Added MCP, Hybrid Memories and Skill support

src/ai.ts

@@ -8,7 +8,7 @@ export type AbortablePromise<T> = Promise<T> & {
 };
 
 export type AiOptions = {
-	/** Token to pull models from hugging face */
+	/** Token to pull diarization models from hugging face */
 	hfToken?: string;
 	/** Path to models */
 	path?: string;

src/audio.ts

@@ -2,7 +2,6 @@ import {execSync, spawn} from 'node:child_process';
 import {mkdtempSync} from 'node:fs';
 import fs from 'node:fs/promises';
 import {tmpdir} from 'node:os';
-import * as path from 'node:path';
 import Path, {join} from 'node:path';
 import {AbortablePromise, Ai} from './ai.ts';
 
@@ -155,7 +154,7 @@ print(json.dumps(segments))
 		const p = new Promise<any>((resolve, reject) => {
 			this.downloadAsrModel(opts.model).then(m => {
 				if(opts.diarization) {
-					let output = path.join(path.dirname(file), 'transcript');
+					let output = join(Path.dirname(file), 'transcript');
 					proc = spawn(<string>this.ai.options.whisper,
 						['-m', m, '-f', file, '-np', '-ml', '1', '-oj', '-of', output],
 						{stdio: ['ignore', 'ignore', 'pipe']}
@@ -226,11 +225,11 @@ print(json.dumps(segments))
 		return <any>Object.assign(p, {abort});
 	}
 
-	asr(file: string, options: { model?: string; diarization?: boolean | 'llm' } = {}): AbortablePromise<string | null> {
+	asr(path: string, options: { model?: string; diarization?: boolean | 'llm' } = {}): AbortablePromise<string | null> {
 		if(!this.ai.options.whisper) throw new Error('Whisper not configured');
 
 		const tmp = join(mkdtempSync(join(tmpdir(), 'audio-')), 'converted.wav');
-		execSync(`ffmpeg -i "${file}" -ar 16000 -ac 1 -f wav "${tmp}"`, { stdio: 'ignore' });
+		execSync(`ffmpeg -i "${path}" -ar 16000 -ac 1 -f wav "${tmp}"`, { stdio: 'ignore' });
 		const clean = () => fs.rm(Path.dirname(tmp), {recursive: true, force: true}).catch(() => {});
 
 		if(!options.diarization) return this.runAsr(tmp, {model: options.model});

src/embedder.ts

@@ -1,13 +1,13 @@
-import { pipeline } from '@xenova/transformers';
+import { pipeline } from '@huggingface/transformers';
 
 const [modelDir, model] = process.argv.slice(2);
 
 let text = '';
 process.stdin.on('data', chunk => text += chunk);
 process.stdin.on('end', async () => {
-	const embedder = await pipeline('feature-extraction', 'Xenova/' + model, {quantized: true, cache_dir: modelDir});
+	const embedder = await pipeline('feature-extraction', 'Xenova/' + model, {cache_dir: modelDir});
 	const output = await embedder(text, { pooling: 'mean', normalize: true });
 	const embedding = Array.from(output.data);
-	console.log(JSON.stringify({embedding}));
+	process.stdout.write(JSON.stringify({embedding}));
 	process.exit();
 });

src/llm.ts

@@ -1,5 +1,3 @@
-import {sum} from '@tensorflow/tfjs';
-import {JSONAttemptParse} from '@ztimson/utils';
 import {AbortablePromise, Ai} from './ai.ts';
 import {Anthropic} from './antrhopic.ts';
 import {OpenAi} from './open-ai.ts';
@@ -7,7 +5,8 @@ import {LLMProvider} from './provider.ts';
 import {AiTool} from './tools.ts';
 import {fileURLToPath} from 'url';
 import {dirname, join} from 'path';
-import { spawn } from 'node:child_process';
+import {spawn} from 'node:child_process';
+import {Memory, MemoryManager} from './memory.ts';
 
 export type AnthropicConfig = {proto: 'anthropic', token: string};
 export type OllamaConfig = {proto: 'ollama', host: string};
@@ -37,16 +36,6 @@ export type LLMMessage = {
 	timestamp?: number;
 }
 
-/** Background information the AI will be fed */
-export type LLMMemory = {
-	/** What entity is this fact about */
-	owner: string;
-	/** The information that will be remembered */
-	fact: string;
-	/** Owner and fact embedding vector */
-	embeddings: [number[], number[]];
-}
-
 export type LLMRequest = {
 	/** System prompt */
 	system?: string;
@@ -63,17 +52,39 @@ export type LLMRequest = {
 	/** Stream response */
 	stream?: (chunk: {text?: string, tool?: string, done?: true}) => any;
 	/** Compress old messages in the chat to free up context */
-	compress?: {
-		/** Trigger chat compression once context exceeds the token count */
-		max: number;
-		/** Compress chat until context size smaller than */
-		min: number
-	},
-	/** Background information the AI will be fed */
-	memory?: LLMMemory[],
+	compress?: {max: number; min: number};
+	/** User's memory documents - RAG injected automatically each turn */
+	memory?: Memory[];
+	/** Model to use for memory operations */
+	memoryModel?: string;
+	/** Skill documents the AI can browse and read on demand */
+	skills?: Skill[];
+	/** MCP servers to connect and expose as tools */
+	mcp?: McpServer[];
 }
 
+export type McpServer = {
+	/** MCP server name for humans */
+	name: string;
+	/** Host URL */
+	host: string;
+	/** Server access token */
+	token?: string;
+}
+
+export type Skill = {
+	/** Name of skill for humans */
+	name: string;
+	/** Description LLM will use to decide to learn a skill */
+	description: string;
+	/** Skill instructions */
+	content: string;
+}
+
+
 class LLM {
+	private memoryManager!: MemoryManager;
+
 	defaultModel!: string;
 	models: {[model: string]: LLMProvider} = {};
 
@@ -85,14 +96,67 @@ class LLM {
 			else if(config.proto == 'ollama') this.models[model] = new OpenAi(this.ai, config.host, 'not-needed', model);
 			else if(config.proto == 'openai') this.models[model] = new OpenAi(this.ai, config.host || null, config.token, model);
 		});
+		this.memoryManager = new MemoryManager(this);
+	}
+
+	private async setupMcp(servers: McpServer[] = []): Promise<{prompt: string, tools: AiTool[]}> {
+		if(!servers?.length) return {prompt: '', tools: []};
+		const allTools: AiTool[] = [];
+		await Promise.all(servers.map(async server => {
+			const res = await fetch(`${server.host}/tools`, {headers: server.token ? {Authorization: `Bearer ${server.token}`} : {}});
+			const mcp: any = await res.json();
+			if(!mcp?.tools) return;
+			for(const t of mcp.tools) {
+				const args: Record<string, any> = {};
+				if(t.inputSchema?.properties) {
+					for(const [key, val] of Object.entries<any>(t.inputSchema.properties)) {
+						args[key] = {type: val.type || 'string', description: val.description || '', required: t.inputSchema.required?.includes(key)};
+					}
+				}
+				allTools.push({
+					name: `${server.name}_${t.name}`,
+					description: t.description || '',
+					args,
+					fn: async (a: any) => {
+						const r = await fetch(`${server.host}/tools/call`, {
+							method: 'POST',
+							headers: {'Content-Type': 'application/json', ...(server.token ? {Authorization: `Bearer ${server.token}`} : {})},
+							body: JSON.stringify({name: t.name, arguments: a})
+						});
+						const data: any = await r.json();
+						return data?.content?.[0]?.text ?? JSON.stringify(data);
+					}
+				});
+			}
+		}));
+
+		const list = allTools.map(t => `- ${t.name}: ${t.description}`).join('\n');
+		return {
+			prompt: `You have access to the following MCP tools:\n${list}`,
+			tools: allTools
+		};
+	}
+
+	private setupSkills(skills: Skill[] = []): {prompt: string, tools: AiTool[]} {
+		if(!skills?.length) return {prompt: '', tools: []};
+		const list = skills.map(s => `- ${s.name}: ${s.description}`).join('\n');
+		return {
+			prompt: `You have access to the following skill documents, use \`read_skill\` to access them:\n${list}`,
+			tools: [{
+				name: 'read_skill',
+				description: 'Read the full content of a skill/knowledge document',
+				args: {
+					name: {type: 'string', description: 'Exact skill name', required: true}
+				},
+				fn: (args: any) => {
+					const skill = skills.find(s => s.name === args.name);
+					if(!skill) return `Skill not found. Available:\n${list}`;
+					return `# ${skill.name}\n${skill.content}`;
+				}
+			}]
+		}
 	}
 
-	/**
-	 * Chat with LLM
-	 * @param {string} message Question
-	 * @param {LLMRequest} options Configuration options and chat history
-	 * @returns {{abort: () => void, response: Promise<string>}} Function to abort response and chat history
-	 */
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
 		options = <any>{
 			system: '',
@@ -106,71 +170,51 @@ class LLM {
 		if(!this.models[m]) throw new Error(`Model does not exist: ${m}`);
 		let abort = () => {};
 		return Object.assign(new Promise<string>(async res => {
+			let tools: AiTool[] = options.tools || this.ai.options.llm?.tools || [];
+			const prompts: string[] = [options.system || this.ai.options.llm?.system || ''];
 			if(!options.history) options.history = [];
-			// If memories were passed, find any relevant ones and add a tool for ADHOC lookups
-			if(options.memory) {
-				const search = async (query?: string | null, subject?: string | null, limit = 10) => {
-					const [o, q] = await Promise.all([
-						subject ? this.embedding(subject) : Promise.resolve(null),
-						query ? this.embedding(query) : Promise.resolve(null),
-					]);
-					return (options.memory || []).map(m => {
-						const score = (o ? this.cosineSimilarity(m.embeddings[0], o[0].embedding) : 0)
-							+ (q ? this.cosineSimilarity(m.embeddings[1], q[0].embedding) : 0);
-						return {...m, score};
-					}).toSorted((a: any, b: any) => a.score - b.score).slice(0, limit)
-						.map(m => `- ${m.owner}: ${m.fact}`).join('\n');
-				}
 
-				options.system += '\nYou have RAG memory and will be given the top_k closest memories regarding the users query. Save anything new you have learned worth remembering from the user message using the remember tool and feel free to recall memories manually.\n';
-				const relevant = await search(message);
-				if(relevant.length) options.history.push({role: 'tool', name: 'recall', id: 'auto_recall_' + Math.random().toString(), args: {}, content: `Things I remembered:\n${relevant}`});
-				options.tools = [{
-					name: 'recall',
-					description: 'Recall the closest memories you have regarding a query using RAG',
-					args: {
-						subject: {type: 'string', description: 'Find information by a subject topic, can be used with or without query argument'},
-						query: {type: 'string', description: 'Search memory based on a query, can be used with or without subject argument'},
-						topK: {type: 'number', description: 'Result limit, default 5'},
-					},
-					fn: (args) => {
-						if(!args.subject && !args.query) throw new Error('Either a subject or query argument is required');
-						return search(args.query, args.subject, args.topK);
-					}
-				}, {
-					name: 'remember',
-					description: 'Store important facts user shares for future recall',
-					args: {
-						owner: {type: 'string', description: 'Subject/person this fact is about'},
-						fact: {type: 'string', description: 'The information to remember'}
-					},
-					fn: async (args) => {
-						if(!options.memory) return;
-						const e = await Promise.all([
-							this.embedding(args.owner),
-							this.embedding(`${args.owner}: ${args.fact}`)
-						]);
-						const newMem = {owner: args.owner, fact: args.fact, embeddings: <any>[e[0][0].embedding, e[1][0].embedding]};
-						options.memory.splice(0, options.memory.length, ...[
-							...options.memory.filter(m => {
-								return !(this.cosineSimilarity(newMem.embeddings[0], m.embeddings[0]) >= 0.9 && this.cosineSimilarity(newMem.embeddings[1], m.embeddings[1]) >= 0.8);
-							}),
-							newMem
-						]);
-						return 'Remembered!';
-					}
-				}, ...options.tools || []];
+			// MCP
+			const mcp = options.mcp || this.ai.options?.llm?.mcp;
+			if(mcp?.length) {
+				const m = await this.setupMcp(mcp);
+				prompts.unshift(m.prompt);
+				tools.push(...m.tools);
 			}
 
-			// Ask
-			const resp = await this.models[m].ask(message, options);
+			// Skills
+			const skills = options.skills || this.ai.options?.llm?.skills;
+			if(skills?.length) {
+				const s = this.setupSkills(skills);
+				prompts.unshift(s.prompt);
+				tools.push(...s.tools);
+			}
 
-			// Remove any memory calls from history
-			if(options.memory) options.history.splice(0, options.history.length, ...options.history.filter(h => h.role != 'tool' || (h.name != 'recall' && h.name != 'remember')));
+			// Memory
+			if(options.memory) {
+				const relevant = await this.memoryManager.recollect(message, options.memory);
+				if(relevant.length) {
+					const context = relevant.map(m => `### ${m.name}\n${m.content}`).join('\n\n');
+					options.history.push({
+						id: 'auto_recall_' + Math.random().toString(), role: 'tool', name: 'recall', args: {},
+						content: `Knowledge Documents:\n\n${context}`
+					});
+				}
+				prompts.unshift('You have access to a knowledge base. Relevant documents are injected automatically before each message. Use this knowledge to inform your responses.');
+			}
 
-			// Compress message history
+			const resp = await this.models[m].ask(message, {...options, tools, system: prompts.filter(Boolean).join('\n\n')});
+
+			// Trim memory injections from history
+			if(options.memory) {
+				options.history.splice(0, options.history.length, ...options.history.filter(h =>
+					h.role !== 'tool' || h.name !== 'recall'));
+			}
+
+			// Auto-memorize before compressing
 			if(options.compress) {
-				const compressed = await this.ai.language.compressHistory(options.history, options.compress.max, options.compress.min, options);
+				if(options.memory) await this.memoryManager.memorize(options.history, options.memory, options);
+				const compressed = await this.compressHistory(options.history, options.compress.max, options.compress.min, options);
 				options.history.splice(0, options.history.length, ...compressed);
 			}
 
@@ -178,13 +222,12 @@ class LLM {
 		}), {abort});
 	}
 
-	async code(message: string, options?: LLMRequest): Promise<any> {
-		const resp = await this.ask(message, {...options, system: [
-			options?.system,
-			'Return your response in a code block'
-		].filter(t => !!t).join(('\n'))});
-		const codeBlock = /```(?:.+)?\s*([\s\S]*?)```/.exec(resp);
-		return codeBlock ? codeBlock[1].trim() : null;
+	/**
+	 * Digest full conversation history into memory documents.
+	 * Call on session end to persist the conversation.
+	 */
+	async updateMemory(history: LLMMessage[], memories: Memory[], options: LLMRequest = {}): Promise<void> {
+		await this.memoryManager.memorize(history, memories, {model: this.defaultModel, ...options});
 	}
 
 	/**
@@ -273,7 +316,7 @@ class LLM {
 	 * @param {maxTokens?: number, overlapTokens?: number} opts Options for embedding such as chunk sizes
 	 * @returns {Promise<Awaited<{index: number, embedding: number[], text: string, tokens: number}>[]>} Chunked embeddings
 	 */
-	embedding(target: object | string, opts: {maxTokens?: number, overlapTokens?: number} = {}): AbortablePromise<any[]> {
+	embedding(target: object | string, opts: {maxTokens?: number, overlapTokens?: number} = {}): AbortablePromise<{index: number, embedding: number[], text: string, tokens: number}[]> {
 		let {maxTokens = 500, overlapTokens = 50} = opts;
 		let aborted = false;
 		const abort = () => { aborted = true; };
@@ -281,7 +324,6 @@ class LLM {
 		const embed = (text: string): Promise<number[]> => {
 			return new Promise((resolve, reject) => {
 				if(aborted) return reject(new Error('Aborted'));
-
 				const args: string[] = [
 					join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'),
 					<string>this.ai.options.path,
@@ -290,7 +332,6 @@ class LLM {
 				const proc = spawn('node', args, {stdio: ['pipe', 'pipe', 'ignore']});
 				proc.stdin.write(text);
 				proc.stdin.end();
-
 				let output = '';
 				proc.stdout.on('data', (data: Buffer) => output += data.toString());
 				proc.on('close', (code: number) => {
@@ -300,7 +341,7 @@ class LLM {
 							const result = JSON.parse(output);
 							resolve(result.embedding);
 						} catch(err) {
-							reject(new Error('Failed to parse embedding output'));
+							reject(err);
 						}
 					} else {
 						reject(new Error(`Embedder process exited with code ${code}`));
@@ -320,7 +361,7 @@ class LLM {
 			}
 			return results;
 		})();
-		return Object.assign(p, { abort });
+		return <any>Object.assign(p, {abort});
 	}
 
 	/**
@@ -346,8 +387,8 @@ class LLM {
 				(char.charCodeAt(0) * (index + 1)) % dimensions / dimensions).slice(0, dimensions);
 		}
 		const v = vector(target);
-		const similarities = searchTerms.map(t => vector(t)).map(refVector => this.cosineSimilarity(v, refVector))
-		return {avg: similarities.reduce((acc, s) => acc + s, 0) / similarities.length, max: Math.max(...similarities), similarities}
+		const similarities = searchTerms.map(t => vector(t)).map(refVector => this.cosineSimilarity(v, refVector));
+		return {avg: similarities.reduce((acc, s) => acc + s, 0) / similarities.length, max: Math.max(...similarities), similarities};
 	}
 
 	/**
@@ -358,7 +399,7 @@ class LLM {
 	 * @returns {Promise<{} | {} | RegExpExecArray | null>}
 	 */
 	async json(text: string, schema: string, options?: LLMRequest): Promise<any> {
-		let system =  `Your job is to convert input to JSON using tool calls. Call the \`submit\` tool at least once with JSON matching this schema:\n\`\`\`json\n${schema}\n\`\`\`\n\nResponses are ignored`;
+		let system = `Your job is to convert input to JSON using tool calls. Call the \`submit\` tool at least once with JSON matching this schema:\n\`\`\`json\n${schema}\n\`\`\`\n\nResponses are ignored`;
 		if(options?.system) system += '\n\n' + options.system;
 		return new Promise(async (resolve, reject) => {
 			let done = false;
@@ -392,7 +433,7 @@ class LLM {
 	 * @returns {Promise<string>} Summary
 	 */
 	async summarize(text: string, length: number = 500, options?: LLMRequest): Promise<string | null> {
-		let system =  `Your job is to summarize the users message using tool calls. Call the \`submit\` tool at least once with the shortest summary possible that's <= ${length} words. The tool call will respond with the token count. Responses are ignored`;
+		let system = `Your job is to summarize the users message using tool calls. Call the \`submit\` tool at least once with the shortest summary possible that's <= ${length} words. The tool call will respond with the token count. Responses are ignored`;
 		if(options?.system) system += '\n\n' + options.system;
 		return new Promise(async (resolve, reject) => {
 			let done = false;

src/memory.ts

@@ -0,0 +1,177 @@
+// memory.ts
+import {LLMRequest, LLMMessage} from './llm.ts';
+
+/** Background information the AI will be fed as a knowledge document */
+export type Memory = {
+	/** Memory subject */
+	name: string;
+	/** Short description of what this document contains - used for RAG retrieval */
+	description: string;
+	/** Full markdown content of the document */
+	content: string;
+	/** Embedding vector of the description - used for similarity search */
+	embedding: number[];
+}
+
+export type MemoryCollection = {
+	/** Memory subject */
+	name: string;
+	/** Short description - required if isNew */
+	description?: string;
+	/** Extracted facts to merge */
+	facts: string[];
+}
+
+export class MemoryManager {
+
+	tools = {
+		edit: (memory: Memory) => ({
+			name: 'edit',
+			description: 'Edit a memory. Omit start/end to append. Pass start only to replace from that line on. Pass start+end to replace a specific range. start=0 replaces the whole document.',
+			args: {
+				content: {type: 'string', description: 'New content', required: true},
+				start: {type: 'number', description: 'First line to replace (0-indexed, inclusive). Omit to append.'},
+				end: {type: 'number', description: 'Last line to replace (0-indexed, inclusive). Omit to replace from start to end of doc.'},
+			},
+			fn: (args: any) => {
+				const lines = memory.content ? memory.content.split('\n') : [];
+				const newLines = args.content.split('\n');
+				if(args.start === undefined) lines.push(...newLines);
+				else if(args.end === undefined) lines.splice(args.start, lines.length - args.start, ...newLines);
+				else lines.splice(args.start, args.end - args.start + 1, ...newLines);
+				memory.content = lines.join('\n');
+				return `Updated memory:\n${memory.content}`;
+			}
+		}),
+		extract: (pools: MemoryCollection[]) => ({
+			name: 'extract',
+			description: 'Extract a list of facts to group into a single memory',
+			args: {
+				name: {type: 'string', description: 'Exact name of an existing memory, or a new name if none fits ([pro]nouns only)', required: true},
+				description: {type: 'string', description: 'One sentence description of the memory subject, only required if new'},
+				facts: {type: 'string', description: 'Comma separated list of extracted facts', required: true},
+			},
+			fn: (args: any) => {
+				pools.push({
+					name: args.name,
+					description: args.description,
+					facts: args.facts.split(',').map((f: string) => f.trim()).filter(Boolean),
+				});
+				return 'Success';
+			}}),
+		read: (memories: Memory[]) => ({
+			name: 'read',
+			description: 'Read entire memory',
+			args: {
+				name: {type: 'string', description: 'Exact memory name', required: true},
+			},
+			fn: (args: any) => {
+				const mem = memories.find(m => m.name === args.name);
+				if(!mem) return 'Document not found';
+				return `Name: ${mem.name}\nDescription: ${mem.description}\n\n${mem.content}`;
+			}
+		}),
+	}
+
+	constructor(private llm: any, private model?: string) {}
+
+	/**
+	 * Extracts facts from conversation and groups them into individual memories
+	 * @param {string} conversation Full conversation formatted as [role]: content
+	 * @param {Memory[]} memories The user's memory documents
+	 * @param {LLMRequest} options LLM options
+	 * @returns {Promise<MemoryCollection[]>} Fact pools grouped by target document
+	 */
+	private async extract(conversation: string, memories: Memory[], options: LLMRequest): Promise<MemoryCollection[]> {
+		const existingDocs = memories.map(m => `Name: ${m.name}\nDescription: ${m.description}`).join('\n\n');
+		const pools: MemoryCollection[] = [];
+		await this.llm.ask(conversation, {
+			model: this.model || options.model,
+			temperature: 0.2,
+			system: `You are a fact extractor. Analyze this conversation and extract facts worth remembering long term.
+Rules:
+- ONLY extract facts the USER explicitly stated about themselves or their business
+- ONLY extract decisions that were MADE during this conversation
+- DO NOT extract anything the AI said, its name, capabilities, or how it introduced itself
+- DO NOT extract greetings, pleasantries or generic exchanges
+- If nothing worth remembering was said, call NO tools
+
+For each fact decide whether it belongs in an existing document or needs a new one, then call the \`extract\` tool.
+
+Existing documents:\n${existingDocs || 'None yet.'}`,
+			tools: [this.tools.extract(pools)]
+		});
+		return pools;
+	}
+
+	/**
+	 * Bot 2 - Editor: merges a pool of facts into a specific document using surgical line-based edits.
+	 * Receives full document content and uses read + amend tools to make precise edits.
+	 * @param {MemoryCollection} newMem The fact pool to merge
+	 * @param {Memory[]} memories The user's memory documents
+	 * @param {LLMRequest} options LLM options
+	 */
+	private async edit(newMem: MemoryCollection, memories: Memory[], options: LLMRequest): Promise<void> {
+		const existing = memories.find(m => m.name === newMem.name);
+		const mem: Memory = existing || {name: newMem.name, description: newMem.description || '', content: '', embedding: []};
+		const isNew = !existing;
+
+		await this.llm.ask(newMem.facts.map(f => `- ${f}`).join('\n'),
+			{
+				model: this.model || options.model,
+				temperature: 0.2,
+				system: `You are a document editor. Merge the users list of facts into the following document using the \`edit\` tool; call it as many times as necessary.
+
+Name: ${mem.name}
+Description: ${mem.description}
+${mem.content}`,
+				tools: [this.tools.edit(mem)]
+			}
+		);
+
+		if(isNew || mem.description !== existing?.description) {
+			const [e] = await this.llm.embedding(mem.description);
+			mem.embedding = e.embedding;
+		}
+
+		if(isNew) memories.push(mem);
+		else {
+			const idx = memories.findIndex(m => m.name === newMem.name);
+			if(idx >= 0) memories[idx] = mem;
+		}
+	}
+
+	/**
+	 * Find relevant memory documents for a query using description embeddings
+	 * @param {string} query The query to search against
+	 * @param {Memory[]} memories The user's memory documents
+	 * @param {number} limit Max number of results to return
+	 * @returns {Promise<Memory[]>} The most relevant memory documents
+	 */
+	async recollect(query: string, memories: Memory[], limit = 5): Promise<Memory[]> {
+		const [e] = await this.llm.embedding(query);
+		return memories
+			.filter(m => m.embedding?.length)
+			.map(m => ({...m, score: this.llm.cosineSimilarity(m.embedding, e.embedding)}))
+			.toSorted((a: any, b: any) => b.score - a.score)
+			.slice(0, limit);
+	}
+
+	/**
+	 * Two-stage memory pipeline: classify facts from conversation history then surgically merge them into documents.
+	 * Bot 1 (classify) extracts and groups facts cheaply. Bot 2 (edit) runs per-document in parallel with full content access.
+	 * @param {LLMMessage[]} history Full conversation history to digest
+	 * @param {Memory[]} memories The user's memory documents — mutated in place
+	 * @param {LLMRequest} options LLM options
+	 */
+	async memorize(history: LLMMessage[], memories: Memory[], options: LLMRequest): Promise<void> {
+		const conversation = history
+			.filter(h => h.role === 'user' || h.role === 'assistant')
+			.map(h => `[${h.role}]: ${h.content}`)
+			.join('\n\n');
+		if(!conversation.trim()) return;
+		const pools = await this.extract(conversation, memories, options);
+		if(!pools.length) return;
+		await Promise.all(pools.map(pool => this.edit(pool, memories, options)));
+	}
+}

src/open-ai.ts

@@ -138,6 +138,7 @@ export class OpenAi extends LLMProvider {
 					}
 				}
 
+				if(resp.error) throw new Error(resp.error);
 				const toolCalls = resp.choices[0].message.tool_calls || [];
 				if(toolCalls.length && !controller.signal.aborted) {
 					history.push(resp.choices[0].message);
@@ -157,7 +158,7 @@ export class OpenAi extends LLMProvider {
 					requestParams.messages = history;
 				}
 			} while (!controller.signal.aborted && resp.choices?.[0]?.message?.tool_calls?.length);
-			history.push({role: 'assistant', content: resp.choices[0].message.content || ''});
+			history.push({role: 'assistant', content: resp.choices[0].message.content.trim() || ''});
 			history = this.toStandard(history);
 
 			if(options.stream) options.stream({done: true});

src/vision.ts

@@ -17,7 +17,7 @@ export class Vision {
 			const {data} = await worker.recognize(path);
 			await worker.terminate();
 			res(data.text.trim() || null);
-		});
+		}).finally(() => worker?.terminate());
 		return Object.assign(p, {abort: () => worker?.terminate()});
 	}
 }