Refined memory management prompts

Better memory management
Added official file support
2026-08-25 10:03:36 -04:00 · 2026-08-24 14:42:10 -04:00 · 2026-08-17 15:50:48 -04:00 · 2026-08-17 15:16:32 -04:00 · 2026-08-16 15:40:50 -04:00 · 2026-08-05 16:16:30 -04:00
18 changed files with 2796 additions and 881 deletions
--- a/README.md
+++ b/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);
--- a/main.mjs
+++ b/main.mjs
@@ -1,25 +0,0 @@
-import {Ai} from './dist/index.mjs';
-
-const ai = new Ai({
-	path: './',
-	llm: {
-		system: 'You are a testbed for developing an AI library',
-		models: {
-			'qwen/qwen3.5-9b': {proto: 'openai', host: 'http://127.0.0.1:1234/v1'}
-		}
-	}
-});
-
-const skills = [{
-	name: 'Momentum',
-	description: 'Learn how to use the Momentum API',
-	content: 'You can initialize it with: new Momentum(url);'
-}];
-
-const history = [], memory = [];
-await ai.language.ask('My favorite color is red', {history, memory});
-await ai.language.updateMemory(history, memory);
-
-history.splice(0, history.length);
-console.log(await ai.language.ask('Whats my favorite color?', {history, memory}));
-console.log(history);
--- a/package-lock.json
+++ b/package-lock.json
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "1.0.4",
+	"version": "1.6.4",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -26,12 +26,13 @@
 	},
 	"dependencies": {
 		"@anthropic-ai/sdk": "^0.102.0",
-		"@tensorflow/tfjs": "^4.22.0",
 		"@huggingface/transformers": "^4.2.0",
+		"@tensorflow/tfjs": "^4.22.0",
 		"@ztimson/node-utils": "^1.0.7",
 		"@ztimson/utils": "^0.29.4",
 		"cheerio": "^1.2.0",
 		"openai": "^6.42.0",
+		"pdf-parse": "^2.4.5",
 		"tesseract.js": "^7.0.0"
 	},
 	"devDependencies": {
--- a/src/ai.ts
+++ b/src/ai.ts
@@ -1,5 +1,5 @@
 import * as os from 'node:os';
-import LLM, {AnthropicConfig, OllamaConfig, OpenAiConfig, LLMRequest} from './llm';
+import LLM, {AnthropicConfig, OpenAiConfig, LLMRequest} from './llm';
 import { Audio } from './audio.ts';
 import {Vision} from './vision.ts';

@@ -18,7 +18,7 @@ export type AiOptions = {
 	embedder?: string;
 	/** Large language models, first is default */
 	llm?: Omit<LLMRequest, 'model'> & {
-		models: {[model: string]: AnthropicConfig | OllamaConfig | OpenAiConfig};
+		models: {[model: string]: AnthropicConfig | OpenAiConfig};
 	}
 	/** OCR model: eng, eng_best, eng_fast */
 	ocr?: string;
--- a/src/antrhopic.ts
+++ b/src/antrhopic.ts
@@ -1,63 +1,65 @@
 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({timestamp, role: h.role, content: textContent});
-				h.content.forEach((c: any) => {
-					if(c.type == 'tool_use') {
-						messages.push({timestamp, role: 'tool', id: c.id, name: c.name, args: c.input, 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,
+	private toWireContent(content: any): any {
+		if(!Array.isArray(content)) return content;
+		return content.map(c => c.type === 'image'
+			? {type: 'image', source: {type: 'base64', media_type: c.mime, data: c.data}}
+			: {type: 'text', text: c.text});
+	}
+
+	/** Convert standard history -> Anthropic wire format */
+	private toWire(history: LLMMessage[]): any[] {
+		const wire: any[] = [];
+		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: this.toWireContent(h.content)});
 			}
 		}
-		return history.map(({timestamp, ...h}) => h);
+		return wire;
 	}

-	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
+	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 || 0.7,
+				temperature: options.temperature || this.ai.options.llm?.temperature || undefined,
 				tools: tools.map(t => ({
 					name: t.name,
 					description: t.description,
@@ -65,75 +67,97 @@ 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,
 			};

-			let resp: any, isFirstMessage = true;
+			if(options.schema) {
+				requestParams.output_config = {format: {type: 'json_schema', schema: convertSchema(options.schema)}};
+			}
+
+			try {
+				let terminal = false;
 				do {
-				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});
-					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 {
-							const result = await tool.fn(toolCall.input, options?.stream, this.ai);
-							return {type: 'tool_result', tool_use_id: toolCall.id, content: typeof result == 'object' ? JSONSanitize(result) : result};
+								const toolStream = options.stream && ((chunk: any) => {
+									if(chunk.done) { terminal = true; return; }
+									options.stream!(chunk);
+								});
+								const result = await tool.fn(entry.args, toolStream, this.ai, tc.id);
+								entry.content = typeof result === 'object' ? JSONSanitize(result) : result;
 							} catch(err: any) {
-							return {type: 'tool_result', tool_use_id: toolCall.id, is_error: true, content: err?.message || err?.toString() || 'Unknown'};
+								entry.error = err?.message || err?.toString() || 'Unknown';
 							}
 						}));
-					history.push({role: 'user', content: results});
-					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 (!controller.signal.aborted && resp.content.some((c: any) => c.type === 'tool_use'));
-			history.push({role: 'assistant', content: resp.content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n')});
-			history = this.toStandard(history);
+				} while(!terminal && !controller.signal.aborted);

 				if(options.stream) options.stream({done: true});
-			if(options.history) options.history.splice(0, options.history.length, ...history);
-			res(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()});
 	}
 }
--- a/src/audio.ts
+++ b/src/audio.ts
@@ -141,11 +141,18 @@ print(json.dumps(segments))
 		if(!llm) return transcript;
 		let chunks = this.ai.language.chunk(transcript, 500, 0);
 		if(chunks.length > 4) chunks = [...chunks.slice(0, 3), <string>chunks.at(-1)];
-		const names = await this.ai.language.json(chunks.join('\n'), '{1: "Detected Name", 2: "Second Name"}', {
-			system: 'Use the following transcript to identify speakers. Only identify speakers you are positive about, dont mention speakers you are unsure about in your response',
+		await this.ai.language.ask(chunks.join('\n'), {
+			system: 'Read the following transcript and attempt to identify every speaker. For every positively identified speaker, call the \`identify\` tool with the speaker\'s ID number & the identified name exactly once.',
 			temperature: 0.1,
+			tools: [
+				{name: 'identify', description: 'Identify a speaker', args: {
+					speaker: {type: 'number', description: 'Speaker number', required: true},
+					name: {type: 'string', description: 'Inferred name', required: true},
+				}, fn: ({speaker, name}) => {
+					transcript = transcript.replaceAll(`[Speaker ${speaker}]`, `[${name}]`);
+				}}
+			]
 		});
-		Object.entries(names).forEach(([speaker, name]) => transcript = transcript.replaceAll(`[Speaker ${speaker}]`, `[${name}]`));
 		return transcript;
 	}

--- a/src/helpers.ts
+++ b/src/helpers.ts
@@ -0,0 +1,85 @@
+import {Memory, MemoryCache} from './memory.ts';
+
+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));
+
+	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: 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);
+		}
+	}
+
+	return [
+		...nodes,
+		...[...ghosts].map(name => ({
+			name,
+			missing: true,
+			links: [],
+			backlinks: nodes.filter(n => n.links.includes(name)).map(n => n.name),
+		})),
+	];
+}
+
+export function renderMemoryGraph(nodes: MemoryNode[]): string {
+	if (!nodes.length) return 'No memories yet.';
+
+	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});
+	}
+
+	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));
+		lines.push(`${group}/`);
+		items.forEach((n, i) => {
+			const last = i === items.length - 1;
+			const branch = last ? '└─' : '├─';
+			const pad = last ? '  ' : '│ ';
+			const tag = n.missing ? ' (ghost)' : '';
+			lines.push(`  ${branch} ${n.label}${tag}`);
+			if (n.links.length) lines.push(`  ${pad}   → ${n.links.join(', ')}`);
+			if (n.backlinks.length) lines.push(`  ${pad}   ← ${n.backlinks.join(', ')}`);
+		});
+		lines.push('');
+	}
+
+	return lines.join('\n').trimEnd();
+}
--- a/src/index.ts
+++ b/src/index.ts
@@ -1,9 +1,11 @@
 export * from './ai';
 export * from './antrhopic';
 export * from './audio';
+export * from './helpers';
 export * from './llm';
 export * from './memory';
 export * from './open-ai';
 export * from './provider';
+export * from './token-pool'
 export * from './tools';
 export * from './vision';
--- a/src/kd-tree.ts
+++ b/src/kd-tree.ts
@@ -0,0 +1,335 @@
+export type DistanceMetric = "euclidean" | "cosine";
+
+export interface KDPoint<T = unknown> {
+	vector: number[];
+	payload: T;
+}
+
+export interface KNNResult<T = unknown> {
+	point: KDPoint<T>;
+	distance: number;
+}
+
+interface KDNode<T> {
+	point: KDPoint<T>;
+	axis: number;
+	left: KDNode<T> | null;
+	right: KDNode<T> | null;
+}
+
+// ─── Distance helpers ─────────────────────────────────────────────────────────
+
+function euclidean(a: number[], b: number[]): number {
+	let sum = 0;
+	for (let i = 0; i < a.length; i++) {
+		const d = a[i] - b[i];
+		sum += d * d;
+	}
+	return Math.sqrt(sum);
+}
+
+function cosine(a: number[], b: number[]): number {
+	let dot = 0, normA = 0, normB = 0;
+	for (let i = 0; i < a.length; i++) {
+		dot   += a[i] * b[i];
+		normA += a[i] * a[i];
+		normB += b[i] * b[i];
+	}
+	const denom = Math.sqrt(normA) * Math.sqrt(normB);
+	return denom === 0 ? 1 : 1 - dot / denom; // distance = 1 - similarity
+}
+
+/**
+ * Keeps the k closest candidates in memory, evicts the furthest when full
+ */
+class BoundedMaxHeap<T> {
+	private heap: KNNResult<T>[] = [];
+
+	constructor(private readonly k: number) {}
+
+	get size(): number { return this.heap.length; }
+
+	get worstDistance(): number {
+		return this.heap.length < this.k ? Infinity : this.heap[0].distance;
+	}
+
+	push(item: KNNResult<T>): void {
+		if (this.heap.length < this.k) {
+			this.heap.push(item);
+			this.bubbleUp(this.heap.length - 1);
+		} else if (item.distance < this.heap[0].distance) {
+			this.heap[0] = item;
+			this.sinkDown(0);
+		}
+	}
+
+	toSortedArray(): KNNResult<T>[] {
+		return [...this.heap].sort((a, b) => a.distance - b.distance);
+	}
+
+	private bubbleUp(i: number): void {
+		while (i > 0) {
+			const parent = (i - 1) >> 1;
+			if (this.heap[parent].distance >= this.heap[i].distance) break;
+			[this.heap[parent], this.heap[i]] = [this.heap[i], this.heap[parent]];
+			i = parent;
+		}
+	}
+
+	private sinkDown(i: number): void {
+		const n = this.heap.length;
+		while (true) {
+			let largest = i;
+			const l = 2 * i + 1, r = 2 * i + 2;
+			if (l < n && this.heap[l].distance > this.heap[largest].distance) largest = l;
+			if (r < n && this.heap[r].distance > this.heap[largest].distance) largest = r;
+			if (largest === i) break;
+			[this.heap[largest], this.heap[i]] = [this.heap[i], this.heap[largest]];
+			i = largest;
+		}
+	}
+}
+
+/**
+ * K-D Tree for efficient nearest-neighbor search over high-dimensional vectors / embeddings.
+ *
+ * Supports:
+ *  - Insertion of labeled points
+ *  - k-nearest-neighbor (KNN) search
+ *  - Radius search (all points within a given distance)
+ *  - Euclidean and cosine distance metrics
+ *  - Bulk construction (balanced tree) for best query performance
+ */
+export class KDTree<T = unknown> {
+	private root: KDNode<T> | null = null;
+	private _size = 0;
+	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".
+	 * @param points         Optional initial set of points. Builds a balanced tree
+	 *                       in O(n log² n) — prefer this over inserting one-by-one
+	 *                       when you have a large corpus.
+	 */
+	constructor(
+		dims: number,
+		metric: DistanceMetric = "euclidean",
+		points?: KDPoint<T>[]
+	) {
+		this.dims = dims;
+		this.distanceFn = metric === "cosine" ? cosine : euclidean;
+
+		if (points && points.length > 0) {
+			this.validateAll(points);
+			this.root = this.buildBalanced([...points], 0);
+			this._size = points.length;
+		}
+	}
+
+	/** Total number of points stored in the tree. */
+	get size(): number { return this._size; }
+
+	// ── Insertion ──────────────────────────────────────────────────────────────
+
+	/**
+	 * Insert a single point. O(log n) average, O(n) worst case on skewed data.
+	 * For bulk loading prefer passing points to the constructor.
+	 */
+	insert(point: KDPoint<T>): void {
+		this.validate(point);
+		this.root = this.insertNode(this.root, point, 0);
+		this._size++;
+	}
+
+	// ── KNN search ─────────────────────────────────────────────────────────────
+
+	/**
+	 * Find the k nearest neighbors to `query`.
+	 * Returns results sorted by distance ascending.
+	 */
+	knn(query: number[], k: number): KNNResult<T>[] {
+		if (k <= 0) throw new RangeError("k must be a positive integer");
+		this.validateVector(query);
+
+		const heap = new BoundedMaxHeap<T>(k);
+		this.searchKNN(this.root, query, k, heap, 0);
+		return heap.toSortedArray();
+	}
+
+	/**
+	 * Nearest single neighbor. Convenience wrapper around knn(query, 1).
+	 * Returns null if the tree is empty.
+	 */
+	nearest(query: number[]): KNNResult<T> | null {
+		const results = this.knn(query, 1);
+		return results[0] ?? null;
+	}
+
+	// ── Radius search ──────────────────────────────────────────────────────────
+
+	/**
+	 * Return all points whose distance to `query` is ≤ `radius`,
+	 * sorted by distance ascending.
+	 */
+	radiusSearch(query: number[], radius: number): KNNResult<T>[] {
+		if (radius < 0) throw new RangeError("radius must be non-negative");
+		this.validateVector(query);
+
+		const results: KNNResult<T>[] = [];
+		this.searchRadius(this.root, query, radius, results, 0);
+		results.sort((a, b) => a.distance - b.distance);
+		return results;
+	}
+
+	// ── Conversion ─────────────────────────────────────────────────────────────
+
+	/** Collect all points in the tree (order not guaranteed). */
+	toArray(): KDPoint<T>[] {
+		const out: KDPoint<T>[] = [];
+		this.collect(this.root, out);
+		return out;
+	}
+
+	/**
+	 * Rebuild the tree from its current points as a balanced tree.
+	 * Useful after many individual insertions to restore O(log n) query time.
+	 */
+	rebalance(): void {
+		const points = this.toArray();
+		this.root = points.length ? this.buildBalanced(points, 0) : null;
+	}
+
+	// ── Private: build ─────────────────────────────────────────────────────────
+
+	private buildBalanced(points: KDPoint<T>[], depth: number): KDNode<T> {
+		const axis = depth % this.dims;
+		points.sort((a, b) => a.vector[axis] - b.vector[axis]);
+
+		const mid = Math.floor(points.length / 2);
+		return {
+			point: points[mid],
+			axis,
+			left:  points.slice(0, mid).length
+				? this.buildBalanced(points.slice(0, mid), depth + 1)
+				: null,
+			right: points.slice(mid + 1).length
+				? this.buildBalanced(points.slice(mid + 1), depth + 1)
+				: null,
+		};
+	}
+
+	// ── Private: insert ────────────────────────────────────────────────────────
+
+	private insertNode(
+		node: KDNode<T> | null,
+		point: KDPoint<T>,
+		depth: number
+	): KDNode<T> {
+		if (node === null) {
+			return { point, axis: depth % this.dims, left: null, right: null };
+		}
+		const axis = depth % this.dims;
+		if (point.vector[axis] < node.point.vector[axis]) {
+			node.left = this.insertNode(node.left, point, depth + 1);
+		} else {
+			node.right = this.insertNode(node.right, point, depth + 1);
+		}
+		return node;
+	}
+
+	// ── Private: KNN traversal ─────────────────────────────────────────────────
+
+	private searchKNN(
+		node: KDNode<T> | null,
+		query: number[],
+		k: number,
+		heap: BoundedMaxHeap<T>,
+		depth: number
+	): void {
+		if (node === null) return;
+
+		const dist = this.distanceFn(query, node.point.vector);
+		heap.push({ point: node.point, distance: dist });
+
+		const axis = node.axis;
+		const diff = query[axis] - node.point.vector[axis];
+		const [near, far] = diff <= 0
+			? [node.left, node.right]
+			: [node.right, node.left];
+
+		this.searchKNN(near, query, k, heap, depth + 1);
+
+		// Only explore the far side if it could contain a closer point.
+		// For cosine distance we can't prune by axis gap alone, so always explore.
+		const shouldExplore =
+			this.distanceFn === cosine
+				? true
+				: Math.abs(diff) < heap.worstDistance;
+
+		if (shouldExplore) {
+			this.searchKNN(far, query, k, heap, depth + 1);
+		}
+	}
+
+	// ── Private: radius traversal ──────────────────────────────────────────────
+
+	private searchRadius(
+		node: KDNode<T> | null,
+		query: number[],
+		radius: number,
+		results: KNNResult<T>[],
+		depth: number
+	): void {
+		if (node === null) return;
+
+		const dist = this.distanceFn(query, node.point.vector);
+		if (dist <= radius) {
+			results.push({ point: node.point, distance: dist });
+		}
+
+		const axis = node.axis;
+		const diff = query[axis] - node.point.vector[axis];
+		const [near, far] = diff <= 0
+			? [node.left, node.right]
+			: [node.right, node.left];
+
+		this.searchRadius(near, query, radius, results, depth + 1);
+
+		const shouldExplore =
+			this.distanceFn === cosine ? true : Math.abs(diff) <= radius;
+
+		if (shouldExplore) {
+			this.searchRadius(far, query, radius, results, depth + 1);
+		}
+	}
+
+	// ── Private: collect ───────────────────────────────────────────────────────
+
+	private collect(node: KDNode<T> | null, out: KDPoint<T>[]): void {
+		if (node === null) return;
+		out.push(node.point);
+		this.collect(node.left, out);
+		this.collect(node.right, out);
+	}
+
+	// ── Private: validation ────────────────────────────────────────────────────
+
+	private validateVector(v: number[]): void {
+		if (v.length !== this.dims) {
+			throw new TypeError(
+				`Vector length ${v.length} does not match tree dimensionality ${this.dims}`
+			);
+		}
+	}
+
+	private validate(point: KDPoint<T>): void {
+		this.validateVector(point.vector);
+	}
+
+	private validateAll(points: KDPoint<T>[]): void {
+		for (const p of points) this.validate(p);
+	}
+}
--- a/src/llm.ts
+++ b/src/llm.ts
@@ -1,24 +1,63 @@
+import {clean, makeUnique, snakeCase} from '@ztimson/utils';
 import {AbortablePromise, Ai} from './ai.ts';
 import {Anthropic} from './antrhopic.ts';
 import {OpenAi} from './open-ai.ts';
 import {LLMProvider} from './provider.ts';
-import {AiTool} from './tools.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';
+import {mkdtempSync} from 'node:fs';
+import fs from 'node:fs/promises';
+import {tmpdir} from 'node:os';
+import {dirname, join, basename, extname} from 'path';
+import { PDFParse } from 'pdf-parse';

-export type AnthropicConfig = {proto: 'anthropic', token: string};
-export type OllamaConfig = {proto: 'ollama', host: string};
-export type OpenAiConfig = {proto: 'openai', host?: string, token: string};
+const MAX_AGENT_DEPTH = 5;
+const PDF_OCR_PAGE_THRESHOLD = 12; // above this many pages, OCR scanned pages instead of feeding images to the model
+
+export type AnthropicConfig = {proto: 'anthropic', token: string | string[]};
+export type OpenAiConfig = {proto: 'openai', host?: string, token: string | string[]};
+
+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 LLMFile = {
+	/** Path to file on disk */
+	path?: string;
+	/** File content: raw text, base64-encoded binary, or a Buffer */
+	content?: string | Buffer;
+	/** Original filename, used to infer type from extension */
+	name?: string;
+	/** Mime type override, inferred from extension if omitted */
+	mime?: string;
+	/** @internal set once extraction has run, skips re-processing next turn */
+	extracted?: boolean;
+};

 export type LLMMessage = {
 	/** Message originator */
 	role: 'assistant' | 'system' | 'user';
 	/** Message content */
 	content: string | any;
+	/** Files attached to request */
+	files?: LLMFile[];
 	/** Timestamp */
 	timestamp?: number;
+	/** Response duration in ms */
+	duration?: number;
+	/** Tokens per second */
+	tps?: number;
 } | {
 	/** Tool call */
 	role: 'tool';
@@ -34,15 +73,21 @@ export type LLMMessage = {
 	error?: undefined | string;
 	/** Timestamp */
 	timestamp?: number;
+	/** Response duration in ms */
+	duration?: number;
+	/** Tokens per second */
+	tps?: number;
 }

 export type LLMRequest = {
+	/** Return a parsed JSON object that matches the schema */
+	schema?: AiToolArg;
 	/** System prompt */
 	system?: string;
 	/** 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 */
@@ -54,13 +99,19 @@ 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[];
+	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[];
+	/** Attach files to request */
+	files?: LLMFile[];
+	/** @internal recursion guard for nested agent delegation */
+	_agentDepth?: number;
 }

 export type McpServer = {
@@ -81,8 +132,12 @@ export type Skill = {
 	content: string;
 }

-
 class LLM {
+	private static AUDIO_EXT = ['wav','mp3','m4a','flac','ogg','aac','wma'];
+	private static IMAGE_EXT = ['png','jpg','jpeg','bmp','gif','tiff','webp'];
+	private static TEXT_EXT = ['txt','md','csv','json','xml','html','js','ts','py','yaml','yml','log'];
+	private static PDF_EXT = ['pdf'];
+
 	private memoryManager!: MemoryManager;

 	defaultModel!: string;
@@ -93,12 +148,171 @@ class LLM {
 		Object.entries(ai.options.llm.models).forEach(([model, config]) => {
 			if(!this.defaultModel) this.defaultModel = model;
 			if(config.proto == 'anthropic') this.models[model] = new Anthropic(this.ai, config.token, model);
-			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 loadBuffer(file: LLMFile, asText: boolean): Promise<Buffer> {
+		if(file.path) return fs.readFile(file.path);
+		if(Buffer.isBuffer(file.content)) return file.content;
+		if(typeof file.content === 'string') return Buffer.from(file.content, asText ? 'utf-8' : 'base64');
+		throw new Error('No path or content provided');
+	}
+
+	private async writeTemp(name: string, buffer: Buffer): Promise<string> {
+		const path = join(mkdtempSync(join(tmpdir(), 'ai-file-')), name);
+		await fs.writeFile(path, buffer);
+		return path;
+	}
+
+	/**
+	 * Extract text from a PDF. Pages with no text layer (scanned/image-only) are handled as either:
+	 * - Rendered to images and returned alongside the text so the (vision-capable) model can read them directly
+	 * - OCR'd via Tesseract when the doc is too large to reasonably pass as images
+	 */
+	private async resolvePdf(buffer: Buffer): Promise<{text: string, images: {mime: string, data: string}[]}> {
+		const parser = new PDFParse({data: buffer});
+		try {
+			const {text, pages} = await parser.getText();
+			const scanned = (pages || []).filter(p => !p.text?.trim());
+			if(!scanned.length) return {text: text.trim() || '[Empty PDF]', images: []};
+			const total = pages.length;
+			const pageNums = scanned.map(p => p.num);
+			const {pages: shots} = await parser.getScreenshot({partial: pageNums});
+			if(total <= PDF_OCR_PAGE_THRESHOLD) {
+				return {
+					text: text.trim(),
+					images: shots.map(s => ({mime: 'image/png', data: Buffer.from(s.data).toString('base64')}))
+				};
+			}
+			const ocrText = await Promise.all(shots.map(async (s, i) => {
+				const path = await this.writeTemp(`page-${pageNums[i]}.png`, Buffer.from(s.data));
+				try {
+					return await this.ai.vision.ocr(path) || '';
+				} finally {
+					fs.rm(dirname(path), {recursive: true, force: true}).catch(() => {});
+				}
+			}));
+			return {text: [text.trim(), ...ocrText].filter(Boolean).join('\n\n'), images: []};
+		} finally {
+			await parser.destroy();
+		}
+	}
+
+	private async resolveFile(file: LLMFile): Promise<{text?: string, images?: {mime: string, data: string}[]}> {
+		const name = file.name || (file.path ? basename(file.path) : 'file');
+
+		// Already resolved on a previous turn, reuse cached text
+		if(file.extracted) return {text: `<file name="${name}">\n${file.content}\n</file>`};
+
+		const ext = extname(name).slice(1).toLowerCase();
+		const mime = file.mime || '';
+		const isAudio = mime.startsWith('audio/') || LLM.AUDIO_EXT.includes(ext);
+		const isImage = mime.startsWith('image/') || LLM.IMAGE_EXT.includes(ext);
+		const isPdf = mime === 'application/pdf' || LLM.PDF_EXT.includes(ext);
+		const isText = mime.startsWith('text/') || LLM.TEXT_EXT.includes(ext);
+
+		let tmpDir: string | null = null;
+		try {
+			if(isImage) {
+				const data = (await this.loadBuffer(file, false)).toString('base64');
+				return {images: [{mime: mime || `image/${ext === 'jpg' ? 'jpeg' : ext}`, data}]};
+			}
+
+			if(isPdf) {
+				const {text, images} = await this.resolvePdf(await this.loadBuffer(file, false));
+				// Only cache/skip re-processing when we didn't need to hand off images (OCR'd or fully text-based)
+				if(!images.length) {
+					file.content = text;
+					file.extracted = true;
+					delete file.path;
+				}
+				return {text: `<file name="${name}">\n${text || '[Scanned PDF - see attached page images]'}\n</file>`, images};
+			}
+
+			let text: string;
+			if(isAudio) {
+				let path = file.path;
+				if(!path) {
+					const buffer = await this.loadBuffer(file, false);
+					path = await this.writeTemp(name, buffer);
+					tmpDir = dirname(path);
+				}
+				text = await this.ai.audio.asr(path) || '';
+			} else if(isText) {
+				text = (await this.loadBuffer(file, true)).toString('utf-8');
+			} else {
+				text = typeof file.content === 'string' ? file.content : `[Binary file, unable to extract: ${name}]`;
+			}
+			file.content = text;
+			file.extracted = true;
+			delete file.path;
+
+			return {text: `<file name="${name}">\n${text}\n</file>`};
+		} catch(err: any) {
+			return {text: `<file name="${name}">Failed to process: ${err.message}</file>`};
+		} finally {
+			if(tmpDir) fs.rm(tmpDir, {recursive: true, force: true}).catch(() => {});
+		}
+	}
+
+	private async resolveFiles(files: LLMFile[]): Promise<{text: string, images: {mime: string, data: string}[]}> {
+		const resolved = await Promise.all(files.map(f => this.resolveFile(f)));
+		return {
+			text: resolved.filter(r => r.text).map(r => r.text).join('\n\n'),
+			images: resolved.flatMap(r => r.images || [])
+		};
+	}
+
+	private setupAgent(agents: Agent[] = [], allAgents: Agent[], history: LLMMessage[], aborts: (() => void)[], depth = 0, delegateState: {resp: string | null}): AiTool[] {
+		return agents.map(a => {
+			const toolName = `${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[] = [];
@@ -132,7 +346,7 @@ 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
 		};
 	}
@@ -141,9 +355,9 @@ 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: 'read_skill',
+				name: 'skill_read',
 				description: 'Read the full content of a skill/knowledge document',
 				args: {
 					name: {type: 'string', description: 'Exact skill name', required: true}
@@ -157,10 +371,23 @@ 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: '',
-			temperature: 0.8,
 			...this.ai.options.llm,
 			models: undefined,
 			history: [],
@@ -168,11 +395,24 @@ class LLM {
 		}
 		const m = options.model || this.defaultModel;
 		if(!this.models[m]) throw new Error(`Model does not exist: ${m}`);
-		let abort = () => {};
-		return Object.assign(new Promise<string>(async res => {
+		let request: AbortablePromise<string> | null = null;
+		let aborted = false;
+		const nestedAborts: (() => void)[] = [];
+		const abort = () => {
+			aborted = true;
+			request?.abort?.();
+			nestedAborts.forEach(a => a());
+		};
+
+		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 || [];
+			const files = options.files || [];
+			if(message || files.length) history.push({role: 'user', content: message || '', timestamp: Date.now()});

 			// MCP
 			const mcp = options.mcp || this.ai.options?.llm?.mcp;
@@ -190,47 +430,110 @@ 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 relevant = await this.memoryManager.recollect(message, options.memory, 1);
-				prompts.unshift(`You have access to the following memory files:
-${options.memory.map(m => `- ${m.name}: ${m.description}`).join('\n')}
-${relevant.length ? `
-The closest memory has been added primitively:
-\`\`\`
-Name: ${relevant[0].name}
-Description: ${relevant[0].description}
-${relevant[0].content}
-\`\`\`
-`: ''}`.trim());
-				tools.push(this.memoryManager.tools.read(<Memory[]>options.memory));
+			const mem = MemoryManager.normalize(options.memory);
+			if(mem) {
+				const mems = mem.memory instanceof MemoryCache ? mem.memory.memories : mem.memory;
+				if(mems.length) {
+					if(mem.inject) {
+						const pool = 15; // candidates considered, cheap since only refs are listed
+						const 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'});
+
+			const lastMsg = history[history.length - 1];
+			if(files.length && lastMsg?.role === 'user') lastMsg.files = files;
+			const restores: {msg: LLMMessage, content: any}[] = [];
+			for(const msg of history) {
+				if(msg.role !== 'user' || !msg.files?.length) continue;
+				const {text, images} = await this.resolveFiles(msg.files);
+				if(!text && !images.length) continue;
+				restores.push({msg, content: msg.content});
+				const merged = text ? [msg.content, text].filter(Boolean).join('\n\n') : msg.content;
+				msg.content = images.length
+					? [...images.map(i => ({type: 'image', mime: i.mime, data: i.data})), {type: 'text', text: merged}]
+					: merged;
+			}
+
+			const toolTimings = new Map<string, {duration: number, tps: number}>();
+			tools = this.wrapToolTiming(tools, toolTimings);
+
+			if(aborted) throw Object.assign(new Error('Aborted'), {name: 'AbortError'});
+
 			prompts.unshift(options.system || this.ai.options.llm?.system || '');
-			const resp = await this.models[m].ask(message, {...options, tools, system: prompts.filter(Boolean).join('\n\n')});
+			request = this.models[m].ask('', {...options, tools, system: prompts.filter(Boolean).join('\n\n')});
+			let resp = await request;

-			// Trim memory injections from history
-			if(options.memory) {
-				history.splice(0, history.length, ...history.filter(h => h.role !== 'tool' || h.name !== 'recall'));
+			// Strip the file injection shim
+			restores.forEach(({msg, content}) => msg.content = content);
+
+			// 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, 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);
 			}

-			return res(resp);
-		}), {abort});
-	}
+			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});

-	/**
-	 * 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});
+			return resp;
+		})();
+
+		return Object.assign(promise, {abort});
 	}

 	/**
@@ -383,49 +686,41 @@ ${relevant[0].content}
 	 * @param {string} searchTerms Multiple search terms to check against target
 	 * @returns {{avg: number, max: number, similarities: number[]}} Similarity values 0-1: 0 = unique, 1 = identical
 	 */
-	fuzzyMatch(target: string, ...searchTerms: string[]) {
+	fuzzyMatch(target, ...searchTerms) {
 		if (searchTerms.length < 2) throw new Error('Requires at least 2 strings to compare');
-		const vector = (text: string, dimensions: number = 10): number[] => {
-			return text.toLowerCase().split('').map((char, index) =>
-				(char.charCodeAt(0) * (index + 1)) % dimensions / dimensions).slice(0, dimensions);
+		const levenshtein = (a, b) => {
+			const m = a.length, n = b.length;
+			if (!m) return n;
+			if (!n) return m;
+			const dp = Array.from({length: m + 1}, (_, i) => [i, ...Array(n).fill(0)]);
+			for (let j = 0; j <= n; j++) dp[0][j] = j;
+			for (let i = 1; i <= m; i++) {
+				for (let j = 1; j <= n; j++) {
+					dp[i][j] = a[i - 1] === b[j - 1]
+						? dp[i - 1][j - 1]
+						: 1 + Math.min(dp[i - 1][j - 1], dp[i - 1][j], dp[i][j - 1]);
 				}
-		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};
+			}
+			return dp[m][n];
+		};
+		const similarity = (a, b) => {
+			a = a.toLowerCase(); b = b.toLowerCase();
+			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,
+			max: Math.max(...similarities),
+			similarities
+		};
 	}

 	/**
-	 * Ask a question with JSON response
-	 * @param {string} text Text to process
-	 * @param {string} schema JSON schema the AI should match
-	 * @param {LLMRequest} options Configuration options and chat history
-	 * @returns {Promise<{} | {} | RegExpExecArray | null>}
+	 * Digest full conversation history into memory documents.
+	 * Call on session end to persist the conversation.
 	 */
-	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${schema}\n\`\`\`\n\nResponses are ignored`;
-		if(options?.system) system += '\n\n' + options.system;
-		return new Promise(async (resolve, reject) => {
-			let done = false;
-			const resp = await this.ask(text, {
-				temperature: 0.3,
-				...options,
-				system,
-				tools: [{
-					name: 'submit',
-					description: 'Submit JSON',
-					args: {json: {type: 'string', description: 'Javascript parsable JSON string', required: true}},
-					fn: (args) => {
-						try {
-							const json = JSON.parse(args.json);
-							resolve(json);
-							done = true;
-						} catch { return 'Invalid JSON'; }
-						return 'Saved';
-					}
-				}, ...(options?.tools || [])],
-			});
-			if(!done) reject(`AI failed to create JSON:\n${resp}`);
-		});
+	async memorize(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest = {}): Promise<Memory[]> {
+		return this.memoryManager.memorize(history, memories, {model: this.defaultModel, ...options});
 	}

 	/**
@@ -462,9 +757,8 @@ ${relevant[0].content}
 		});
 	}

-	addModel(name: string, config: AnthropicConfig | OllamaConfig | OpenAiConfig, setDefault = false) {
+	addModel(name: string, config: AnthropicConfig | OpenAiConfig, setDefault = false) {
 		if(config.proto == 'anthropic') this.models[name] = new Anthropic(this.ai, config.token, name);
-		else if(config.proto == 'ollama') this.models[name] = new OpenAi(this.ai, config.host, 'not-needed', name);
 		else if(config.proto == 'openai') this.models[name] = new OpenAi(this.ai, config.host || null, config.token, name);
 		if(setDefault || !this.defaultModel) this.defaultModel = name;
 	}
@@ -476,12 +770,11 @@ ${relevant[0].content}
 		}
 	}

-	setModels(models: {[model: string]: AnthropicConfig | OllamaConfig | OpenAiConfig}, replace = true) {
+	setModels(models: {[model: string]: AnthropicConfig | OpenAiConfig}, replace = true) {
 		if(replace) this.models = {};
 		Object.entries(models).forEach(([model, config]) => {
 			if(!this.defaultModel) this.defaultModel = model;
 			if(config.proto == 'anthropic') this.models[model] = new Anthropic(this.ai, config.token, model);
-			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.defaultModel = Object.keys(this.models)[0] ?? '';
--- a/src/memory.ts
+++ b/src/memory.ts
@@ -1,177 +1,644 @@
-// memory.ts
+import {MemoryNode, rebuildGraph} from './helpers.ts';
 import {LLMRequest, LLMMessage} from './llm.ts';
 import {AiTool} from './tools.ts';
+import {KDPoint, KDTree} from './kd-tree.ts';
+import {escapeRegex} from '@ztimson/utils';
+
+const MERGE_THRESHOLD = 0.88;
+const PENDING_HEADING = '## Pending';
+const GENERIC_TEMPLATE = `# {{Title}}
+
+## Summary
+
+## Details
+
+## Related`;

-/** 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[];
+	links: string[];
+	backlinks: string[];
 }

-export type MemoryCollection = {
-	/** Memory subject */
+type MemoryRef = {
 	name: string;
-	/** Short description - required if isNew */
-	description?: string;
-	/** Extracted facts to merge */
+	description: string;
+}
+
+type FactBucket = {
+	subject: string;
 	facts: string[];
 }

+type FactAgentResult = {
+	buckets: FactBucket[];
+	journal: string;
+}
+
+function dedupeFacts(facts: string[]): string[] {
+	const seen = new Map<string, string>();
+	for (const f of facts) {
+		const clean = f.trim();
+		if (clean) seen.set(clean.toLowerCase(), clean);
+	}
+	return [...seen.values()];
+}
+
+function cosineDistance(a: number[], b: number[]): number {
+	let dot = 0, normA = 0, normB = 0;
+	for (let i = 0; i < a.length; i++) {
+		dot   += a[i] * b[i];
+		normA += a[i] * a[i];
+		normB += b[i] * b[i];
+	}
+	const denom = Math.sqrt(normA) * Math.sqrt(normB);
+	return denom === 0 ? 1 : 1 - dot / denom;
+}
+
+function cosineSearch(query: number[], memories: Memory[], limit: number): MemoryRef[] {
+	return memories
+		.filter(m => m.embedding?.length)
+		.map(m => ({ref: {name: m.name, description: m.description}, distance: cosineDistance(query, m.embedding)}))
+		.sort((a, b) => a.distance - b.distance)
+		.slice(0, limit)
+		.map(s => s.ref);
+}
+
+export function stripHeader(content: string): string {
+	return content.replace(/^---[\s\S]*?\n---\n?/, '').trimStart();
+}
+
+export class MemoryCache {
+	private tree!: KDTree<MemoryRef>;
+	public memories: Memory[];
+	public nodes: MemoryNode[] = [];
+
+	get length() { return this.memories.length; }
+
+	constructor(memories: Memory[]) {
+		this.memories = memories;
+		this.rebuild();
+	}
+
+	private buildTree(): KDTree<MemoryRef> {
+		const embedded = this.memories.filter(m => m.embedding?.length);
+		if (!embedded.length) return new KDTree<MemoryRef>(0);
+
+		const dims = embedded[0].embedding.length;
+		const points: KDPoint<MemoryRef>[] = embedded.map(m => ({
+			vector: m.embedding,
+			payload: {name: m.name, description: m.description},
+		}));
+
+		return new KDTree<MemoryRef>(dims, 'cosine', points);
+	}
+
+	search(query: number[], limit: number): MemoryRef[] {
+		if (!this.tree || this.tree.dims === 0) return [];
+		return this.tree.knn(query, limit).map(r => r.point.payload);
+	}
+
+	add(memory: Memory): void {
+		this.memories.push(memory);
+		this.rebuild();
+	}
+
+	update(memory: Memory): void {
+		const existing = this.memories.find(m => m.name === memory.name);
+		if (existing) Object.assign(existing, 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(`${node.description}\n\n${stripHeader(node.content)}`.trim());
+			if (e) node.embedding = e.embedding;
+		}));
+		this.commit();
+		return missing.length;
+	}
+}
+
+export type MemoryOptions = {
+	/** Memory object */
+	memory: Memory[] | MemoryCache;
+	/** Inject N memories into the system prompt */
+	inject?: boolean;
+	/** expose recall tool to LLM */
+	tool?: boolean;
+	/** Update memory on compression */
+	update?: boolean;
+	/** Max context size of memories to inject to each call (removed immediately after use) */
+	maxTokens?: number;
+}
+
 export class MemoryManager {
+	private recentlyTouched = new Map<string, number>();
+
+	private queues = new Map<string, {
+		dirty: boolean,
+		request: {abort?: () => void} | null,
+		task: Promise<void>,
+	}>();

 	tools = {
-		edit: (memory: Memory): AiTool => ({
-			name: 'edit_memory',
-			description: 'Edit a memory. Omit start/end to append. Pass start only to replace from that line on (Note line 0 = first line of content / line AFTER description). Pass start+end to replace a specific range. start=0 replaces the whole document. Returns updated document',
+		forget: (memories: Memory[] | MemoryCache): AiTool => ({
+			name: 'memory_forget',
+			description: 'Permanently delete a memory document and clean up all references to it',
 			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.'},
+				name: {type: 'string', description: 'Exact memory name to forget', required: true}
 			},
 			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 memory.content;
-			}
+				const result = this.forget(args.name, memories);
+				return result ? `Forgotten: ${args.name}` : `Not found: ${args.name}`;
+			},
 		}),
-		extract: (pools: MemoryCollection[]): AiTool => ({
-			name: 'extract_facts',
-			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', required: true},
-				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[]): AiTool => ({
-			name: 'read_memory',
-			description: 'Read entire memory',
+
+		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 = memories.find(m => m.name === args.name);
+				const mem = this.access(memories).find(args.name);
 				if (!mem) return 'Document not found';
-				return `Name: ${mem.name}\nDescription: ${mem.description}\n\n${mem.content}`;
-			}
+				this.touch(mem.name);
+				return 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: ${m.description}`).join('\n\n');
-		const pools: MemoryCollection[] = [];
-		await this.llm.ask(conversation, {
-			model: this.model || options.model,
-			temperature: 0.2,
-			system: `You are a fact extractor. Analyze this conversation and extract facts worth remembering long term.
-Rules:
- ONLY extract facts the USER explicitly stated about themselves or their business
- ONLY extract decisions that were MADE during this conversation
- DO NOT extract anything the AI said, its name, capabilities, or how it introduced itself
- DO NOT extract greetings, pleasantries or generic exchanges
- If nothing worth remembering was said, dont do anything, skip calling tools
-
-For each fact decide whether it belongs in an existing document or needs a new one, then call the \`extract_facts\` tool.
-
-Existing documents:\n${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_memory\` tool; call it as many times as necessary:
+		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: ${[...m.links, ...m.backlinks].join(', ')}
 \`\`\`
-${mem.content}
+${m.content}
+\`\`\``).join('\n\n');
+			},
+		}),
+	};
+
+	constructor(private llm: any) {}
+
+	static normalize(m?: Memory[] | MemoryCache | MemoryOptions) {
+		if (!m) return null;
+		const raw = m instanceof MemoryCache || Array.isArray(m);
+		return raw ? {memory: <Memory[] | MemoryCache>m, inject: true, tool: true, update: true} : {inject: true, tool: true, update: true, ...m};
+	}
+
+	private access(memories: Memory[] | MemoryCache): MemoryAccessor {
+		return new MemoryAccessor(memories);
+	}
+
+	private stage(node: Memory, block: string): void {
+		this.ensureDoc(node);
+		const body = stripHeader(node.content);
+		const idx = body.indexOf(PENDING_HEADING);
+		const newBody = idx === -1
+			? `${body.trimEnd()}\n\n${PENDING_HEADING}\n${block}\n`
+			: `${body.slice(0, idx + PENDING_HEADING.length)}\n${block}${body.slice(idx + PENDING_HEADING.length)}`;
+		node.content = this.touchHeader(node, newBody);
+	}
+
+	private ensureDoc(node: Memory): void {
+		if (node.content) return;
+		const title = node.name.split('/').pop() ?? node.name;
+		node.content = this.touchHeader(node, `# ${title}\n`);
+	}
+
+	private sanitizeDescription(text: string): string {
+		return (text ?? '').replace(/\s+/g, ' ').trim().slice(0, 240);
+	}
+
+	private relink(memories: Memory[], from: string, to: string): void {
+		const pattern = new RegExp(`\\[\\[${escapeRegex(from)}\\]\\]`, 'g');
+		for (const m of memories) if (pattern.test(m.content)) m.content = m.content.replace(pattern, `[[${to}]]`);
+	}
+
+	private async factAgent(conversation: string, store: MemoryAccessor, options: LLMRequest): Promise<FactAgentResult> {
+		const ghosts = store.ghosts();
+
+		const response = await this.llm.ask(conversation, {
+			model: options.model,
+			temperature: 0.2,
+			system: `You are a fact extractor for Obsidian-style knowledge vaults. Analyze the conversation and produce:
+
+1. Journal recap (single paragraph)
+   - "Captains Log" style record keeping
+   - What was discussed/worked on, decisions, user's events/state/mood, general context
+   - Leave empty only for trivial/empty exchanges/small talk
+
+2. Fact buckets
+   - ONLY facts the USER explicitly stated about themselves, their work, projects, or decisions made during this conversation
+   - NEVER extract greetings, pleasantries, or anything the assistant itself said
+   - Extract the final/end state, not deltas
+
+Path assignment rules:
+   - Reuse existing node names whenever possible
+   - Documents should be grouped and named by the root subject
+     - Person → People/Name
+     - Project → Projects/Name
+     - Concept → Concepts/Name
+   - A bug report, its investigation, should be nested and attached to the same root subject node
+   - Tickets/one-off tasks → file under the project/name/component they belong to
+   - Only create a new top-level node when the fact belongs to a genuinely new subject (person/project/concept)\`
+
+Available nodes:
+${this.listNodes(store.list).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}
+${ghosts.length ? `${ghosts.map(g => `- ${g}: (Ghost)`).join('\n')}` : ''}`,
+			schema: {
+				journal: {type: 'string', description: 'Short day-to-day recap; empty if nothing happened.', required: false},
+				buckets: {type: 'array', description: 'Groups of facts to remember; empty array if nothing worth storing.', items: {
+						type: 'object', items: {
+							subject: {type: 'string', description: 'Exact node name or new path (e.g. "People/Sarah", "Projects/Oxide")', required: true},
+							facts: {type: 'array', description: 'Facts to store here', items: {type: 'string'}},
+						},
+					},
+				},
+			},
+		});
+
+		const buckets = new Map<string, string[]>();
+		for (const bucket of response.buckets ?? []) {
+			const subject = bucket.subject.trim();
+			const facts = buckets.get(subject) ?? [];
+			facts.push(...dedupeFacts(bucket.facts));
+			buckets.set(subject, facts);
+		}
+
+		return {
+			buckets: buckets.entries().toArray().map(([subject, facts]) => ({subject, facts})),
+			journal: (response.journal ?? '').trim(),
+		};
+	}
+
+	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}));
+	}
+
+	/** Find the nearest node above the similarity threshold and fold the smaller/less-connected one into
+	 * the other. Journals are exempt — they're partitioned by date, not topic, and merging across weeks
+	 * would wreck the timeline. Returns 'merged' if `node` absorbed another (caller should re-run the doc
+	 * agent), 'absorbed' if `node` itself got folded away (caller should stop touching it), or null. */
+	private async checkMerge(node: Memory, memories: Memory[] | MemoryCache, options: LLMRequest, threshold = MERGE_THRESHOLD): Promise<Memory | null> {
+		if (!node.embedding?.length || node.name.startsWith('Journal/')) return null;
+		const store = this.access(memories);
+
+		let closest: Memory | null = null, closestDist = Infinity;
+		for (const other of store.list) {
+			if (other.name === node.name || other.name.startsWith('Journal/') || !other.embedding?.length) continue;
+			const d = cosineDistance(node.embedding, other.embedding);
+			if (d < closestDist) { closestDist = d; closest = other; }
+		}
+		if (!closest || closestDist > threshold) return null;
+
+		const result = await this.mergeAgent(node, closest, options);
+		const merged: Memory = {name: result.name, description: this.sanitizeDescription(result.description), content: '', embedding: [], links: [], backlinks: []};
+		merged.content = this.touchHeader(merged, result.content);
+		const [e] = await this.llm.embedding(`${merged.description}\n\n${result.content}`.trim());
+		if (e) merged.embedding = e.embedding;
+
+		this.relink(store.list, node.name, merged.name);
+		this.relink(store.list, closest.name, merged.name);
+
+		this.queues.get(closest.name)?.request?.abort?.();
+		this.queues.delete(closest.name);
+
+		store.forget(node.name);
+		store.forget(closest.name);
+		store.list.push(merged);
+		store.commit();
+
+		return merged;
+	}
+
+	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 () => {
+			let current = node;
+			do {
+				entry.dirty = false;
+				await this.docAgent(current, store.list, options, entry);
+				const merged = await this.checkMerge(current, memories, options);
+				if (merged) { current = merged; entry.dirty = true; }
+			} 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> {
+		if(!memories.includes(node)) return;
+		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 factual sentence describing the document\'s ENTIRE SUBJECT MATTER — for use as a search/merge fingerprint', required: true},
+						content: {type: 'string', description: 'Rewritten document body in markdown, without the frontmatter block', required: true},
+					},
+					system: `You are a knowledge base editor maintaining one Obsidian-style document.
+
+If it has a "## Pending" section, fold all new material into the appropriate part, resolve overlap, then remove the section entirely. If no section, just tidy per the rules below.
+
+Use this loose structure, adapting headings to what the content needs:
+
+# Title
+## Summary
+## Details
+## Related
+
+Rules:
+- Contradictions: newer facts always win — delete outdated statements entirely
+- Journals (Journal/...): keep entries as a chronological timeline; clean up grammar within entries but never delete history
+- Use Obsidian markdown: # headings, **bold**, bullet/numbered lists, tables for 2D data
+- Link specific entities and concepts with [[WikiLink]] (e.g., [[Projects/KiwixServer]]); skip generics
+- Keep concise, factual, human-readable
+- NO frontmatter, filler, preamble, or AI commentary
+
+Available nodes to link to (don't duplicate their content):
+${this.listNodes(memories).filter(n => n.name !== node.name).map(n => n.name).join(', ') || 'none'}
+
+Current document:
+\`\`\`markdown
+${currentBody}
 \`\`\``,
-				tools: [this.tools.edit(mem)]
+				});
+				entry.request = request;
+				update = await request;
 			}
-		);
-
-		if(isNew || mem.description !== existing?.description) {
-			const e = await this.llm.embedding(mem.description);
-			mem.embedding = e?.[0]?.embedding;
+		} catch (err: any) {
+			if (err?.name === 'AbortError') return;
+			throw err;
+		} finally {
+			entry.request = null;
 		}

-		if(isNew) memories.push(mem);
-		else {
-			const idx = memories.findIndex(m => m.name === newMem.name);
-			if(idx >= 0) memories[idx] = mem;
+		if (!update?.content) return;
+		node.description = node.name !== 'People/User' ? this.sanitizeDescription(update.description) : 'All information about the current user';
+		node.content = this.touchHeader(node, update.content);
+		const [e] = await this.llm.embedding(`${node.description}\n\n${update.content}`.trim());
+		if (e) node.embedding = e.embedding;
+	}
+
+	private async mergeAgent(a: Memory, b: Memory, options: LLMRequest): Promise<{name: string, description: string, content: string}> {
+		return this.llm.ask('', {
+			model: options.model,
+			temperature: 0.3,
+			schema: {
+				name: {type: 'string', description: 'New path for the merged doc, collection/subject format (e.g. Projects/Oxide) — only reuse an old title if it\'s genuinely the best fit', required: true},
+				description: {type: 'string', description: 'One factual sentence describing the merged document\'s subject matter', required: true},
+				content: {type: 'string', description: 'Fully reconciled body in markdown, without frontmatter', required: true},
+			},
+			system: `You are a knowledge base editor merging two overlapping Obsidian documents into one. Newer facts win on contradiction.
+
+Structure loosely:
+
+# Title
+## Summary
+## Details
+## Related
+
+Combine both documents, resolve duplication and contradictions.
+
+Document A ("${a.name}"):
+\`\`\`markdown
+${stripHeader(a.content)}
+\`\`\`
+
+Document B ("${b.name}"):
+\`\`\`markdown
+${stripHeader(b.content)}
+\`\`\``,
+		});
+	}
+
+	private parseFrontmatter(content: string): {fm: Map<string, string>, body: string} {
+		const match = content.match(/^---\n([\s\S]*?)\n---\n?([\s\S]*)$/);
+		if (!match) return {fm: new Map(), body: content};
+		const fm = new Map<string, string>();
+		for (const line of match[1].split('\n')) {
+			const i = line.indexOf(':');
+			if (i === -1) continue;
+			const key = line.slice(0, i).trim();
+			const raw = line.slice(i + 1).trim();
+			let value = raw;
+			try { value = JSON.parse(raw); } catch { /* legacy unquoted value, keep raw */ }
+			fm.set(key, value);
+		}
+		return {fm, body: match[2]};
+	}
+
+	private touchHeader(node: Memory, body: string): string {
+		const {fm} = this.parseFrontmatter(node.content);
+		fm.set('name', node.name);
+		fm.set('description', node.description || '');
+		fm.set('modified', new Date().toISOString());
+		return this.writeFrontmatter(fm, body);
+	}
+
+	private writeFrontmatter(fm: Map<string, string>, body: string): string {
+		const lines = [...fm.entries()].map(([k, v]) => `${k}: ${JSON.stringify(String(v).replace(/\s+/g, ' ').trim())}`);
+		return `---\n${lines.join('\n')}\n---\n\n${body.trimStart()}`;
+	}
+
+	decay() {
+		for (const [name, ttl] of this.recentlyTouched) {
+			if (ttl <= 1) this.recentlyTouched.delete(name);
+			else this.recentlyTouched.set(name, ttl - 1);
 		}
 	}

-	/**
-	 * 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[]> {
+	touch(name: string, ttl = 2) {
+		this.recentlyTouched.set(name, ttl);
+	}
+
+	forget(name: string, memories: Memory[] | MemoryCache): boolean {
+		return this.access(memories).forget(name);
+	}
+
+	async recollect(query: string, memories: Memory[] | MemoryCache, limit = 5, graphDepth = 1): Promise<Memory[]> {
+		const store = this.access(memories);
+		if (!store.list.length) return [];
+
+		await store.backfillEmbeddings(this.llm);
+
 		const [e] = await this.llm.embedding(query);
-		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);
+		if (!e) return [];
+
+		const vectorResults = store.search(e.embedding, limit);
+		const found = new Set<string>(vectorResults.map(r => r.name));
+
+		if (graphDepth > 0) {
+			let frontier = [...found];
+			for (let depth = 0; depth < graphDepth && frontier.length; depth++) {
+				const next: string[] = [];
+				for (const name of frontier) {
+					const node = store.find(name);
+					if (!node) continue;
+					for (const link of node.links) {
+						if (!found.has(link) && store.find(link)) {
+							found.add(link);
+							next.push(link);
+						}
+					}
+				}
+				frontier = next;
+			}
 		}

-	/**
-	 * 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 vectorOrder = vectorResults.map(r => r.name);
+		const graphExpansions = [...found].filter(n => !vectorOrder.includes(n));
+		return [...vectorOrder, ...graphExpansions].map(n => store.find(n)!).filter(Boolean);
+	}
+
+	async memorize(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest): Promise<Memory[]> {
 		const conversation = history
 			.filter(h => h.role === 'user' || h.role === 'assistant')
-			.map(h => `[${h.role}]: ${h.content}`)
-			.join('\n\n');
-		if(!conversation.trim()) return;
-		const pools = await this.extract(conversation, memories, options);
-		if(!pools.length) return;
-		await Promise.all(pools.map(pool => this.edit(pool, memories, options)));
+			.map(h => `[${h.role}]: ${h.content}`).join('\n\n').trim();
+		if (!conversation) return [];
+
+		const uid = `${Date.now()}_${Math.random().toString(36).slice(2)}`;
+		const pending = {role: 'tool', name: 'memory_process', id: uid, content: conversation} as unknown as LLMMessage;
+		history.push(pending);
+
+		const store = this.access(memories);
+		const {buckets, journal} = await this.factAgent(conversation, store, options);
+		const touched: Memory[] = [];
+
+		if (journal) {
+			const journalName = `Journal/${this.getWeekMonday()}`;
+			let jnode = store.find(journalName);
+			if (!jnode) {
+				jnode = {name: journalName, description: '', content: '', embedding: [], links: [], backlinks: []};
+				store.list.push(jnode);
+			}
+			this.stage(jnode, `### ${new Date().toISOString().slice(0, 10)}\n${journal}`);
+			touched.push(jnode);
+		}
+
+		for (const {subject, facts} of buckets) {
+			let node = store.find(subject);
+			if (!node) {
+				node = {name: subject, description: '', content: '', embedding: [], links: [], backlinks: []};
+				store.list.push(node);
+			}
+			this.stage(node, facts.map(f => `- ${f}`).join('\n'));
+			touched.push(node);
+		}
+
+		for (const node of touched) {
+			const [e] = await this.llm.embedding(`${node.description}\n\n${stripHeader(node.content)}`.trim());
+			if (e) node.embedding = e.embedding;
+			this.touch(node.name);
+		}
+
+		if (touched.length) {
+			store.commit();
+			(pending as any).content = `Saved to ${touched.map(n => `[[${n.name}]]`).join(', ')}`;
+			await Promise.all(touched.map(node => this.reconcile(node, memories, options).catch(() => {})));
+		} else {
+			(pending as any).content = 'Nothing worth remembering.';
+		}
+
+		(touched as any).uid = uid;
+		return touched;
+	}
+
+	async reconcileAll(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(PENDING_HEADING));
+		await Promise.all(targets.map(node => this.reconcile(node, memories, options)));
+		store.commit();
 	}
 }
--- a/src/open-ai.ts
+++ b/src/open-ai.ts
@@ -1,84 +1,72 @@
 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' && h.content) {
-				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) => {
+	private toWireContent(content: any): any {
+		if(!Array.isArray(content)) return content;
+		return content.map(c => c.type === 'image'
+			? {type: 'image_url', image_url: {url: `data:${c.mime};base64,${c.data}`}}
+			: {type: 'text', text: c.text});
+	}
+
+	/** Convert standard history -> OpenAI wire format */
+	private toWire(history: LLMMessage[], system?: string): any[] {
+		const wire: any[] = [];
+		if(system) wire.push({role: 'system', content: system});
+		for(const h of history) {
 			if(h.role === 'tool') {
-				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: []
 				}, {
 					role: 'tool',
 					tool_call_id: h.id,
-					content: h.error || h.content
+					content: h.error || h.content || '',
 				});
 			} else {
-				const {timestamp, ...rest} = h;
-				result.push(rest);
+				wire.push({role: h.role, content: this.toWireContent(h.content)});
 			}
-			return result;
-		}, [] as any[]);
+		}
+		return wire;
 	}

-	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
+	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_tokens: options.max_tokens || this.ai.options.llm?.max_tokens || 4096,
-				temperature: options.temperature || this.ai.options.llm?.temperature || 0.7,
+				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',
 					function: {
@@ -93,77 +81,96 @@ export class OpenAi extends LLMProvider {
 				}))
 			};

-			let resp: any, isFirstMessage = true;
+			if(options.schema) {
+				const schema = convertSchema(options.schema);
+				requestParams.response_format = {type: 'json_schema', json_schema: {name: 'response', strict: true, schema}};
+			}
+			if(options.stream) requestParams.stream_options = {include_usage: true};
+
+			try {
+				let terminal = false;
 				do {
-				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: []}}];
 						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"}'};
+						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, {});
-							const result = await tool.fn(args, options.stream, this.ai);
-							return {role: 'tool', tool_call_id: toolCall.id, content: typeof result == 'object' ? JSONSanitize(result) : result};
+								const toolStream = options.stream && ((chunk: any) => {
+									if(chunk.done) { terminal = true; return; }
+									options.stream!(chunk);
+								});
+								const result = await tool.fn(entry.args, toolStream, this.ai, tc.id);
+								entry.content = typeof result === 'object' ? JSONSanitize(result) : result;
 							} catch(err: any) {
-							return {role: 'tool', tool_call_id: toolCall.id, content: JSONSanitize({error: err?.message || err?.toString() || 'Unknown'})};
+								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 (!controller.signal.aborted && resp.choices?.[0]?.message?.tool_calls?.length);
-			history.push({role: 'assistant', content: resp.choices[0].message.content.trim() || ''});
-			history = this.toStandard(history);
+				} while(!terminal && !controller.signal.aborted);

 				if(options.stream) options.stream({done: true});
-			if(options.history) options.history.splice(0, options.history.length, ...history);
-			res(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()});
 	}
 }
--- a/src/provider.ts
+++ b/src/provider.ts
@@ -1,5 +1,5 @@
 import {AbortablePromise} from './ai.ts';
-import {LLMMessage, LLMRequest} from './llm.ts';
+import {LLMRequest} from './llm.ts';

 export abstract class LLMProvider {
 	abstract ask(message: string, options: LLMRequest): AbortablePromise<string>;
--- a/src/token-pool.ts
+++ b/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)}...${token.slice(-4)}`;
+	}
+
+	/** Anthropic & OpenAI SDKs both attach `status` to thrown errors */
+	private statusCode(err: any): number {
+		return err?.status ?? err?.response?.status ?? err?.statusCode;
+	}
+
+	private retryAfter(err: any): number {
+		const headers = err?.headers || err?.response?.headers;
+		const raw = headers?.get?.('retry-after') ?? headers?.['retry-after'];
+		if(raw) {
+			const seconds = Number(raw);
+			if(!isNaN(seconds)) return Date.now() + seconds * 1000;
+			const date = new Date(raw).getTime();
+			if(!isNaN(date)) return date;
+		}
+		return Date.now() + DEFAULT_COOLDOWN;
+	}
+
+	async run<T>(fn: (token: string) => Promise<T>): Promise<T> {
+		const now = Date.now();
+		for(const state of this.states) {
+			if(state.cooldownUntil > now) continue;
+			try {
+				const result = await fn(state.token);
+				state.cooldownUntil = 0;
+				state.lastError = undefined;
+				return result;
+			} catch(err: any) {
+				const code = this.statusCode(err);
+				if(![401, 403, 429].includes(code)) throw err;
+				state.cooldownUntil = code === 429 ? this.retryAfter(err) : Date.now() + DEFAULT_COOLDOWN;
+				state.lastError = {code, message: err?.message || 'Unknown error'};
+			}
+		}
+
+		const failures: Record<string, {code: number, message: string}> = {};
+		this.states.forEach(s => { if(s.lastError) failures[this.preview(s.token)] = s.lastError; });
+		throw new TokenPoolExhaustedError(failures);
+	}
+}
--- a/src/tools.ts
+++ b/src/tools.ts
@@ -1,6 +1,6 @@
 import  * as cheerio from 'cheerio';
 import {$Sync} from '@ztimson/node-utils';
-import {ASet, consoleInterceptor, Http, fn as Fn, decodeHtml} from '@ztimson/utils';
+import {ASet, consoleInterceptor, Http, fn as Fn, decodeHtml, objectMap} from '@ztimson/utils';
 import * as os from 'node:os';
 import {Ai} from './ai.ts';
 import {LLMRequest} from './llm.ts';
@@ -41,28 +41,83 @@ 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 const CliTool: AiTool = {
+export function convertSchema(schema: any): any {
+	if(!schema) return null;
+
+	const convertProp = (prop: any): any => {
+		const converted: any = {
+			type: prop.type || 'string',
+		};
+
+		if(prop.description) converted.description = prop.description;
+		if(prop.default !== undefined) converted.default = prop.default;
+		if(prop.enum) converted.enum = prop.enum;
+		if(prop.pattern) converted.pattern = prop.pattern;
+
+		// Handle array items
+		if(prop.type === 'array' && prop.items) {
+			converted.items = convertProp(prop.items);
+		}
+
+		// Handle object properties
+		if(prop.type === 'object' && prop.items) {
+			converted.properties = objectMap(prop.items, (key, value) => convertProp(value));
+			const required = Object.entries(prop.items).filter(([_, v]: any) => v.required).map(([k]) => k);
+			if(required.length) converted.required = required;
+			converted.additionalProperties = false;
+		}
+
+		// Handle min/max based on type
+		if(prop.min !== undefined) {
+			if(prop.type === 'string' || prop.type === 'array') converted.minLength = prop.min;
+			else converted.minimum = prop.min;
+		}
+		if(prop.max !== undefined) {
+			if(prop.type === 'string' || prop.type === 'array') converted.maxLength = prop.max;
+			else converted.maximum = prop.max;
+		}
+
+		return converted;
+	};
+
+	return {
+		type: 'object',
+		properties: objectMap(schema, (key, value) => convertProp(value)),
+		required: Object.entries(schema).filter(([_, v]: any) => v.required).map(([k]) => k),
+		additionalProperties: false
+	};
+}
+
+export const ExecCliTool: AiTool = {
 	name: 'cli',
 	description: 'Use the command line interface, returns any output',
 	args: {command: {type: 'string', description: 'Command to run', required: true}},
 	fn: (args: {command: string}) => $Sync`${args.command}`
 }

-export const DateTimeTool: AiTool = {
-	name: 'get_datetime',
-	description: 'Get local date / time',
-	args: {},
-	fn: async () => new Date().toString()
+export const ExecJSTool: AiTool = {
+	name: 'exec_javascript',
+	description: 'Execute commonjs javascript',
+	args: {
+		code: {type: 'string', description: 'CommonJS javascript', required: true}
+	},
+	fn: async (args: {code: string}) => {
+		const c = consoleInterceptor(null);
+		const resp = await Fn<any>({console: c}, args.code, true).catch((err: any) => c.output.error.push(err));
+		return {...c.output, return: resp, stdout: undefined, stderr: undefined};
+	}
 }

-export const DateTimeUTCTool: AiTool = {
-	name: 'get_datetime_utc',
-	description: 'Get current UTC date / time',
-	args: {},
-	fn: async () => new Date().toUTCString()
+export const ExecPythonTool: AiTool = {
+	name: 'exec_python',
+	description: 'Execute commonjs javascript',
+	args: {
+		code: {type: 'string', description: 'CommonJS javascript', required: true}
+	},
+	fn: async (args: {code: string}) => ({result: $Sync`python -c "${args.code}"`})
 }

 export const ExecTool: AiTool = {
@@ -76,11 +131,11 @@ export const ExecTool: AiTool = {
 		try {
 			switch(args.language) {
 				case 'cli':
-					return await CliTool.fn({command: args.code}, stream, ai);
+					return await ExecCliTool.fn({command: args.code}, stream, ai);
 				case 'node':
-					return await JSTool.fn({code: args.code}, stream, ai);
+					return await ExecJSTool.fn({code: args.code}, stream, ai);
 				case 'python':
-					return await PythonTool.fn({code: args.code}, stream, ai);
+					return await ExecPythonTool.fn({code: args.code}, stream, ai);
 				default:
 					throw new Error(`Unsupported language: ${args.language}`);
 			}
@@ -90,8 +145,483 @@ export const ExecTool: AiTool = {
 	}
 }

-export const FetchTool: AiTool = {
-	name: 'fetch',
+export const FsDeleteTool = (whitelist: null | string[] = null): AiTool => {
+	return {
+		name: 'fs_delete',
+		description: 'Delete a file or directory',
+		args: {
+			path: {type: 'string', description: 'Path to file or directory', required: true},
+			recursive: {type: 'boolean', description: 'Delete all children', required: false}
+		},
+		fn: async ({path, recursive = false}) => {
+			const {existsSync, rmSync} = await import('fs');
+			const normalizePath = p => p.replace(/\\/g, '/');
+
+			path = normalizePath(path);
+			if(whitelist && !whitelist.some(p => path.startsWith(p))) return {error: 'Permission denied'};
+			if(!existsSync(path)) return {error: 'Path does not exist'};
+
+			rmSync(path, {recursive, force: true});
+			return {success: true, path};
+		}
+	}
+}
+
+export const FsMoveTool = (whitelist: null | string[] = null): AiTool => {
+	return {
+		name: 'fs_move',
+		description: 'Move or rename a file or directory',
+		args: {
+			source: {type: 'string', description: 'Path to source file or directory', required: true},
+			destination: {type: 'string', description: 'Path to destination file or directory', required: true}
+		},
+		fn: async ({source, destination}) => {
+			const {existsSync, renameSync} = await import('fs');
+			const normalizePath = p => p.replace(/\\/g, '/');
+
+			source = normalizePath(source);
+			destination = normalizePath(destination);
+			if(whitelist && !whitelist.some(p => source.startsWith(p) && destination.startsWith(p))) return {error: 'Permission denied'};
+
+			if(!existsSync(source)) return {error: 'Source path does not exist'};
+			if(existsSync(destination)) return {error: 'Destination path already exists'};
+
+			renameSync(source, destination);
+			return {success: true, source, destination};
+		}
+	}
+}
+
+export const FsReadTool = (whitelist: null | string[] = null): AiTool => {
+	return {
+		name: 'fs_read',
+		description: 'Read the contents of a provided path. Works with files and directories',
+		args: {path: {type: 'string', description: 'Path to file or directory', required: true}},
+		fn: async ({path}) => {
+			const {existsSync, lstatSync, readdirSync, readFileSync} = await import('fs');
+			const {join} = await import('path');
+			const normalizePath = p => p.replace(/\\/g, '/');
+
+			path = normalizePath(path);
+			if(whitelist && !whitelist.some(p => path.startsWith(p))) return {error: 'Permission denied'};
+
+			if(!existsSync(path)) return {error: 'Path does not exist'};
+			const stats = lstatSync(path);
+			if(stats.isDirectory()) {
+				const children = readdirSync(path).map(name => {
+					const childPath = normalizePath(join(path, name));
+					const childStats = lstatSync(childPath);
+					return {name, type: childStats.isDirectory() ? 'directory' : 'file', size: childStats.size};
+				});
+				return {type: 'directory', children};
+			}
+			const content = readFileSync(path, 'utf-8');
+			return {type: 'file', content};
+		}
+	}
+}
+
+export const FsSearchTool = (whitelist: null | string[] = null): AiTool => {
+	return {
+		name: 'fs_search',
+		description: 'Scan a directory for matching glob patterns (e.g. "**/*.js", "src/**/*.test.ts")',
+		args: {
+			pattern: {type: 'string', description: 'Glob pattern to match against paths', required: true},
+			root: {type: 'string', description: 'Directory to search from', required: false, default: '.'}
+		},
+		fn: async ({pattern, root = '.'}) => {
+			const {existsSync, lstatSync, readdirSync} = await import('fs');
+			const {join, relative} = await import('path');
+			const normalizePath = p => p.replace(/\\/g, '/');
+
+			root = normalizePath(root);
+			if(!existsSync(root)) return {error: 'Root path does not exist'};
+			if(!lstatSync(root).isDirectory()) return {error: 'Root path is not a directory'};
+
+			if(whitelist && !whitelist.some(p => root.startsWith(p))) return {error: 'Permission denied'};
+
+			const globToRegex = (glob) => {
+				let re = '';
+				for(let i = 0; i < glob.length; i++) {
+					const c = glob[i];
+					if(c === '*') {
+						if(glob[i + 1] === '*') {
+							const isSlash = glob[i + 2] === '/';
+							re += '.*';
+							i += isSlash ? 2 : 1;
+						} else {
+							re += '[^/]*';
+						}
+					} else if(c === '?') {
+						re += '[^/]';
+					} else if('.+^$(){}|[]\\'.includes(c)) {
+						re += '\\' + c;
+					} else {
+						re += c;
+					}
+				}
+				return new RegExp('^' + re + '$');
+			};
+			const regex = globToRegex(pattern);
+
+			const results: any = [];
+			const walk = (dir) => {
+				for(const name of readdirSync(dir)) {
+					const fullPath = normalizePath(join(dir, name));
+					const stats = lstatSync(fullPath);
+					const relPath = normalizePath(relative(root, fullPath));
+					if(regex.test(relPath)) {
+						results.push({path: relPath, type: stats.isDirectory() ? 'directory' : 'file', size: stats.size});
+					}
+					if(stats.isDirectory()) walk(fullPath);
+				}
+			};
+			walk(root);
+
+			return results;
+		}
+	}
+}
+
+export const FsWriteTool = (whitelist: null | string[] = null): AiTool => {
+	return {
+		name: 'fs_write',
+		description: 'Create a directory, write content to a file or preform a find & replace',
+		args: {
+			path: {type: 'string', description: 'Path to file or directory', required: true},
+			content: {type: 'string', description: 'Content to write or replace (Omit to create a directory)'},
+			find: {type: 'string', description: 'Text or regex pattern to match (regex must match pattern: "/pattern/g")'}
+		},
+		fn: async ({path, content, find}) => {
+			const {existsSync, mkdirSync, readFileSync, writeFileSync} = await import('fs');
+			const {dirname} = await import('path');
+			const normalizePath = p => p.replace(/\\/g, '/');
+
+			path = normalizePath(path);
+			if(whitelist && !whitelist.some(p => path.startsWith(p))) return {error: 'Permission denied'};
+
+			if(content === undefined) {
+				mkdirSync(path, {recursive: true});
+				return {success: true, type: 'directory', path};
+			}
+
+			const dir = normalizePath(dirname(path));
+			if(!existsSync(dir)) mkdirSync(dir, {recursive: true});
+
+			if(find && existsSync(path)) {
+				const existing = readFileSync(path, 'utf-8');
+				const regexMatch = find.match(/^\/(.+)\/([gimuy]*)$/);
+				const pattern = regexMatch ? new RegExp(regexMatch[1], regexMatch[2]) : find;
+
+				if(!existing.match(pattern)) return {error: 'Find pattern not found in file'};
+
+				const updated = existing.replace(pattern, content);
+				writeFileSync(path, updated, 'utf-8');
+				return {success: true, type: 'file', path, replaced: true, content: updated};
+			}
+
+			writeFileSync(path, content, 'utf-8');
+			return {success: true, type: 'file', path, content};
+		}
+	}
+}
+
+export const GetPathsTool: AiTool = {
+	name: 'get_paths',
+	description: 'Get the current working directory, and paths to the users home directory',
+	fn: async () => {
+		return {
+			home: os.homedir(),
+			cwd: process.cwd()
+		};
+	}
+}
+
+export const GetDatetimeTool: AiTool = {
+	name: 'get_datetime',
+	description: 'Get local/UTC timestamp',
+	args: {
+		timezone: {type: 'string', description: 'Which timezone to return, defaults to local', enum: ['local', 'utc'], default: 'local'}
+	},
+	fn: ({timezone}) => new Date()[timezone === 'local' ? 'toString' : 'toUTCString']()
+}
+
+export const GetDevice: AiTool = {
+	name: 'get_device',
+	description: 'Get comprehensive system information including hostname, specs, load, storage, and network status',
+	args: {},
+	fn: async () => {
+		const platform = os.platform();
+		const hostname = os.hostname();
+
+		// CPU Info
+		const cpus = os.cpus();
+		const cpuModel = cpus[0].model;
+		const cpuCores = cpus.length;
+
+		// Memory Info
+		const totalMem: any = (os.totalmem() / 1024 / 1024 / 1024).toFixed(2);
+		const freeMem: any = (os.freemem() / 1024 / 1024 / 1024).toFixed(2);
+		const usedMem: any = (totalMem - freeMem).toFixed(2);
+		const memUsage: any = ((usedMem / totalMem) * 100).toFixed(1);
+
+		// Load Average (not available on Windows)
+		const loadAvg = platform === 'win32' ? ['N/A', 'N/A', 'N/A'] : os.loadavg().map(l => l.toFixed(2));
+
+		// Storage Usage
+		let storage = {};
+		if(platform === 'win32') {
+			const ps = $Sync`powershell "Get-PSDrive C | Select-Object Used,Free | ConvertTo-Json"`.trim();
+			const drive = JSON.parse(ps);
+			const used: any = (drive.Used / 1024 / 1024 / 1024).toFixed(2);
+			const free: any = (drive.Free / 1024 / 1024 / 1024).toFixed(2);
+			const total: any = (parseFloat(used) + parseFloat(free)).toFixed(2);
+			const usage: any = ((used / total) * 100).toFixed(1);
+			storage = {
+				filesystem: 'C:',
+				size: `${total} GB`,
+				used: `${used} GB`,
+				available: `${free} GB`,
+				usage: `${usage}%`
+			};
+		} else {
+			const df = $Sync`df -h / | tail -1`.trim();
+			const s = df.split(/\s+/);
+			storage = {
+				filesystem: s[0],
+				size: s[1],
+				used: s[2],
+				available: s[3],
+				usage: s[4]
+			};
+		}
+
+		// Network Status
+		const interfaces = os.networkInterfaces();
+		const activeIfaces = Object.entries(interfaces)
+			.filter(([name]) => name !== 'lo' && !name.includes('Loopback'))
+			.map(([name, addrs]) => {
+				const ipv4 = addrs?.find(a => a.family === 'IPv4');
+				return ipv4 ? {name, ip: ipv4.address} : null;
+			})
+			.filter(Boolean);
+
+		// Internet connectivity check
+		let internet = false;
+		try {
+			if(platform === 'win32') {
+				$Sync`powershell "Test-Connection -ComputerName 8.8.8.8 -Count 1 -Quiet"`;
+			} else {
+				$Sync`ping -c 1 -W 2 8.8.8.8 > /dev/null 2>&1`;
+			}
+			internet = true;
+		} catch {}
+
+		// Uptime
+		const uptime = os.uptime();
+		const days = Math.floor(uptime / 86400);
+		const hours = Math.floor((uptime % 86400) / 3600);
+		const minutes = Math.floor((uptime % 3600) / 60);
+
+		return {
+			hostname,
+			cpu: {
+				model: cpuModel,
+				cores: cpuCores
+			},
+			memory: {
+				total: `${totalMem} GB`,
+				used: `${usedMem} GB`,
+				free: `${freeMem} GB`,
+				usage: `${memUsage}%`
+			},
+			load: {
+				'1min': loadAvg[0],
+				'5min': loadAvg[1],
+				'15min': loadAvg[2]
+			},
+			storage,
+			network: {
+				interfaces: activeIfaces,
+				internet: internet ? 'connected' : 'disconnected'
+			},
+			uptime: `${days}d ${hours}h ${minutes}m`,
+			platform: `${os.type()} ${os.release()}`
+		};
+	}
+}
+
+export const GetWikipediaTool: AiTool = {
+	name: 'get_wikipedia',
+	description: 'Search Wikipedia for matching articles',
+	args: {
+		query: {type: 'string', description: 'Search term or article title', required: true},
+		mode: {type: 'string', description: 'search - look for articles, summary - intro of first found article (default), full - complete first found article', enum: ['search', 'summary', 'full'], default: 'summary'},
+		ua: {type: 'string', description: 'User Agent'},
+	},
+	fn: async ({query, mode, ua}) => {
+		class WikipediaClient {
+			useragent = 'Mozilla/5.0 (Windows NT 10.0; Win64; x64)';
+
+			constructor(useragent: string) {
+				this.useragent = useragent;
+			}
+
+			async get(url) {
+				const resp = await fetch(url, {headers: {'User-Agent': this.useragent}});
+				return resp.json();
+			}
+
+			api(params) {
+				const qs = new URLSearchParams({...params, format: 'json', utf8: '1'}).toString();
+				return this.get(`https://en.wikipedia.org/w/api.php?${qs}`);
+			}
+
+			clean(text) {
+				const cutoffs = ['== See also ==', '== References ==', '== Bibliography ==', '== External links =='];
+				for (const marker of cutoffs) {
+					const idx = text.indexOf(marker);
+					if (idx !== -1) text = text.slice(0, idx);
+				}
+
+				return text
+					.replace(/^={4}\s*(.+?)\s*={4}$/gm, '#### $1')
+					.replace(/^={3}\s*(.+?)\s*={3}$/gm, '### $1')
+					.replace(/^={2}\s*(.+?)\s*={2}$/gm, '## $1')
+					.replace(/\n{3,}/g, '\n\n')
+					.replace(/ {2,}/g, ' ')
+					.replace(/\[\d+]/g, '')
+					.trim();
+			}
+
+			async searchTitles(query: string, limit = 6) {
+				const data = await this.api({action: 'query', list: 'search', srsearch: query, srlimit: limit, srprop: 'snippet'});
+				return data.query?.search || [];
+			}
+
+			async fetchExtract(title: string, introOnly = false) {
+				const params: any = {action: 'query', prop: 'extracts', titles: title, explaintext: 1, redirects: 1};
+				if(introOnly) params.exintro = 1;
+				const data = await this.api(params);
+				const page: any = Object.values(data.query?.pages || {})[0];
+				return this.clean(page?.extract || '');
+			}
+
+			pageUrl(title: string) {
+				return `https://en.wikipedia.org/wiki/${encodeURIComponent(title.replace(/ /g, '_'))}`;
+			}
+
+			stripHtml(text: string) {
+				return text.replace(/<[^>]+>/g, '');
+			}
+
+			async lookup(query: string, detail = 'summary') {
+				const results = await this.searchTitles(query, 6);
+				if(!results.length) return `❌ No Wikipedia articles found for "${query}"`;
+				const title = results[0].title;
+				const url = this.pageUrl(title);
+				const introOnly = detail !== 'full';
+				const content = await this.fetchExtract(title, introOnly);
+				return `## ${title}\n🔗 ${url}\n\n${content}`;
+			}
+
+			async search(query: string) {
+				const results = await this.searchTitles(query, 8);
+				if(!results.length) return `❌ No results for "${query}"`;
+				const lines = [`### Search results for "${query}"\n`];
+				for(let i = 0; i < results.length; i++) {
+					const r = results[i];
+					const snippet = this.stripHtml(r.snippet || '').trim();
+					lines.push(`**${i + 1}. ${r.title}**\n${snippet}\n${this.pageUrl(r.title)}`);
+				}
+				return lines.join('\n\n');
+			}
+		}
+
+		const wiki = new WikipediaClient(ua);
+		if(mode === 'search') return wiki.search(query);
+		return wiki.lookup(query, mode || 'summary');
+	}
+};
+
+export const GeoCodeTool: AiTool = {
+	name: 'geo_code',
+	description: 'Converts coordinates to address OR vice versa',
+	args: {
+		query: {type: 'string', description: 'Search query - coordinates (lat,lon) or address string', required: true},
+	},
+	fn: async ({query}) => {
+		const coordinates = /(-?\d+(?:\.\d+)?).*?,.*?(-?\d+(?:\.\d+)?)/.exec(query);
+		if(coordinates) { // Geolocate
+			const url = `https://nominatim.openstreetmap.org/reverse?format=json&lat=${encodeURIComponent(coordinates[1])}&lon=${encodeURIComponent(coordinates[2])}`;
+			const response = await fetch(url, {headers: {'User-Agent': 'OpenSight/1.0', 'Accept-Language': 'en'}});
+			const data = await response.json();
+			if(data.display_name) return {address: data.display_name, mode: 'geolocate'};
+		} else { // Geocode
+			const url = `https://nominatim.openstreetmap.org/search?format=json&q=${encodeURIComponent(query)}`;
+			const response = await fetch(url, {headers: {'User-Agent': 'OpenSight/1.0'}});
+			const data = await response.json();
+			if(data[0]) return {latitude: parseFloat(data[0].lat), longitude: parseFloat(data[0].lon), mode: 'geocode'};
+		}
+		return {error: 'Not found'};
+	},
+}
+
+export const GeoWeatherTool: AiTool = {
+	name: 'geo_weather',
+	description: 'Gets weather and air quality info for a location and time',
+	args: {
+		query: {type: 'string', description: 'Location - address or place name', required: true},
+		day: {type: 'string', description: 'Date to retrieve (YYYY-MM-DD), defaults to today'},
+	},
+	fn: async ({query, day}) => {
+		day = day || new Date().toISOString().slice(0, 10);
+
+		const geoUrl = `https://nominatim.openstreetmap.org/search?format=json&q=${encodeURIComponent(query)}`;
+		const geoResponse = await fetch(geoUrl, {headers: {'User-Agent': 'OpenSight/1.0'}});
+		const geoData = await geoResponse.json();
+		if(!geoData[0]) return {error: 'Location not found'};
+
+		const lat = parseFloat(geoData[0].lat);
+		const lon = parseFloat(geoData[0].lon);
+
+		const weatherUrl = `https://api.open-meteo.com/v1/forecast?latitude=${lat}&longitude=${lon}&start_date=${day}&end_date=${day}&daily=weathercode,temperature_2m_max,temperature_2m_min,apparent_temperature_max,apparent_temperature_min,precipitation_sum,precipitation_probability_max,windspeed_10m_max,winddirection_10m_dominant,uv_index_max,sunrise,sunset&timezone=auto`;
+		const airUrl = `https://air-quality-api.open-meteo.com/v1/air-quality?latitude=${lat}&longitude=${lon}&start_date=${day}&end_date=${day}&hourly=us_aqi,european_aqi,pm10,pm2_5&timezone=auto`;
+
+		const [weatherResponse, airResponse] = await Promise.all([fetch(weatherUrl), fetch(airUrl)]);
+		const weatherData = await weatherResponse.json();
+		const airData = await airResponse.json();
+
+		const avg = arr => (arr && arr.length) ? arr.reduce((a, b) => a + b, 0) / arr.length : null;
+
+		return {
+			location: geoData[0].display_name,
+			latitude: lat,
+			longitude: lon,
+			elevation: weatherData.elevation,
+			date: day,
+			weatherCode: weatherData.daily?.weathercode?.[0],
+			tempMax: weatherData.daily?.temperature_2m_max?.[0],
+			tempMin: weatherData.daily?.temperature_2m_min?.[0],
+			feelsLikeMax: weatherData.daily?.apparent_temperature_max?.[0],
+			feelsLikeMin: weatherData.daily?.apparent_temperature_min?.[0],
+			precipitation: weatherData.daily?.precipitation_sum?.[0],
+			precipitationChance: weatherData.daily?.precipitation_probability_max?.[0],
+			windSpeedMax: weatherData.daily?.windspeed_10m_max?.[0],
+			windDirection: weatherData.daily?.winddirection_10m_dominant?.[0],
+			uvIndexMax: weatherData.daily?.uv_index_max?.[0],
+			sunrise: weatherData.daily?.sunrise?.[0],
+			sunset: weatherData.daily?.sunset?.[0],
+			usAqi: avg(airData.hourly?.us_aqi),
+			europeanAqi: avg(airData.hourly?.european_aqi),
+			pm10: avg(airData.hourly?.pm10),
+			pm2_5: avg(airData.hourly?.pm2_5),
+		};
+	},
+}
+
+export const WebFetchTool: AiTool = {
+	name: 'web_fetch',
 	description: 'Make HTTP request to URL',
 	args: {
 		url: {type: 'string', description: 'URL to fetch', required: true},
@@ -107,30 +637,59 @@ export const FetchTool: AiTool = {
 	}) => new Http({url: args.url, headers: args.headers}).request({method: args.method || 'GET', body: args.body})
 }

-export const JSTool: AiTool = {
-	name: 'exec_javascript',
-	description: 'Execute commonjs javascript',
+export const WebFlareSolverTool = (host: string) => {
+	return {
+		name: 'web_flaresolverr',
+		description: 'Use a flaresolverr proxy to bypass cloudflare bot detection',
 		args: {
-		code: {type: 'string', description: 'CommonJS javascript', required: true}
+			url: {type: 'string', description: 'URL to fetch', required: true},
+			cmd: {type: 'string', description: 'Flaresolverr cmd', enum: ['request.get', 'request.post'], default: 'request.get'},
+			maxTimeout: {type: 'number', description: 'Fetch time limit', default: 60_000},
+			postData: {type: 'object', description: 'Data to send during request.post requests'},
 		},
-	fn: async (args: {code: string}) => {
-		const c = consoleInterceptor(null);
-		const resp = await Fn<any>({console: c}, args.code, true).catch((err: any) => c.output.error.push(err));
-		return {...c.output, return: resp, stdout: undefined, stderr: undefined};
+		fn: async ({url, cmd, maxTimeout, postData}) => {
+			function toFormUrlEncoded(obj, prefix = '') {
+				const pairs: any = [];
+				for (const key in obj) {
+					if (!obj.hasOwnProperty(key)) continue;
+
+					const value = obj[key];
+					const encodedKey = prefix
+						? `${prefix}[${encodeURIComponent(key)}]`
+						: encodeURIComponent(key);
+
+					if (value === null || value === undefined) {
+						pairs.push(`${encodedKey}=`);
+					} else if (typeof value === 'object' && !Array.isArray(value)) {
+						pairs.push(toFormUrlEncoded(value, encodedKey));
+					} else if (Array.isArray(value)) {
+						value.forEach(item => {
+							pairs.push(`${encodedKey}[]=${encodeURIComponent(item)}`);
+						});
+					} else {
+						pairs.push(`${encodedKey}=${encodeURIComponent(value)}`);
 					}
 				}

-export const PythonTool: AiTool = {
-	name: 'exec_javascript',
-	description: 'Execute commonjs javascript',
-	args: {
-		code: {type: 'string', description: 'CommonJS javascript', required: true}
-	},
-	fn: async (args: {code: string}) => ({result: $Sync`python -c "${args.code}"`})
+				return pairs.join('&');
 			}

-export const ReadWebpageTool: AiTool = {
-	name: 'read_webpage',
+			const res = await fetch(host + '/v1', {
+				method: 'POST',
+				headers: {'Content-Type': 'application/json'},
+				body: JSON.stringify({cmd, url, maxTimeout, postData: postData ? toFormUrlEncoded(postData) : undefined}),
+			});
+
+			if(!res.ok) throw new Error(`FlareSolverr HTTP error: ${res.status} ${res.statusText}`);
+			const data = await res.json();
+			if(data.status !== 'ok') throw new Error(`FlareSolverr error: ${data.message ?? data.status}`);
+			return data.solution.response;
+		}
+	}
+}
+
+export const WebReadTool: AiTool = {
+	name: 'web_read',
 	description: 'Extract clean content from webpages, or convert media/documents to accessible formats',
 	args: {
 		url: {type: 'string', description: 'URL to read', required: true},
@@ -258,94 +817,3 @@ export const WebSearchTool: AiTool = {
 		return results;
 	}
 }
-
-class WikipediaClient {
-	private async get(url: string): Promise<any> {
-		const resp = await fetch(url, {headers: {'User-Agent': UA}});
-		return resp.json();
-	}
-
-	private api(params: Record<string, any>): Promise<any> {
-		const qs = new URLSearchParams({...params, format: 'json', utf8: '1'}).toString();
-		return this.get(`https://en.wikipedia.org/w/api.php?${qs}`);
-	}
-
-	private clean(text: string): string {
-		return text.replace(/\n{3,}/g, '\n\n').replace(/ {2,}/g, ' ').replace(/\[\d+\]/g, '').trim();
-	}
-
-	private truncate(text: string, max: number): string {
-		if(text.length <= max) return text;
-		const cut = text.slice(0, max);
-		const lastPara = cut.lastIndexOf('\n\n');
-		return lastPara > max * 0.7 ? cut.slice(0, lastPara) : cut;
-	}
-
-	private async searchTitles(query: string, limit = 6): Promise<any[]> {
-		const data = await this.api({action: 'query', list: 'search', srsearch: query, srlimit: limit, srprop: 'snippet'});
-		return data.query?.search || [];
-	}
-
-	private async fetchExtract(title: string, intro = false): Promise<string> {
-		const params: any = {action: 'query', prop: 'extracts', titles: title, explaintext: 1, redirects: 1};
-		if(intro) params.exintro = 1;
-		const data = await this.api(params);
-		const page = Object.values(data.query?.pages || {})[0] as any;
-		return this.clean(page?.extract || '');
-	}
-
-	private pageUrl(title: string): string {
-		return `https://en.wikipedia.org/wiki/${encodeURIComponent(title.replace(/ /g, '_'))}`;
-	}
-
-	private stripHtml(text: string): string {
-		return text.replace(/<[^>]+>/g, '');
-	}
-
-	async lookup(query: string, detail: 'intro' | 'full' = 'intro'): Promise<string> {
-		const results = await this.searchTitles(query, 6);
-		if(!results.length) return `❌ No Wikipedia articles found for "${query}"`;
-		const title = results[0].title;
-		const url = this.pageUrl(title);
-		const content = await this.fetchExtract(title, detail === 'intro');
-		const text = this.truncate(content, detail === 'intro' ? 2000 : 8000);
-		return `## ${title}\n🔗 ${url}\n\n${text}`;
-	}
-
-	async search(query: string): Promise<string> {
-		const results = await this.searchTitles(query, 8);
-		if(!results.length) return `❌ No results for "${query}"`;
-		const lines = [`### Search results for "${query}"\n`];
-		for(let i = 0; i < results.length; i++) {
-			const r = results[i];
-			const snippet = this.truncate(this.stripHtml(r.snippet || ''), 150);
-			lines.push(`**${i + 1}. ${r.title}**\n${snippet}\n${this.pageUrl(r.title)}`);
-		}
-		return lines.join('\n\n');
-	}
-}
-
-export const WikipediaLookupTool: AiTool = {
-	name: 'wikipedia_lookup',
-	description: 'Get Wikipedia article content',
-	args: {
-		query: {type: 'string', description: 'Topic or article title', required: true},
-		detail: {type: 'string', description: 'Content level: "intro" (summary, default) or "full" (complete article)', enum: ['intro', 'full'], default: 'intro'}
-	},
-	fn: async (args: {query: string; detail?: 'intro' | 'full'}) => {
-		const wiki = new WikipediaClient();
-		return wiki.lookup(args.query, args.detail || 'intro');
-	}
-};
-
-export const WikipediaSearchTool: AiTool = {
-	name: 'wikipedia_search',
-	description: 'Search Wikipedia for matching articles',
-	args: {
-		query: {type: 'string', description: 'Search terms', required: true}
-	},
-	fn: async (args: {query: string}) => {
-		const wiki = new WikipediaClient();
-		return wiki.search(args.query);
-	}
-};
--- a/src/vision.ts
+++ b/src/vision.ts
@@ -12,12 +12,31 @@ export class Vision {
 	 */
 	ocr(path: string): AbortablePromise<string | null> {
 		let worker: any;
+		let reject: (err: any) => void;
+
+		const handler = (err: Error) => {
+			if(err.stack?.includes('tesseract.js')) {
+				process.off('uncaughtException', handler);
+				reject?.(err);
+				return;
+			}
+			throw err;
+		};
+		process.on('uncaughtException', handler);
+
 		const p = (async () => {
 			worker = await createWorker(this.ai.options.ocr || 'eng', 2, {cachePath: this.ai.options.path});
-			const {data} = await worker.recognize(path);
-			await worker.terminate();
-			return data.text.trim() || null;
-		})().finally(() => worker?.terminate());
+			return await new Promise<string | null>((res, rej) => {
+				reject = rej;
+				worker.recognize(path)
+					.then(({data}: any) => res(data.text.trim() || null))
+					.catch(rej);
+			});
+		})().finally(() => {
+			process.off('uncaughtException', handler);
+			worker?.terminate();
+		});
+
 		return Object.assign(p, {abort: () => worker?.terminate()});
 	}
 }
--- a/tsconfig.json
+++ b/tsconfig.json
@@ -4,7 +4,10 @@
 		"target": "ESNext",
 		"useDefineForClassFields": true,
 		"module": "ESNext",
-		"lib": ["ESNext"],
+		"lib": [
+			"ESNext",
+			"dom"
+		],
 		"skipLibCheck": true,

 		/* Bundler mode */
Author	SHA1	Message	Date
ztimson	0a6f1e4d62	Refined memory management prompts All checks were successful Publish Library / Build NPM Project (push) Successful in 1m18s Details Publish Library / Tag Version (push) Successful in 20s Details	2026-08-25 10:03:36 -04:00
ztimson	08a351e028	Better memory management All checks were successful Publish Library / Build NPM Project (push) Successful in 59s Details Publish Library / Tag Version (push) Successful in 11s Details	2026-08-24 14:42:10 -04:00
ztimson	85c01d3ef1	Added official file support All checks were successful Publish Library / Build NPM Project (push) Successful in 30s Details Publish Library / Tag Version (push) Successful in 10s Details	2026-08-17 15:50:48 -04:00
ztimson	5826573d5c	Added official file support All checks were successful Publish Library / Build NPM Project (push) Successful in 50s Details Publish Library / Tag Version (push) Successful in 13s Details	2026-08-17 15:16:32 -04:00
ztimson	797a40a566	Added official file support All checks were successful Publish Library / Build NPM Project (push) Successful in 58s Details Publish Library / Tag Version (push) Successful in 13s Details	2026-08-16 15:40:50 -04:00
ztimson	7308927a3c	max token rename All checks were successful Publish Library / Build NPM Project (push) Successful in 35s Details Publish Library / Tag Version (push) Successful in 14s Details	2026-08-05 16:16:30 -04:00
ztimson	04f038ba65	Memory prompt refinement All checks were successful Publish Library / Build NPM Project (push) Successful in 38s Details Publish Library / Tag Version (push) Successful in 19s Details	2026-08-05 13:14:21 -04:00
ztimson	d42f58d710	Memory refinement All checks were successful Publish Library / Build NPM Project (push) Successful in 54s Details Publish Library / Tag Version (push) Successful in 11s Details	2026-08-05 12:22:13 -04:00
ztimson	878a8794ee	Rebuild graph edges on changes All checks were successful Publish Library / Build NPM Project (push) Successful in 46s Details Publish Library / Tag Version (push) Successful in 19s Details	2026-08-04 17:05:58 -04:00
ztimson	3f1289d993	Small agent tweaks All checks were successful Publish Library / Build NPM Project (push) Successful in 49s Details Publish Library / Tag Version (push) Successful in 9s Details	2026-08-04 14:33:28 -04:00
ztimson	077f75cdd9	Fixed delegate agent history... again All checks were successful Publish Library / Build NPM Project (push) Successful in 48s Details Publish Library / Tag Version (push) Successful in 13s Details	2026-08-04 13:58:47 -04:00
ztimson	566d84fd7a	Added memory graph traversal helpers All checks were successful Publish Library / Build NPM Project (push) Successful in 43s Details Publish Library / Tag Version (push) Successful in 14s Details	2026-08-04 12:58:39 -04:00
ztimson	4230b534fc	bump 1.4.0 All checks were successful Publish Library / Build NPM Project (push) Successful in 1m17s Details Publish Library / Tag Version (push) Successful in 14s Details	2026-08-04 12:44:41 -04:00
ztimson	119f8472f2	token pools Some checks failed Publish Library / Tag Version (push) Has been cancelled Details Publish Library / Build NPM Project (push) Has been cancelled Details	2026-08-04 12:44:21 -04:00
ztimson	9c04e58c63	Pass deligate subagents full history, improved memory managment	2026-08-04 12:24:23 -04:00
ztimson	7fbb42c26a	improved subagent instructions	2026-08-04 12:03:31 -04:00
ztimson	be08db8e2c	Attach tps to response promise All checks were successful Publish Library / Build NPM Project (push) Successful in 42s Details Publish Library / Tag Version (push) Successful in 9s Details	2026-08-04 09:48:20 -04:00
ztimson	497f051c62	bump 1.3.5 All checks were successful Publish Library / Build NPM Project (push) Successful in 52s Details Publish Library / Tag Version (push) Successful in 7s Details	2026-08-04 09:30:45 -04:00
ztimson	62fbe73b22	Added tps + duration to AI history All checks were successful Publish Library / Build NPM Project (push) Successful in 52s Details Publish Library / Tag Version (push) Successful in 11s Details	2026-08-04 09:26:57 -04:00
ztimson	d53b1c6328	Removed <tool> blocks from responses All checks were successful Publish Library / Build NPM Project (push) Successful in 39s Details Publish Library / Tag Version (push) Successful in 11s Details	2026-08-03 20:23:22 -04:00
ztimson	89619e211e	Fixed message history and response All checks were successful Publish Library / Build NPM Project (push) Successful in 59s Details Publish Library / Tag Version (push) Successful in 22s Details	2026-08-03 19:30:39 -04:00
ztimson	afc6653364	fixed openai system calls in history breaking anthropic calls All checks were successful Publish Library / Build NPM Project (push) Successful in 55s Details Publish Library / Tag Version (push) Successful in 21s Details	2026-08-02 22:35:17 -04:00
ztimson	68e72445a2	Keep recent memories in context All checks were successful Publish Library / Build NPM Project (push) Successful in 53s Details Publish Library / Tag Version (push) Successful in 17s Details	2026-08-01 21:42:05 -04:00
ztimson	1aa6cdf329	Agent/subagent support All checks were successful Publish Library / Build NPM Project (push) Successful in 45s Details Publish Library / Tag Version (push) Successful in 15s Details	2026-08-01 18:28:16 -04:00
ztimson	d022a5ef4d	Improved levenshtein fuzzy match All checks were successful Publish Library / Build NPM Project (push) Successful in 52s Details Publish Library / Tag Version (push) Successful in 17s Details	2026-08-01 12:00:26 -04:00
ztimson	a1d438a20a	Tools can now emit "done" event and end chat early gracefully All checks were successful Publish Library / Build NPM Project (push) Successful in 1m0s Details Publish Library / Tag Version (push) Successful in 9s Details	2026-07-31 17:49:06 -04:00
ztimson	52a9e3aaa4	Fixed history poisoning on empty tool response All checks were successful Publish Library / Build NPM Project (push) Successful in 51s Details Publish Library / Tag Version (push) Successful in 13s Details	2026-07-30 22:12:49 -04:00
ztimson	a7aec4ee29	Improved memory prompt slightly All checks were successful Publish Library / Build NPM Project (push) Successful in 1m9s Details Publish Library / Tag Version (push) Successful in 19s Details	2026-07-30 16:00:03 -04:00
ztimson	dda2d4c2a3	Bump 1.2.8 All checks were successful Publish Library / Build NPM Project (push) Successful in 49s Details Publish Library / Tag Version (push) Successful in 7s Details	2026-07-29 22:35:29 -04:00
ztimson	58e0e488e4	Added Geo, FS and flarescraperr tools Some checks failed Publish Library / Tag Version (push) Has been cancelled Details Publish Library / Build NPM Project (push) Has been cancelled Details	2026-07-29 22:34:51 -04:00
ztimson	8dfcd06752	More memory fixes All checks were successful Publish Library / Build NPM Project (push) Successful in 43s Details Publish Library / Tag Version (push) Successful in 14s Details	2026-07-29 22:11:09 -04:00
ztimson	14f6cdd313	Personal file memory organization instructions All checks were successful Publish Library / Build NPM Project (push) Successful in 33s Details Publish Library / Tag Version (push) Successful in 12s Details	2026-07-27 22:47:48 -04:00
ztimson	73d6ee0f2a	Personal file memory organization instructions All checks were successful Publish Library / Build NPM Project (push) Successful in 45s Details Publish Library / Tag Version (push) Successful in 12s Details	2026-07-27 22:39:06 -04:00
ztimson	bee4085666	updatememory awaits full result Some checks failed Publish Library / Tag Version (push) Has been cancelled Details Publish Library / Build NPM Project (push) Has been cancelled Details	2026-07-27 22:34:36 -04:00
ztimson	3b5c71de7c	Improved memory management All checks were successful Publish Library / Build NPM Project (push) Successful in 40s Details Publish Library / Tag Version (push) Successful in 14s Details	2026-07-27 20:10:09 -04:00
ztimson	8229e02a52	Improved memory management All checks were successful Publish Library / Build NPM Project (push) Successful in 44s Details Publish Library / Tag Version (push) Successful in 11s Details	2026-07-27 14:25:24 -04:00
ztimson	a6fb8ae828	New memory system All checks were successful Publish Library / Build NPM Project (push) Successful in 1m5s Details Publish Library / Tag Version (push) Successful in 11s Details	2026-07-27 03:59:39 -04:00
ztimson	d1230bcaad	Updated wiki tool All checks were successful Publish Library / Build NPM Project (push) Successful in 55s Details Publish Library / Tag Version (push) Successful in 14s Details	2026-07-26 12:18:57 -04:00
ztimson	2d49c9aa80	Removed redundant llama protocol (Use openai) All checks were successful Publish Library / Build NPM Project (push) Successful in 44s Details Publish Library / Tag Version (push) Successful in 13s Details	2026-07-11 19:33:02 -04:00
ztimson	9a39f00f94	Diarization fix All checks were successful Publish Library / Build NPM Project (push) Successful in 44s Details Publish Library / Tag Version (push) Successful in 13s Details	2026-07-11 18:36:24 -04:00
ztimson	436757daad	Added new json output support Some checks failed Publish Library / Build NPM Project (push) Failing after 1m2s Details Publish Library / Tag Version (push) Has been skipped Details	2026-07-11 18:27:55 -04:00
ztimson	69b3297bb3	Proper error handling for OCR All checks were successful Publish Library / Build NPM Project (push) Successful in 1m25s Details Publish Library / Tag Version (push) Successful in 10s Details	2026-06-09 11:21:12 -04:00
ztimson	710c6ce52c	Proper error handling for OCR Some checks failed Publish Library / Tag Version (push) Has been cancelled Details Publish Library / Build NPM Project (push) Has been cancelled Details	2026-06-09 11:20:50 -04:00
ztimson	4ac3036000	Proper error handling for OCR All checks were successful Publish Library / Build NPM Project (push) Successful in 43s Details Publish Library / Tag Version (push) Successful in 11s Details	2026-06-09 09:41:09 -04:00