@ -18,28 +20,6 @@ interface KDNode<T> {
 	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
-}
-
 /**
  * Keeps the k closest candidates in memory, evicts the furthest when full
  */
@@ -123,7 +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);
@@ -300,11 +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;
 
@@ -340,7 +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);

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/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));
+}