max token rename

README.md

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

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "1.2.13",
+	"version": "1.5.0",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",

src/antrhopic.ts

@@ -1,62 +1,56 @@
 import {Anthropic as anthropic} from '@anthropic-ai/sdk';
-import {findByProp, objectMap, JSONSanitize, JSONAttemptParse} from '@ztimson/utils';
+import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, makeArray} from '@ztimson/utils';
 import {AbortablePromise, Ai} from './ai.ts';
 import {LLMMessage, LLMRequest} from './llm.ts';
 import {LLMProvider} from './provider.ts';
+import {TokenPool} from './token-pool.ts';
 import {convertSchema} from './tools.ts';
 
 export class Anthropic extends LLMProvider {
-	client!: anthropic;
+	private clients = new Map<string, anthropic>();
+	tokenPool!: TokenPool;
 
-	constructor(public readonly ai: Ai, public readonly apiToken: string, public model: string) {
+	constructor(public readonly ai: Ai, public readonly apiToken: string | string[], public model: string) {
 		super();
-		this.client = new anthropic({apiKey: apiToken});
+		this.tokenPool = new TokenPool(...makeArray(apiToken).filter(Boolean));
 	}
 
-	private toStandard(history: any[]): LLMMessage[] {
-		const timestamp = Date.now();
-		const messages: LLMMessage[] = [];
-		for(let h of history) {
-			if(typeof h.content == 'string') {
-				messages.push(<any>{timestamp, ...h});
-			} else {
-				const textContent = h.content?.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
-				if(textContent) messages.push({role: h.role, content: textContent, timestamp: timestamp});
-				h.content.forEach((c: any) => {
-					if(c.type == 'tool_use') {
-						messages.push({role: 'tool', id: c.id, name: c.name, args: c.input, timestamp: c.timestamp, content: undefined});
-					} else if(c.type == 'tool_result') {
-						const m: any = messages.findLast(m => (<any>m).id == c.tool_use_id);
-						if(m) m[c.is_error ? 'error' : 'content'] = c.content;
+	private getClient(token: string): anthropic {
+		let client = this.clients.get(token);
+		if(!client) {
+			client = new anthropic({apiKey: token});
+			this.clients.set(token, client);
 		}
-				});
-			}
-		}
-		return messages;
+		return client;
 	}
 
-	private fromStandard(history: LLMMessage[]): any[] {
-		for(let i = 0; i < history.length; i++) {
-			if(history[i].role == 'tool') {
-				const h: any = history[i];
-				history.splice(i, 1,
+	/** Convert standard history -> Anthropic wire format */
+	private toWire(history: LLMMessage[]): any[] {
+		const wire: any[] = [];
+		for(const h of history) {
+			if(h.role === 'tool') {
+				wire.push(
 					{role: 'assistant', content: [{type: 'tool_use', id: h.id, name: h.name, input: h.args}]},
-					{role: 'user', content: [{type: 'tool_result', tool_use_id: h.id, is_error: !!h.error, content:  h.error || h.content}]}
-				)
-				i++;
+					{role: 'user', content: [{type: 'tool_result', tool_use_id: h.id, is_error: !!h.error, content: h.error || h.content || ''}]}
+				);
+			} else {
+				wire.push({role: h.role, content: h.content});
 			}
 		}
-		return history;
+		return wire;
 	}
 
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<string | any> {
 		const controller = new AbortController();
-		return Object.assign(new Promise<any>(async (res) => {
-			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
+		return Object.assign(new Promise<any>(async (res, rej) => {
+			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,
-				max_tokens: options.max_tokens || this.ai.options.llm?.max_tokens || 4096,
+				max_tokens: options.maxTokens || this.ai.options.llm?.maxTokens || 4096,
 				system: options.system || this.ai.options.llm?.system || '',
 				temperature: options.temperature || this.ai.options.llm?.temperature || undefined,
 				tools: tools.map(t => ({
@@ -66,95 +60,97 @@ export class Anthropic extends LLMProvider {
 						type: 'object',
 						properties: t.args ? objectMap(t.args, (key, value) => ({...value, required: undefined})) : {},
 						required: t.args ? Object.entries(t.args).filter(t => t[1].required).map(t => t[0]) : []
-					},
-					fn: undefined
+					}
 				})),
-				messages: history,
 				stream: !!options.stream,
 			};
 
-			// Add structured output support
 			if(options.schema) {
-				requestParams.output_config = {
-					format: {
-						type: 'json_schema',
-						schema: convertSchema(options.schema)
-					}
-				};
+				requestParams.output_config = {format: {type: 'json_schema', schema: convertSchema(options.schema)}};
 			}
 
-			let resp: any, isFirstMessage = true, terminal = false;
+			try {
+				let terminal = false;
 				do {
-				requestParams.messages = history.map(({timestamp, ...m}) => m);
-				resp = await this.client.messages.create(requestParams).catch(err => {
-					err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
+					requestParams.messages = this.toWire(history.filter(h => h.role !== 'system'));
+
+					const callStart = Date.now();
+					const resp: any = await this.tokenPool.run(token => this.getClient(token).messages.create(requestParams)).catch(err => {
+						err.message += `\n\nMessages:\n${JSON.stringify(requestParams.messages, null, 2)}`;
 						throw err;
 					});
 
-				// Streaming mode
+					let usage: any, content: any[] = [];
 					if(options.stream) {
-					if(!isFirstMessage) options.stream({text: '\n\n'});
-					else isFirstMessage = false;
-					resp.content = [];
 						for await (const chunk of resp) {
 							if(controller.signal.aborted) break;
 							if(chunk.type === 'content_block_start') {
-							if(chunk.content_block.type === 'text') {
-								resp.content.push({type: 'text', text: ''});
-							} else if(chunk.content_block.type === 'tool_use') {
-								resp.content.push({type: 'tool_use', id: chunk.content_block.id, name: chunk.content_block.name, input: <any>''});
-							}
+								if(chunk.content_block.type === 'text') content.push({type: 'text', text: ''});
+								else if(chunk.content_block.type === 'tool_use') content.push({type: 'tool_use', id: chunk.content_block.id, name: chunk.content_block.name, input: ''});
 							} else if(chunk.type === 'content_block_delta') {
 								if(chunk.delta.type === 'text_delta') {
-								const text = chunk.delta.text;
-								resp.content.at(-1).text += text;
-								options.stream({text});
+									content.at(-1).text += chunk.delta.text;
+									options.stream({text: chunk.delta.text});
 								} else if(chunk.delta.type === 'input_json_delta') {
-								resp.content.at(-1).input += chunk.delta.partial_json;
+									content.at(-1).input += chunk.delta.partial_json;
 								}
 							} else if(chunk.type === 'content_block_stop') {
-							const last = resp.content.at(-1);
-							if(last?.input != null) last.input = last.input ? JSONAttemptParse(last.input, {}) : {};
+								const last = content.at(-1);
+								if(last?.type === 'tool_use') last.input = last.input ? JSONAttemptParse(last.input, {}) : {};
+							} else if(chunk.type === 'message_delta') {
+								if(chunk.usage) usage = chunk.usage;
 							} else if(chunk.type === 'message_stop') {
 								break;
 							}
 						}
+					} else {
+						usage = resp.usage;
+						content = resp.content;
 					}
+					const duration = Date.now() - callStart;
+					const tps = usage?.output_tokens && duration > 0 ? usage.output_tokens / (duration / 1000) : 0;
 
-				// Run tools
-				const toolCalls = resp.content.filter((c: any) => c.type === 'tool_use');
+					const toolCalls = content.filter((c: any) => c.type === 'tool_use');
 					if(toolCalls.length && !controller.signal.aborted) {
-					history.push({role: 'assistant', content: resp.content, timestamp: Date.now()});
-					const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
-						const tool = tools.find(findByProp('name', toolCall.name));
-						if(options.stream) options.stream({tool: toolCall.name});
-						if(!tool) return {tool_use_id: toolCall.id, is_error: true, content: 'Tool not found'};
+						const text = content.filter((c: any) => c.type === 'text').map((c: any) => c.text).join('\n\n').trim();
+						if(text) history.push({role: 'assistant', content: text, timestamp: Date.now(), duration, tps});
+
+						const entries = toolCalls.map((tc: any) => {
+							const entry: any = {role: 'tool', id: tc.id, name: tc.name, args: tc.input, content: undefined, timestamp: Date.now()};
+							history.push(entry);
+							return {tc, entry};
+						});
+
+						await Promise.all(entries.map(async ({tc, entry}: any) => {
+							const tool = tools.find(findByProp('name', tc.name));
+							if(options.stream) options.stream({tool: tc.name});
+							if(!tool) { entry.error = 'Tool not found'; return; }
 							try {
-							// Wrap stream so a tool's `done` ends turn gracefully
 								const toolStream = options.stream && ((chunk: any) => {
 									if(chunk.done) { terminal = true; return; }
 									options.stream!(chunk);
 								});
-							const result = await tool.fn(toolCall.input, toolStream, this.ai);
-							return {type: 'tool_result', tool_use_id: toolCall.id, content: typeof result == 'object' ? JSONSanitize(result) : result};
-						} catch (err: any) {
-							return {type: 'tool_result', tool_use_id: toolCall.id, is_error: true, content: err?.message || err?.toString() || 'Unknown'};
+								const result = await tool.fn(entry.args, toolStream, this.ai, tc.id);
+								entry.content = typeof result === 'object' ? JSONSanitize(result) : result;
+							} catch(err: any) {
+								entry.error = err?.message || err?.toString() || 'Unknown';
 							}
 						}));
-					history.push({role: 'user', content: results, timestamp: Date.now()});
-					requestParams.messages = history;
+					} else {
+						terminal = true;
+						const text = content.filter((c: any) => c.type === 'text').map((c: any) => c.text).join('\n\n').trim();
+						if(text) history.push({role: 'assistant', content: text, timestamp: Date.now(), duration, tps});
 					}
-			} while (!terminal && !controller.signal.aborted && resp.content.some((c: any) => c.type === 'tool_use'));
+				} while(!terminal && !controller.signal.aborted);
 
-			if(!terminal) {
-				const textContent = resp.content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
-				history.push({role: 'assistant', content: textContent, timestamp: Date.now()});
-			}
-			history = this.toStandard(history);
 				if(options.stream) options.stream({done: true});
-			if(options.history) options.history.splice(0, options.history.length, ...history);
-			const finalContent = history.at(-1)?.content;
+
+				const turnStart = history.map(h => h.role).lastIndexOf('user');
+				const finalContent = history.slice(turnStart + 1).reduce((str, h) => h.role === 'assistant' ? str + (h.content || '') : str, '').trim();
 				res(options.schema ? JSONAttemptParse(finalContent, finalContent) : finalContent);
+			} catch(err) {
+				rej(err);
+			}
 		}), {abort: () => controller.abort()});
 	}
 }

src/helpers.ts

@@ -1,21 +1,39 @@
-import {MemoryCache} from './memory-cache.ts';
-import {extractMetadata, Memory, MemoryNode} from './memory.ts';
+import {Memory, MemoryCache} from './memory.ts';
 
-export function buildMemoryGraph(memories: Memory[] | MemoryCache): MemoryNode[] {
+export type MemoryNode = {
+	name: string;
+	missing: boolean;
+	links: string[];
+	backlinks: string[];
+}
+
+export function extractLinks(content: string): string[] {
+	if (!content) return [];
+	const matches = content.matchAll(/\[\[([^\]|]+)(?:\|[^\]]*)?\]\]/g);
+	return [...new Set([...matches].map(m => m[1].trim()))];
+}
+
+export function rebuildGraph(memories: Memory[] | MemoryCache): MemoryNode[] {
 	const mems = memories instanceof MemoryCache ? memories.memories : memories;
 	const nameSet = new Set(mems.map(m => m.name));
-	const ghosts = new Set<string>();
 
-	const nodes: MemoryNode[] = mems.map(m => {
-		const {links, backlinks} = extractMetadata(m.content);
-		return {
+	for (const m of mems) m.links = extractLinks(m.content).filter(l => l !== m.name);
+	for (const m of mems) m.backlinks = [];
+	for (const m of mems) {
+		for (const link of m.links) {
+			const target = mems.find(t => t.name === link);
+			if (target) target.backlinks.push(m.name);
+		}
+	}
+
+	const nodes: MemoryNode[] = mems.map(m => ({
 		name: m.name,
 		missing: false,
-			links,
-			backlinks,
-		};
-	});
+		links: m.links,
+		backlinks: m.backlinks,
+	}));
 
+	const ghosts = new Set<string>();
 	for (const node of nodes) {
 		for (const link of node.links) {
 			if (!nameSet.has(link)) ghosts.add(link);
@@ -28,30 +46,28 @@ export function buildMemoryGraph(memories: Memory[] | MemoryCache): MemoryNode[]
 			name,
 			missing: true,
 			links: [],
-			backlinks: nodes
-				.filter(n => n.links.includes(name))
-				.map(n => n.name),
-		}))
+			backlinks: nodes.filter(n => n.links.includes(name)).map(n => n.name),
+		})),
 	];
 }
 
-export function renderMemoryGraph(nodes) {
+export function renderMemoryGraph(nodes: MemoryNode[]): string {
 	if (!nodes.length) return 'No memories yet.';
 
-	const groups = new Map();
+	const groups = new Map<string, (MemoryNode & {label: string})[]>();
 	for (const node of nodes) {
 		const [prefix, ...rest] = node.name.split('/');
 		const group = rest.length ? prefix : 'Root';
 		const label = rest.length ? rest.join('/') : node.name;
 		if (!groups.has(group)) groups.set(group, []);
-		groups.get(group).push({...node, label});
+		groups.get(group)!.push({...node, label});
 	}
 
 	const ghostCount = nodes.filter(n => n.missing).length;
 	const lines = [`Memory Graph (${nodes.length} nodes, ${ghostCount} ghost${ghostCount === 1 ? '' : 's'})`, ''];
 
 	for (const group of [...groups.keys()].sort()) {
-		const items = groups.get(group).sort((a, b) => a.label.localeCompare(b.label));
+		const items = groups.get(group)!.sort((a, b) => a.label.localeCompare(b.label));
 		lines.push(`${group}/`);
 		items.forEach((n, i) => {
 			const last = i === items.length - 1;

src/index.ts

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

src/kd-tree.ts

@@ -103,9 +103,10 @@ class BoundedMaxHeap<T> {
 export class KDTree<T = unknown> {
 	private root: KDNode<T> | null = null;
 	private _size = 0;
-	private readonly dims: number;
 	private readonly distanceFn: (a: number[], b: number[]) => number;
 
+	readonly dims: number;
+
 	/**
 	 * @param dims           Dimensionality of all vectors (must be consistent).
 	 * @param metric         Distance metric to use. Default: "euclidean".

src/llm.ts

@@ -1,16 +1,31 @@
+import {clean, makeUnique, snakeCase} from '@ztimson/utils';
 import {AbortablePromise, Ai} from './ai.ts';
 import {Anthropic} from './antrhopic.ts';
-import {MemoryCache} from './memory-cache.ts';
 import {OpenAi} from './open-ai.ts';
 import {LLMProvider} from './provider.ts';
 import {AiTool, AiToolArg} from './tools.ts';
 import {fileURLToPath} from 'url';
 import {dirname, join} from 'path';
 import {spawn} from 'node:child_process';
-import {Memory, MemoryManager} from './memory.ts';
+import {Memory, MemoryCache, MemoryManager, MemoryOptions, stripHeader} from './memory.ts';
 
-export type AnthropicConfig = {proto: 'anthropic', token: string};
-export type OpenAiConfig = {proto: 'openai', host?: string, token: string};
+const MAX_AGENT_DEPTH = 5;
+
+export type AnthropicConfig = {proto: 'anthropic', token: string | string[]};
+export type OpenAiConfig = {proto: 'openai', host?: string, token: string | string[]};
+
+export type Agent = {
+	name: string;
+	description?: string;
+	model?: string | null;
+	temperature?: number;
+	system: string;
+	delegate?: boolean;
+	skills?: Skill[] | null;
+	tools?: AiTool[] | null;
+	mcp?: McpServer[] | null;
+	agents?: string[] | null;
+}
 
 export type LLMMessage = {
 	/** Message originator */
@@ -19,6 +34,10 @@ export type LLMMessage = {
 	content: string | any;
 	/** Timestamp */
 	timestamp?: number;
+	/** Response duration in ms */
+	duration?: number;
+	/** Tokens per second */
+	tps?: number;
 } | {
 	/** Tool call */
 	role: 'tool';
@@ -34,6 +53,10 @@ export type LLMMessage = {
 	error?: undefined | string;
 	/** Timestamp */
 	timestamp?: number;
+	/** Response duration in ms */
+	duration?: number;
+	/** Tokens per second */
+	tps?: number;
 }
 
 export type LLMRequest = {
@@ -44,7 +67,7 @@ export type LLMRequest = {
 	/** Message history */
 	history?: LLMMessage[];
 	/** Max tokens for request */
-	max_tokens?: number;
+	maxTokens?: number;
 	/** 0 = Rigid Logic, 1 = Balanced, 2 = Hyper Creative **/
 	temperature?: number;
 	/** Available tools */
@@ -56,13 +79,17 @@ export type LLMRequest = {
 	/** Compress old messages in the chat to free up context */
 	compress?: {max: number; min: number};
 	/** User's memory documents - RAG injected automatically each turn */
-	memory?: Memory[] | MemoryCache;
+	memory?: Memory[] | MemoryCache | MemoryOptions;
 	/** 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[];
+	/** Subagents exposed as delegatable/wrapped tools */
+	agents?: Agent[];
+	/** @internal recursion guard for nested agent delegation */
+	_agentDepth?: number;
 }
 
 export type McpServer = {
@@ -83,7 +110,6 @@ export type Skill = {
 	content: string;
 }
 
-
 class LLM {
 	private memoryManager!: MemoryManager;
 
@@ -100,6 +126,54 @@ class LLM {
 		this.memoryManager = new MemoryManager(this);
 	}
 
+	private setupAgent(agents: Agent[] = [], allAgents: Agent[], history: LLMMessage[], aborts: (() => void)[], depth = 0, delegateState: {resp: string | null}): AiTool[] {
+		return agents.map(a => {
+			const toolName = `${a.delegate ? '' : 'sub'}agent_${snakeCase(a.name)}`;
+			return {
+				name: toolName,
+				description: `${a.delegate ? 'Delegate to ' : ''}Subagent: ${a.description || a.name}`,
+				args: clean<any>({
+					context: !a.delegate ? {type: 'string', description: 'Summary of related messages, samples, files, etc...', required: true} : undefined,
+					instructions: {type: 'string', description: 'Detailed instructions for subagent to complete', required: true},
+				}),
+				fn: async (args: any, stream: any, ai: any, id?: string) => {
+					if(depth >= MAX_AGENT_DEPTH) return 'Max agent delegation depth exceeded';
+
+					const nested = (a.agents || [])
+						.map(name => allAgents.find(x => x.name === name))
+						.filter((x): x is Agent => !!x && x.name !== a.name);
+
+					const q = a.delegate ? '' : `${args.instructions}${args.context ? `\n\n<context>${args.context}</context>` : ''}`;
+
+					const request = this.ask(q, {
+						system: `You are a specialized subagent being called from an orchestrator
+${a.delegate ? 'Your output streams directly to the user for the remainder of this turn. You are mid conversation' : 'You are wrapped in a tool call that will be analysis by an LLM'}
+Dispense with greetings and focus on your instructions using available tools and returning only the final result unless specifically instructed to converse
+
+${a.system}`,
+						model: a.model || undefined,
+						temperature: a.temperature,
+						stream: a.delegate ? stream : undefined,
+						history: a.delegate ? history : [],
+						mcp: a.mcp || undefined,
+						skills: a.skills || undefined,
+						tools: a.tools || undefined,
+						agents: nested,
+						_agentDepth: depth + 1,
+					} as any);
+					aborts.push(request.abort);
+					const resp = await request;
+
+					if(a.delegate) {
+						delegateState.resp = resp;
+						return '';
+					}
+					return resp;
+				}
+			};
+		});
+	}
+
 	private async setupMcp(servers: McpServer[] = []): Promise<{prompt: string, tools: AiTool[]}> {
 		if(!servers?.length) return {prompt: '', tools: []};
 		const allTools: AiTool[] = [];
@@ -133,7 +207,7 @@ class LLM {
 
 		const list = allTools.map(t => `- ${t.name}: ${t.description}`).join('\n');
 		return {
-			prompt: `You have access to the following MCP tools:\n${list}`,
+			prompt: `## MCP\nYou have access to the following MCP tools:\n${list}`,
 			tools: allTools
 		};
 	}
@@ -142,7 +216,7 @@ class LLM {
 		if(!skills?.length) return {prompt: '', tools: []};
 		const list = skills.map(s => `- ${s.name}: ${s.description}`).join('\n');
 		return {
-			prompt: `You have access to the following skill documents, use \`read_skill\` to access them:\n${list}`,
+			prompt: `## Skills\nYou have access to the following skill documents, whenever there is overlap between a question and a skill file, use \`skill_read\` to get instructions and background knowledge:\n${list}`,
 			tools: [{
 				name: 'skill_read',
 				description: 'Read the full content of a skill/knowledge document',
@@ -158,6 +232,20 @@ class LLM {
 		}
 	}
 
+	private wrapToolTiming(tools: AiTool[], timings: Map<string, {duration: number, tps: number}>): AiTool[] {
+		return tools.map(t => ({
+			...t,
+			fn: async (args: any, stream: any, ai: any, id?: string) => {
+				const start = Date.now();
+				const result = await t.fn(args, stream, ai, id);
+				const duration = Date.now() - start;
+				const tps = duration > 0 ? this.estimateTokens(result) / (duration / 1000) : 0;
+				if(id) timings.set(id, {duration, tps});
+				return result;
+			}
+		}));
+	}
+
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
 		options = <any>{
 			system: '',
@@ -170,15 +258,21 @@ class LLM {
 		if(!this.models[m]) throw new Error(`Model does not exist: ${m}`);
 		let request: AbortablePromise<string> | null = null;
 		let aborted = false;
+		const nestedAborts: (() => void)[] = [];
 		const abort = () => {
 			aborted = true;
 			request?.abort?.();
+			nestedAborts.forEach(a => a());
 		};
 
-		const promise = (async () => {
+		let promise: any;
+		const requestStart = Date.now();
+
+		promise = (async () => {
 			let tools: AiTool[] = options.tools || this.ai.options.llm?.tools || [];
 			const prompts: string[] = [];
 			let history = options.history || [];
+			if(message) history.push({role: 'user', content: message, timestamp: Date.now()});
 
 			// MCP
 			const mcp = options.mcp || this.ai.options?.llm?.mcp;
@@ -196,57 +290,95 @@ class LLM {
 				tools.push(...s.tools);
 			}
 
+			// 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));
+
 			// Memory
-			if (options.memory) {
-				const mems = options.memory instanceof MemoryCache ? options.memory.memories : options.memory;
+			const mem = MemoryManager.normalize(options.memory);
+			if(mem) {
+				const mems = mem.memory instanceof MemoryCache ? mem.memory.memories : mem.memory;
 				if(mems.length) {
-					const relevant = await this.memoryManager.recollect(message, options.memory, 5);
-					prompts.unshift(`You have access to the following memory files:
-	${mems.map(m => `- ${m.name}: ${m.description}`).join('\n')}
-	${relevant.length ? `
-	Relevant memories have been preloaded:
-	${relevant.map(r => `
-	**${r.name}**
-	${r.description}
-	${r.content}
-	`).join('\n---\n')}
-	` : ''}`.trim());
-					tools.push(this.memoryManager.tools.read(options.memory));
+					if(mem.inject) {
+						const pool = 15; // candidates considered, cheap since only refs are listed
+						const budget = mem.maxTokens ?? 2000; // actual content injected
+						const relevant = await this.memoryManager.recollect(message, mem.memory, pool);
+
+						let used = 0;
+						const preloaded: typeof relevant = [];
+						const listed: typeof relevant = [];
+						for(const r of relevant) {
+							const t = this.estimateTokens(r.content);
+							if(used + t <= budget || preloaded.length === 0) {
+								preloaded.push(r);
+								used += t;
+							} else listed.push(r);
+						}
+
+						prompts.unshift(`## Memory
+You have a background memory process which has prefetched relevant information${mem.update ? ' and will create new memories from this conversation' : ''} for you
+Assume it is perfect and never mention this process to anyone ever
+Always use your memories to craft a personalized response, they contain links / [[wiki links]] which you use navigate between them
+${mem.tool ? `You can access memory files via the \`memory_search\` and \`memory_recall\` tools
+When you need information about the user, \`memory_recall\` \`People/User\` before asking (fetch if not included bellow)
+When you need information not provided, attempt 1-3 \`memory_search\` calls with distinct queries before asking` : ''}
+
+${preloaded.length ? `### Prefetched Memories (Most relevant first):
+
+${preloaded.map(r => `Memory: ${r.name}
+Description: ${r.description}
+Linked: ${makeUnique([...r.links, ...r.backlinks]).join(', ')}
+\`\`\`
+${stripHeader(r.content)}
+\`\`\``).join('\n\n')}` : ''}
+${mem.tool && listed.length ? '\n' + listed.map(r => `Memory: ${r.name}
+Description: ${r.description}
+Linked: ${makeUnique([...r.links, ...r.backlinks]).join(', ')}
+<!-- Truncated -->`).join('\n\n') : ''}`.trim())
+					}
+					if(mem.tool) tools.push(this.memoryManager.tools.read(mem.memory));
 				}
 			}
 
 			if(aborted) throw Object.assign(new Error('Aborted'), {name: 'AbortError'});
 
-			prompts.unshift(options.system || this.ai.options.llm?.system || '');
-			request = this.models[m].ask(message, {...options, tools, system: prompts.filter(Boolean).join('\n\n')});
-			const resp = await request;
+			const toolTimings = new Map<string, {duration: number, tps: number}>();
+			tools = this.wrapToolTiming(tools, toolTimings);
 
-			// Trim memory injections from history
-			if(options.memory) {
-				history.splice(0, history.length, ...history.filter(h => h.role !== 'tool' || h.name !== 'recall'));
+			if(aborted) throw Object.assign(new Error('Aborted'), {name: 'AbortError'});
+
+			prompts.unshift(options.system || this.ai.options.llm?.system || '');
+			request = this.models[m].ask('', {...options, tools, system: prompts.filter(Boolean).join('\n\n')});
+			let resp = await request;
+
+			// Capture meta (duration / tps)
+			for(const h of history) {
+				if(h.role === 'tool' && toolTimings.has(h.id)) Object.assign(h, toolTimings.get(h.id));
 			}
 
-			// Auto-memorize before compressing
+			if(typeof resp === 'string' && !resp.trim() && delegateState.resp !== null) resp = delegateState.resp;
+
+			if(mem?.tool) history.splice(0, history.length, ...history.filter(h => h.role !== 'tool' || h.name !== 'memory_recall'));
 			if(options.compress && this.estimateTokens(history) >= options.compress.max) {
-				if(options.memory) await this.memoryManager.memorize(history, options.memory, {model: options.memoryModel || this.defaultModel, ...options});
+				if(mem?.update) await this.memoryManager.memorize(history, mem.memory, {model: options.memoryModel || this.defaultModel, ...options});
 				const compressed = await this.compressHistory(history, options.compress.max, options.compress.min, options);
 				if(options.history) options.history.splice(0, options.history.length, ...compressed);
 			}
 
+			const requestDuration = Date.now() - requestStart;
+			const totalTokens = history
+				.filter((h: any) => h.role === 'assistant' && h.duration && h.tps)
+				.reduce((sum: number, h: any) => sum + h.tps * (h.duration / 1000), 0);
+			const requestTps = requestDuration > 0 ? totalTokens / (requestDuration / 1000) : 0;
+			Object.assign(promise, {duration: requestDuration, tps: requestTps});
+
 			return resp;
 		})();
 
 		return Object.assign(promise, {abort});
 	}
 
-	/**
-	 * Digest full conversation history into memory documents.
-	 * Call on session end to persist the conversation.
-	 */
-	async updateMemory(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest = {}): Promise<Memory[]> {
-		return this.memoryManager.memorize(history, memories, {model: this.defaultModel, ...options});
-	}
-
 	/**
 	 * Compress chat history to reduce context size
 	 * @param {LLMMessage[]} history Chatlog that will be compressed
@@ -399,7 +531,6 @@ class LLM {
 	 */
 	fuzzyMatch(target, ...searchTerms) {
 		if (searchTerms.length < 2) throw new Error('Requires at least 2 strings to compare');
-
 		const levenshtein = (a, b) => {
 			const m = a.length, n = b.length;
 			if (!m) return n;
@@ -415,13 +546,10 @@ class LLM {
 			}
 			return dp[m][n];
 		};
-
 		const similarity = (a, b) => {
 			a = a.toLowerCase(); b = b.toLowerCase();
-			const dist = levenshtein(a, b);
-			return 1 - dist / Math.max(a.length, b.length, 1);
+			return 1 - levenshtein(a, b) / Math.max(a.length, b.length, 1);
 		};
-
 		const similarities = searchTerms.map(t => similarity(target, t));
 		return {
 			avg: similarities.reduce((acc, s) => acc + s, 0) / similarities.length,
@@ -430,6 +558,14 @@ class LLM {
 		};
 	}
 
+	/**
+	 * Digest full conversation history into memory documents.
+	 * Call on session end to persist the conversation.
+	 */
+	async memorize(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest = {}): Promise<Memory[]> {
+		return this.memoryManager.memorize(history, memories, {model: this.defaultModel, ...options});
+	}
+
 	/**
 	 * Create a summary of some text
 	 * @param {string} text Text to summarize

src/memory-cache.ts

@@ -1,59 +0,0 @@
-import {KDPoint, KDTree} from './kd-tree.ts';
-import {Memory, MemoryRef} from './memory.ts';
-
-export class MemoryCache {
-	private tree: KDTree<MemoryRef>;
-	public memories: Memory[];
-
-	get length() { return this.memories.length; }
-
-	constructor(memories: Memory[]) {
-		this.memories = memories;
-		this.tree = this.buildTree();
-	}
-
-	private buildTree(): KDTree<MemoryRef> {
-		const embedded = this.memories.filter(m => m.embedding?.length);
-		if (!embedded.length) return new KDTree<MemoryRef>(0);
-
-		const dims = embedded[0].embedding.length;
-		const points: KDPoint<MemoryRef>[] = embedded.map(m => ({
-			vector: m.embedding,
-			payload: {name: m.name, description: m.description},
-		}));
-
-		return new KDTree<MemoryRef>(dims, 'cosine', points);
-	}
-
-	search(query: number[], limit: number): MemoryRef[] {
-		const results = this.tree.knn(query, limit);
-		return results.map(r => r.point.payload);
-	}
-
-	add(memory: Memory): void {
-		this.memories.push(memory);
-		this.rebuild();
-	}
-
-	update(memory: Memory): void {
-		const idx = this.memories.findIndex(m => m.name === memory.name);
-		if (idx !== -1) {
-			this.memories[idx] = memory;
-		} else {
-			this.memories.push(memory);
-		}
-		this.rebuild();
-	}
-
-	remove(name: string): void {
-		const idx = this.memories.findIndex(m => m.name === name);
-		if (idx !== -1) {
-			this.memories.splice(idx, 1);
-			this.rebuild();
-		}
-	}
-
-	rebuild(): void {
-		this.tree = this.buildTree();
-	}
-}

src/memory.ts

@@ -1,52 +1,35 @@
+import {MemoryNode, rebuildGraph} from './helpers.ts';
 import {LLMRequest, LLMMessage} from './llm.ts';
-import {MemoryCache} from './memory-cache.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[];
-}
-
-export type MemoryRef = {
-	name: string;
-	description: string;
-}
-
-export type FactBucket = {
-	subject: string;
-	facts: string[];
-}
-
-export type MemoryNode = {
-	name: string;
-	missing: boolean;
 	links: string[];
 	backlinks: string[];
 }
 
-function extractLinks(content: string): string[] {
-	if(!content) return [];
-	const matches = content.matchAll(/\[\[([^\]]+)\]\]/g);
-	return [...new Set([...matches].map(m => m[1].trim()))];
+type MemoryRef = {
+	name: string;
+	description: string;
 }
 
-export function extractMetadata(content: string): {links: string[], backlinks: string[]} {
-	const match = content.match(/^---\n([\s\S]*?)\n---/);
-	if (!match) return {links: [], backlinks: []};
-
-	const fm = match[1];
-	const getList = (key: string): string[] => {
-		const m = fm.match(new RegExp(`^${key}:\\s*\\[(.*)\\]$`, 'm'));
-		if (!m || !m[1].trim()) return [];
-		return m[1].split(',').map(s => s.trim().replace(/^"|"$/g, '')).filter(Boolean);
-	};
-
-	return {
-		links: getList('links'),
-		backlinks: getList('backlinks'),
-	};
+type FactBucket = {
+	subject: string;
+	facts: string[];
 }
 
 function dedupeFacts(facts: string[]): string[] {
@@ -69,50 +52,148 @@ function cosineDistance(a: number[], b: number[]): number {
 	return denom === 0 ? 1 : 1 - dot / denom;
 }
 
-function getWeekMonday(date: Date = new Date()): string {
-	const d = new Date(Date.UTC(date.getFullYear(), date.getMonth(), date.getDate()));
-	const day = d.getUTCDay();
-	const diff = day === 0 ? -6 : 1 - day;
-	d.setUTCDate(d.getUTCDate() + diff);
-	return d.toISOString().slice(0, 10);
+function 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);
 }
 
-function getWeekSunday(monday: string): string {
-	const d = new Date(`${monday}T00:00:00Z`);
-	d.setUTCDate(d.getUTCDate() + 6);
-	return d.toISOString().slice(0, 10);
+export 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 pendingMemorizations = new Map<string, {
-		memories: Memory[] | MemoryCache,
-		tempMemoryName: string,
-		timestamp: number,
-	}>();
+	private recentlyTouched = new Map<string, number>();
 
 	private queues = new Map<string, {
-		pending: string[],
+		dirty: boolean,
 		request: {abort?: () => void} | null,
 		task: Promise<void>,
 	}>();
 
 	tools = {
-		read: (memories: Memory[] | MemoryCache): AiTool => ({
-			name: 'memory_recall',
-			description: 'Read the full content of a memory document',
-			args: {
-				name: {type: 'string', description: 'Exact memory name', required: true},
-			},
-			fn: (args: any) => {
-				const mems = memories instanceof MemoryCache ? memories.memories : memories;
-				const mem = mems.find(m => m.name === args.name);
-				if (!mem) return 'Document not found';
-				return mem.content;
-			},
-		}),
-
 		forget: (memories: Memory[] | MemoryCache): AiTool => ({
 			name: 'memory_forget',
 			description: 'Permanently delete a memory document and clean up all references to it',
@@ -124,360 +205,331 @@ export class MemoryManager {
 				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) {}
 
-	private async createTempMemory(conversation: string): Promise<Memory> {
-		const timestamp = Date.now();
-		const content = `---
-name: _temp_${timestamp}
-description: Temporary memory - processing in background
-tags: [_temporary]
-links: []
-backlinks: []
-modified: ${new Date().toISOString()}
----
-
-# Recent Conversation (Processing)
-
-${conversation}`;
-		const [e] = await this.llm.embedding(content);
-		return {
-			name: `_temp_${timestamp}`,
-			description: 'Temporary memory - processing in background',
-			content,
-			embedding: e?.embedding || [],
-		};
+	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};
 	}
 
-	forget(name: string, memories: Memory[] | MemoryCache): boolean {
-		const mem = memories instanceof MemoryCache ? memories.memories : memories;
-		const idx = mem.findIndex(m => m.name === name);
-		if (idx === -1) return false;
+	private access(memories: Memory[] | MemoryCache): MemoryAccessor {
+		return new MemoryAccessor(memories);
+	}
 
-		for (const node of mem) {
-			const {links, backlinks} = extractMetadata(node.content);
-			const newBacklinks = backlinks.filter(b => b !== name);
-			const newLinks = links.filter(l => l !== name);
+	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);
+	}
 
-			if (newBacklinks.length !== backlinks.length || newLinks.length !== links.length) {
-				node.content = this.updateFrontmatter(node.content, {
-					links: newLinks,
-					backlinks: newBacklinks,
+	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);
 		}
 
-		mem.splice(idx, 1);
-
-		if (memories instanceof MemoryCache) memories.rebuild();
-		return true;
+		return buckets.entries().toArray().map(([subject, facts]) => ({subject, facts}));
 	}
 
-	private cosineSearch(query: number[], memories: Memory[], limit: number): MemoryRef[] {
-		const scored = memories
-			.filter(m => m.embedding?.length)
-			.map(m => ({
-				ref: {name: m.name, description: m.description},
-				distance: cosineDistance(query, m.embedding),
-			}))
-			.sort((a, b) => a.distance - b.distance)
-			.slice(0, limit);
-		return scored.map(s => s.ref);
+	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 mem: Memory[] = memories instanceof MemoryCache ? memories.memories : memories;
-		if (!mem.length) return [];
+		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 [];
 
-		let vectorResults: MemoryRef[];
-		if (memories instanceof MemoryCache) vectorResults = memories.search(e.embedding, limit);
-		else vectorResults = this.cosineSearch(e.embedding, mem, limit);
+		const vectorResults = store.search(e.embedding, limit);
 		const found = new Set<string>(vectorResults.map(r => r.name));
 
 		if (graphDepth > 0) {
-			const frontier = [...found];
-			for (let depth = 0; depth < graphDepth; depth++) {
+			let frontier = [...found];
+			for (let depth = 0; depth < graphDepth && frontier.length; depth++) {
 				const next: string[] = [];
 				for (const name of frontier) {
-					const node = mem.find(m => m.name === name);
+					const node = store.find(name);
 					if (!node) continue;
-					const {links} = extractMetadata(node.content);
-					for (const link of links) {
-						if (!found.has(link) && mem.find(m => m.name === link)) {
+					for (const link of node.links) {
+						if (!found.has(link) && store.find(link)) {
 							found.add(link);
 							next.push(link);
 						}
 					}
 				}
-				frontier.splice(0, frontier.length, ...next);
-				if (!frontier.length) break;
+				frontier = next;
 			}
 		}
 
 		const vectorOrder = vectorResults.map(r => r.name);
 		const graphExpansions = [...found].filter(n => !vectorOrder.includes(n));
-		const ordered = [...vectorOrder, ...graphExpansions];
-		return ordered.map(n => mem.find(m => m.name === n)!).filter(Boolean);
+		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 [];
+		if (!conversation) return [];
 
-		const trackingId = `${Date.now()}_${Math.random()}`;
-		const tempMemory = await this.createTempMemory(conversation);
-		const mem = memories instanceof MemoryCache ? memories.memories : memories;
-		mem.push(tempMemory);
-		if (memories instanceof MemoryCache) memories.rebuild();
-		this.pendingMemorizations.set(trackingId, {
-			memories,
-			tempMemoryName: tempMemory.name,
-			timestamp: Date.now(),
-		});
+		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);
 
-		try {
-			await this._memorizeBackground(conversation, memories, options, tempMemory.name);
-			const finalMem = memories instanceof MemoryCache ? memories.memories : memories;
-			return finalMem.filter(m => !m.name.startsWith('_temp_'));
-		} finally {
-			const pending = this.pendingMemorizations.get(trackingId);
-			if (pending) {
-				const cleanMem = pending.memories instanceof MemoryCache
-					? pending.memories.memories
-					: pending.memories;
-				const idx = cleanMem.findIndex(m => m.name === pending.tempMemoryName);
-				if (idx !== -1) cleanMem.splice(idx, 1);
-				if (pending.memories instanceof MemoryCache) pending.memories.rebuild();
-			}
-			this.pendingMemorizations.delete(trackingId);
-		}
-	}
+		const store = this.access(memories);
+		const buckets = await this.factAgent(conversation, store, options, this.getWeekMonday());
+		const touched: Memory[] = [];
 
-	private async _memorizeBackground(conversation: string, memories: Memory[] | MemoryCache, options: LLMRequest, tempName: string): Promise<void> {
-		const mem = memories instanceof MemoryCache ? memories.memories : memories;
-		const monday = getWeekMonday();
-		const sunday = getWeekSunday(monday);
-		const buckets = await this.factAgent(conversation, mem, options, monday);
-		if(!buckets.length) return;
-		const jobs = [...buckets].map(({subject, facts}) => {
-			let node = mem.find(m => m.name === subject);
-			if(!node) {
-				node = {name: subject, description: '', content: '', embedding: [],};
-				mem.push(node);
+		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);
 			}
-			const week = subject.startsWith('Journal/') ? {monday, sunday} : undefined;
-			return this.enqueue(node, facts, mem, options, tempName, week);
-		});
-		await Promise.all(jobs);
-	}
-
-	/**
-	 * Coalescing queue: if a doc is already compiling, abort the in-flight run, merge its
-	 * facts with the new ones and restart. Never blocks a pending update, never drops facts.
-	 */
-	private enqueue(node: Memory, facts: string[], memories: Memory[] | MemoryCache, options: LLMRequest, tempName: string, week?: {monday: string, sunday: string}): Promise<void> {
-		const key = node.name;
-		const existing = this.queues.get(key);
-		if (existing) {
-			existing.pending.push(...facts);
-			existing.request?.abort?.();
-			return existing.task;
-		}
-
-		const entry: {pending: string[], request: {abort?: () => void} | null, task: Promise<void>} = {pending: [...facts], request: null, task: Promise.resolve()};
-		this.queues.set(key, entry);
-		const m = memories instanceof MemoryCache ? memories.memories : memories;
-		entry.task = (async () => {
-			while (entry.pending.length) {
-				const batch = dedupeFacts(entry.pending.splice(0, entry.pending.length));
-				const written = await this.docAgent(node, batch, m, options, tempName, week, entry);
-				if (!written) entry.pending.unshift(...batch);
-			}
-		})().finally(() => {
-			this.queues.delete(key);
-			if(!this.queues.size && memories instanceof MemoryCache) memories.rebuild();
-		});
-		return entry.task;
-	}
-
-	private buildHeader(node: Memory, week?: {monday: string, sunday: string}, links: string[] = [], backlinks: string[] = []): string {
-		const tags = node.name.split('/')[0]?.toLowerCase();
-		const lines = [
-			'---',
-			`name: ${node.name}`,
-			`description: ${node.description || ''}`,
-			tags ? `tags: [${tags}]` : '',
-			links.length ? `links: [${links.map(l => `"${l}"`).join(', ')}]` : 'links: []',
-			backlinks.length ? `backlinks: [${backlinks.map(l => `"${l}"`).join(', ')}]` : 'backlinks: []',
-			week ? `week: ${week.monday} – ${week.sunday}` : '',
-			`modified: ${new Date().toISOString()}`,
-			'---',
-		].filter(Boolean);
-		return lines.join('\n');
-	}
-
-	private applyHeader(content: string, header: string): string {
-		return `${header}\n\n${this.stripHeader(content)}`;
-	}
-
-	private updateFrontmatter(content: string, updates: {links?: string[], backlinks?: string[]}): string {
-		const match = content.match(/^---\n([\s\S]*?)\n---\n\n?([\s\S]*)$/);
-		if (!match) return content;
-
-		const [, fm, body] = match;
-		let newFm = fm;
-
-		if (updates.links !== undefined) {
-			const linksList = updates.links.length ? `[${updates.links.map(l => `"${l}"`).join(', ')}]` : '[]';
-			newFm = newFm.replace(/^links:.*$/m, `links: ${linksList}`);
-		}
-
-		if (updates.backlinks !== undefined) {
-			const backlinksList = updates.backlinks.length ? `[${updates.backlinks.map(l => `"${l}"`).join(', ')}]` : '[]';
-			newFm = newFm.replace(/^backlinks:.*$/m, `backlinks: ${backlinksList}`);
-		}
-
-		newFm = newFm.replace(/^modified:.*$/m, `modified: ${new Date().toISOString()}`);
-
-		return `---\n${newFm}\n---\n\n${body}`;
-	}
-
-	private stripHeader(content: string): string {
-		return content.replace(/^---[\s\S]*?\n---\n?/, '').trimStart();
-	}
-
-	private async docAgent(node: Memory, facts: string[], memories: Memory[], options: LLMRequest, tempName: string, week: {monday: string, sunday: string} | undefined, entry: {request: {abort?: () => void} | null}): Promise<boolean> {
-		const {links: oldLinks} = extractMetadata(node.content);
-		const currentBody = this.stripHeader(node.content);
-		let update;
-		try {
-			for(let i = 0; i < 3 && !update?.content; i++) {
-				const request = this.llm.ask(`New Facts:\n${facts.map(f => `- ${f}`).join('\n')}`, {
-					model: options.model,
-					temperature: 0.3,
-					schema: {
-						description: {type: 'string', description: 'One-line description of what this document covers, no formatting or emojis', required: true},
-						content: {type: 'string', description: 'Rewritten document in markdown, without the frontmatter block', required: true},
-					},
-					system: `You are a knowledge base editor. Rewrite the current document below so it incorporates the new facts.
-
-Formatting rules:
-- Use Obsidian-style markdown: # headings, **bold** to add emphasis, __italics__ for titles, terms, etc, bullet & numbered lists for grouped 1D data and tables for 2D data
-- Link related concepts with [[WikiLink]] notation using full paths like [[People/Sarah]] or [[Projects/Website]]
-- Create links for specific entities (person, place, project, program) and abstract concepts (quantum mechanics, entropy) but skip generics (car, red, dog)
-- Keep the document concise, factual, and human-readable
-- Resolve contradictions: the new facts always win — delete the outdated statement entirely, never keep both
-- Later facts in the list override earlier ones
-- Do not add frontmatter blocks, filler, preamble, or AI commentary
-${week ? '- This is a weekly journal entry.\n' : ''}
-All nodes:
-${this.listNodes(memories).map(n => n.name).join(', ') || 'none'}
-
-Current document:
-\`\`\`markdown
-${currentBody}
-\`\`\``}
-				);
-				entry.request = request;
-				update = await request;
-			}
-		} catch (err: any) {
-			if (err?.name === 'AbortError') return false;
-			throw err;
-		} finally {
-			entry.request = null;
-		}
-
-		if(!update?.content) return false;
-		const newLinks = extractLinks(update.content).filter(l => l !== node.name && l !== tempName);
-		const newLinkSet = new Set(newLinks);
-		const oldLinkSet = new Set(oldLinks);
-
-		for (const added of newLinkSet) {
-			if (!oldLinkSet.has(added)) {
-				const target = memories.find(m => m.name === added);
-				if (target) {
-					const {backlinks} = extractMetadata(target.content);
-					if (!backlinks.includes(node.name)) {
-						target.content = this.updateFrontmatter(target.content, {
-							backlinks: [...backlinks, node.name],
-						});
-					}
-				}
-			}
-		}
-		for (const removed of oldLinkSet) {
-			if (!newLinkSet.has(removed)) {
-				const target = memories.find(m => m.name === removed);
-				if (target) {
-					const {backlinks} = extractMetadata(target.content);
-					target.content = this.updateFrontmatter(target.content, {
-						backlinks: backlinks.filter(b => b !== node.name),
-					});
-				}
-			}
-		}
-
-		const {backlinks} = extractMetadata(node.content);
-		node.description = node.name !== 'Person/User' ? update.description : 'All information about the current user';
-		node.content = this.applyHeader(update.content, this.buildHeader(node, week, newLinks, backlinks));
+			this.appendFacts(node, facts);
 			const [e] = await this.llm.embedding(node.content);
-		if(e) node.embedding = e.embedding;
-		return true;
+			if (e) node.embedding = e.embedding;
+			this.touch(node.name);
+			touched.push(node);
 		}
 
-	private async factAgent(conversation: string, memories: Memory[], options: LLMRequest, weekKey: string): Promise<FactBucket[]> {
-		const buckets = new Map<string, string[]>();
-		await this.llm.ask(conversation, {
-			model: options.model,
-			temperature: 0.2,
-			system: `You are a fact extractor. Analyze this conversation and extract facts worth remembering long-term.
+		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.';
+		}
 
-Rules:
-- ONLY extract current facts the USER explicitly stated about themselves, their work, or their projects
-- ONLY extract decisions that were MADE during this conversation
-- DO NOT extract anything the AI said, its capabilities, or meta-conversation about the AI
-- DO NOT extract greetings, pleasantries, or generic exchanges
-- DO NOT extract deltas or changes in facts; ONLY the end fact
-- If nothing worth remembering was said, do not call any tools
+		(touched as any).uid = uid;
+		return touched;
+	}
 
-When extracting facts, you MUST also decide the exact destination path:
-- Use an existing node name if the facts clearly belong there
-- All information primary about the user should go under "People/User"
-- When required, create a new path following collection/subject format (e.g., People/Sarah, Projects/Oxide)
-- For journal entries, use "Journal"
-
-Available nodes:
-- Journal
-${this.listNodes(memories).filter(n => !n.name.includes('_temp_') && !n.name.includes('Journal')).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}`,
-			tools: [{
-				name: 'facts_extract',
-				description: 'Submit facts with their destination',
-				args: {
-					destination: {type: 'string', description: 'Exact existing node name OR new path (e.g. "People/Sarah", "Projects/Oxide")', required: true},
-					facts: {type: 'string', description: 'Comma-separated facts', required: true},
-				},
-				fn: (args: any) => {
-					const subject = args.destination.trim().toLowerCase() === 'journal'
-						? `Journal/${weekKey}` : args.destination.trim();
-					const facts = buckets.get(subject) ?? [];
-					facts.push(...dedupeFacts(String(args.facts).split(',')));
-					buckets.set(subject, facts);
-					return 'Recorded';
-				},
-			}],
-		});
-		return buckets.entries().toArray().map(([subject, facts]) => ({subject, facts}));
+	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/open-ai.ts

@@ -1,85 +1,64 @@
 import {OpenAI as openAI} from 'openai';
-import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, clean} from '@ztimson/utils';
+import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, clean, makeArray} from '@ztimson/utils';
 import {AbortablePromise, Ai} from './ai.ts';
 import {LLMMessage, LLMRequest} from './llm.ts';
 import {LLMProvider} from './provider.ts';
+import {TokenPool} from './token-pool.ts';
 import {convertSchema} from './tools.ts';
 
 export class OpenAi extends LLMProvider {
-	client!: openAI;
+	tokenPool!: TokenPool;
+	private clients = new Map<string, openAI>();
 
-	constructor(public readonly ai: Ai, public readonly host: string | null, public readonly token: string, public model: string) {
+	constructor(public readonly ai: Ai, public readonly host: string | null, public readonly token: string | string[], public model: string) {
 		super();
-		this.client = new openAI(clean({
-			baseURL: host,
-			apiKey: token || (host ? 'ignored' : undefined)
-		}));
+		const tokens = makeArray(token).filter(Boolean);
+		this.tokenPool = new TokenPool(...(tokens.length ? tokens : [host ? 'ignored' : '']));
 	}
 
-	private toStandard(history: any[]): LLMMessage[] {
-		for(let i = 0; i < history.length; i++) {
-			const h = history[i];
-			if(h.role === 'assistant' && h.tool_calls) {
-				const tools = h.tool_calls.map((tc: any) => ({
-					role: 'tool',
-					id: tc.id,
-					name: tc.function.name,
-					args: JSONAttemptParse(tc.function.arguments, {}),
-					timestamp: h.timestamp
-				}));
-				history.splice(i, 1, ...tools);
-				i += tools.length - 1;
-			} else if(h.role === 'tool') {
-				const record = history.find(h2 => h.tool_call_id == h2.id);
-				if(record) {
-					if(h.content?.includes('"error":')) record.error = h.content;
-					else record.content = h.content || '';
+	private getClient(token: string): openAI {
+		let client = this.clients.get(token);
+		if(!client) {
+			client = new openAI(clean({baseURL: this.host, apiKey: token || undefined}));
+			this.clients.set(token, client);
 		}
-				history.splice(i, 1);
-				i--;
-			}
-			if(!history[i]?.timestamp) history[i].timestamp = Date.now();
-		}
-		return history;
+		return client;
 	}
 
-	private fromStandard(history: LLMMessage[]): any[] {
-		return history.reduce((result, h) => {
+	/** Convert standard history -> OpenAI wire format */
+	private toWire(history: LLMMessage[], system?: string): any[] {
+		const wire: any[] = [];
+		if(system) wire.push({role: 'system', content: system});
+		for(const h of history) {
 			if(h.role === 'tool') {
-				result.push({
+				wire.push({
 					role: 'assistant',
 					content: null,
-					tool_calls: [{ id: h.id, type: 'function', function: { name: h.name, arguments: JSON.stringify(h.args) } }],
-					refusal: null,
-					annotations: [],
-					timestamp: h.timestamp,
+					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,
-					timestamp: h.timestamp,
+					content: h.error || h.content || '',
 				});
 			} else {
-				result.push(h);
+				wire.push({role: h.role, content: h.content});
 			}
-			return result;
-		}, [] as any[]);
+		}
+		return wire;
 	}
 
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<string | any> {
 		const controller = new AbortController();
 		return Object.assign(new Promise<any>(async (res, rej) => {
-			if(options.system) {
-				if(options.history?.[0]?.role != 'system') options.history?.splice(0, 0, {role: 'system', content: options.system, timestamp: Date.now()});
-				else options.history[0].content = options.system;
-			}
-			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
+			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,
-				messages: history,
 				stream: !!options.stream,
-				max_completion_tokens: options.max_tokens || this.ai.options.llm?.max_tokens || undefined,
+				max_completion_tokens: options.maxTokens || this.ai.options.llm?.maxTokens || undefined,
 				temperature: options.temperature || this.ai.options.llm?.temperature || undefined,
 				tools: tools.map(t => ({
 					type: 'function',
@@ -97,97 +76,94 @@ 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};
 
-			let resp: any, isFirstMessage = true, terminal = false;
+			try {
+				let terminal = false;
 				do {
-				requestParams.messages = history.map(({timestamp, ...m}) => m);
-				resp = await this.client.chat.completions.create(requestParams).catch(err => {
-					err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
+					requestParams.messages = this.toWire(history.filter(h => h.role !== 'system'), options.system);
+
+					const callStart = Date.now();
+					const resp: any = await this.tokenPool.run(token => this.getClient(token).chat.completions.create(requestParams)).catch(err => {
+						err.message += `\n\nMessages:\n${JSON.stringify(requestParams.messages, null, 2)}`;
 						throw err;
 					});
 
+					let usage: any, msg: any = {content: '', tool_calls: []};
 					if(options.stream) {
-					if(!isFirstMessage) options.stream({text: '\n\n'});
-					else isFirstMessage = false;
-					resp.choices = [{message: {role: 'assistant', content: '', tool_calls: [], timestamp: Date.now()}}];
 						for await (const chunk of resp) {
 							if(controller.signal.aborted) break;
-						if(chunk.choices[0].delta.content) {
-							resp.choices[0].message.content += chunk.choices[0].delta.content;
+							if(chunk.usage) usage = chunk.usage;
+							if(chunk.choices[0]?.delta?.content) {
+								msg.content += chunk.choices[0].delta.content;
 								options.stream({text: chunk.choices[0].delta.content});
 							}
-
-						if(chunk.choices[0].delta.tool_calls) {
+							if(chunk.choices[0]?.delta?.tool_calls) {
 								for(const deltaTC of chunk.choices[0].delta.tool_calls) {
-								const existing = resp.choices[0].message.tool_calls.find(tc => tc.index === deltaTC.index);
+									const existing = msg.tool_calls.find((tc: any) => tc.index === deltaTC.index);
 									if(existing) {
 										if(deltaTC.id) existing.id = deltaTC.id;
-									if(deltaTC.type) existing.type = deltaTC.type;
-									if(deltaTC.function) {
-										if(!existing.function) existing.function = {};
-										if(deltaTC.function.name) existing.function.name = deltaTC.function.name;
-										if(deltaTC.function.arguments) existing.function.arguments = (existing.function.arguments || '') + deltaTC.function.arguments;
-									}
+										if(deltaTC.function?.name) existing.function.name = deltaTC.function.name;
+										if(deltaTC.function?.arguments) existing.function.arguments += deltaTC.function.arguments;
 									} else {
-									resp.choices[0].message.tool_calls.push({
+										msg.tool_calls.push({
 											index: deltaTC.index,
 											id: deltaTC.id || '',
-										type: deltaTC.type || 'function',
-										function: {
-											name: deltaTC.function?.name || '',
-											arguments: deltaTC.function?.arguments || ''
-										}
+											function: {name: deltaTC.function?.name || '', arguments: deltaTC.function?.arguments || ''}
 										});
 									}
 								}
 							}
 						}
+					} else {
+						usage = resp.usage;
+						msg = resp.choices[0].message;
 					}
+					const duration = Date.now() - callStart;
+					const tps = usage?.completion_tokens && duration > 0 ? usage.completion_tokens / (duration / 1000) : 0;
 
-				if(resp.error) throw new Error(resp.error);
-				const toolCalls = resp.choices[0].message.tool_calls || [];
+					const toolCalls = msg.tool_calls || [];
 					if(toolCalls.length && !controller.signal.aborted) {
-					history.push(resp.choices[0].message);
-					const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
-						const tool = tools?.find(findByProp('name', toolCall.function.name));
-						if(options.stream) options.stream({tool: toolCall.function.name});
-						if(!tool) return {role: 'tool', tool_call_id: toolCall.id, content: '{"error": "Tool not found"}', timestamp: Date.now()};
+						if(msg.content?.trim()) history.push({role: 'assistant', content: msg.content.trim(), timestamp: Date.now(), duration, tps});
+
+						const entries = toolCalls.map((tc: any) => {
+							const entry: any = {role: 'tool', id: tc.id, name: tc.function.name, args: JSONAttemptParse(tc.function.arguments, {}), content: undefined, timestamp: Date.now()};
+							history.push(entry);
+							return {tc, entry};
+						});
+
+						await Promise.all(entries.map(async ({tc, entry}: any) => {
+							const tool = tools.find(findByProp('name', tc.function.name));
+							if(options.stream) options.stream({tool: tc.function.name});
+							if(!tool) { entry.error = 'Tool not found'; return; }
 							try {
-							const args = JSONAttemptParse(toolCall.function.arguments, {});
-							// Wrap stream so a tool's `done` ends turn gracefully
 								const toolStream = options.stream && ((chunk: any) => {
 									if(chunk.done) { terminal = true; return; }
 									options.stream!(chunk);
 								});
-							const result = await tool.fn(args, toolStream, this.ai);
-							return {role: 'tool', tool_call_id: toolCall.id, content: typeof result == 'object' ? JSONSanitize(result) : result, timestamp: Date.now()};
-						} catch (err: any) {
-							return {role: 'tool', tool_call_id: toolCall.id, content: JSONSanitize({error: err?.message || err?.toString() || 'Unknown'}), timestamp: Date.now()};
+								const result = await tool.fn(entry.args, toolStream, this.ai, tc.id);
+								entry.content = typeof result === 'object' ? JSONSanitize(result) : result;
+							} catch(err: any) {
+								entry.error = err?.message || err?.toString() || 'Unknown';
 							}
 						}));
-					history.push(...results);
-					requestParams.messages = history;
+					} else {
+						terminal = true;
+						const text = (msg.content || '').trim();
+						if(text) history.push({role: 'assistant', content: text, timestamp: Date.now(), duration, tps});
 					}
-			} while (!terminal && !controller.signal.aborted && resp.choices?.[0]?.message?.tool_calls?.length);
+				} while(!terminal && !controller.signal.aborted);
 
-			if(!terminal) {
-				const textContent = resp.choices[0].message.content?.trim() || '';
-				history.push({role: 'assistant', content: textContent, timestamp: Date.now()});
-			}
-			history = this.toStandard(history);
 				if(options.stream) options.stream({done: true});
-			if(options.history) options.history.splice(0, options.history.length, ...history);
-			const finalContent = history.at(-1)?.content;
+
+				const turnStart = history.map(h => h.role).lastIndexOf('user');
+				const finalContent = history.slice(turnStart + 1).reduce((str, h) => h.role === 'assistant' ? str + (h.content || '') : str, '').trim();
 				res(options.schema ? JSONAttemptParse(finalContent, finalContent) : finalContent);
+			} catch(err) {
+				rej(err);
+			}
 		}), {abort: () => controller.abort()});
 	}
 }

src/token-pool.ts

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

src/tools.ts

@@ -41,7 +41,7 @@ export type AiTool = {
 	/** Tool arguments */
 	args?: AiToolArg,
 	/** Callback function */
-	fn: (args: any, stream: LLMRequest['stream'], ai: Ai) => any | Promise<any>,
+	fn: (args: any, stream: LLMRequest['stream'], ai: Ai, toolId?: string) => any | Promise<any>,
 };
 
 export function convertSchema(schema: any): any {