Improved memory management

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "1.1.0",
+	"version": "1.2.3",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",

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;

src/kd-tree.ts

@@ -0,0 +1,334 @@
+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 dims: number;
+	private readonly distanceFn: (a: number[], b: number[]) => 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);
+	}
+}

src/llm.ts

@@ -6,10 +6,9 @@ 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} from './memory.ts';
 
 export type AnthropicConfig = {proto: 'anthropic', token: string};
-export type OllamaConfig = {proto: 'llama', host: string};
 export type OpenAiConfig = {proto: 'openai', host?: string, token: string};
 
 export type LLMMessage = {
@@ -56,7 +55,7 @@ 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;
 	/** Model to use for memory operations */
 	memoryModel?: string;
 	/** Skill documents the AI can browse and read on demand */
@@ -95,7 +94,6 @@ 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 == 'llama') this.models[model] = new OpenAi(this.ai, config.host, 'ignored', model, true);
 			else if(config.proto == 'openai') this.models[model] = new OpenAi(this.ai, config.host || null, config.token, model);
 		});
 		this.memoryManager = new MemoryManager(this);
@@ -192,19 +190,20 @@ class LLM {
 			}
 
 			// Memory
-			if(options.memory) {
+			if (options.memory) {
-				const relevant = await this.memoryManager.recollect(message, options.memory, 1);
+				const mems = options.memory instanceof MemoryCache ? options.memory.memories : options.memory;
+				const relevant = await this.memoryManager.recollect(message, options.memory, 5);
 				prompts.unshift(`You have access to the following memory files:
-${options.memory.map(m => `- ${m.name}: ${m.description}`).join('\n')}
+${mems.map(m => `- ${m.name}: ${m.description}`).join('\n')}
 ${relevant.length ? `
-The closest memory has been added primitively:
+Relevant memories have been preloaded:
-\`\`\`
+${relevant.map(r => `
-Name: ${relevant[0].name}
+**${r.name}**
-Description: ${relevant[0].description}
+${r.description}
-${relevant[0].content}
+${r.content}
-\`\`\`
+`).join('\n---\n')}
-`: ''}`.trim());
+` : ''}`.trim());
-				tools.push(this.memoryManager.tools.read(<Memory[]>options.memory));
+				tools.push(this.memoryManager.tools.read(options.memory));
 			}
 
 			prompts.unshift(options.system || this.ai.options.llm?.system || '');
@@ -217,7 +216,7 @@ ${relevant[0].content}
 
 			// Auto-memorize before compressing
 			if(options.compress && this.estimateTokens(history) >= options.compress.max) {
-				if(options.memory) await this.memoryManager.memorize(history, options.memory, options);
+				if(options.memory) await this.memoryManager.memorize(history, options.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);
 			}
@@ -230,7 +229,7 @@ ${relevant[0].content}
 	 * 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> {
+	async updateMemory(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest = {}): Promise<void> {
 		await this.memoryManager.memorize(history, memories, {model: this.defaultModel, ...options});
 	}
 
@@ -429,9 +428,8 @@ ${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 == 'llama') this.models[name] = new OpenAi(this.ai, config.host, 'not-needed', name, true);
 		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;
 	}
@@ -443,12 +441,11 @@ ${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 == 'llama') this.models[model] = new OpenAi(this.ai, config.host, 'not-needed', model, true);
 			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] ?? '';

src/memory.ts

@@ -1,177 +1,501 @@
-// memory.ts
 import {LLMRequest, LLMMessage} from './llm.ts';
 import {AiTool} from './tools.ts';
+import {KDTree, KDPoint} from './kd-tree.ts';
 
-/** Background information the AI will be fed as a knowledge document */
 export type Memory = {
-	/** Memory subject */
 	name: string;
-	/** Short description of what this document contains - used for RAG retrieval */
 	description: string;
-	/** Full markdown content of the document */
 	content: string;
-	/** Embedding vector of the description - used for similarity search */
 	embedding: number[];
+	links: string[];
+	backlinks: string[];
 }
 
-export type MemoryCollection = {
+type MemoryRef = {
-	/** Memory subject */
 	name: string;
-	/** Short description - required if isNew */
+	description: string;
-	description?: string;
+}
-	/** Extracted facts to merge */
+
+type FactBucket = {
+	subject: string;
 	facts: string[];
 }
 
+export type MemoryNode = {
+	name: string;
+	missing: boolean;
+	links: string[];
+	backlinks: string[];
+}
+
+export function buildMemoryGraph(memories: Memory[] | MemoryCache): MemoryNode[] {
+	const mems = memories instanceof MemoryCache ? memories.memories : memories;
+	const nameSet = new Set(mems.map(m => m.name));
+	const ghosts = new Set<string>();
+
+	for (const m of mems) {
+		for (const link of m.links) {
+			if (!nameSet.has(link)) ghosts.add(link);
+		}
+	}
+
+	return [
+		...mems.map(m => ({
+			name: m.name,
+			missing: false,
+			links: m.links,
+			backlinks: m.backlinks,
+		})),
+		...[...ghosts].map(name => ({
+			name,
+			missing: true,
+			links: [],
+			backlinks: mems
+				.filter(m => m.links.includes(name))
+				.map(m => m.name),
+		}))
+	];
+}
+
+function extractLinks(content: string): string[] {
+	const matches = content.matchAll(/\[\[([^\]]+)\]\]/g);
+	return [...new Set([...matches].map(m => m[1].trim()))];
+}
+
+function rebuildBacklinks(memories: Memory[]): void {
+	for (const m of memories) m.backlinks = [];
+	for (const m of memories) {
+		for (const link of m.links) {
+			const target = memories.find(t => t.name === link);
+			if (target) target.backlinks.push(m.name);
+		}
+	}
+}
+
+function cosineDistance(a: number[], b: number[]): number {
+	let dot = 0, normA = 0, normB = 0;
+	for (let i = 0; i < a.length; i++) {
+		dot   += a[i] * b[i];
+		normA += a[i] * a[i];
+		normB += b[i] * b[i];
+	}
+	const denom = Math.sqrt(normA) * Math.sqrt(normB);
+	return denom === 0 ? 1 : 1 - dot / denom;
+}
+
+export class MemoryCache {
+	private tree: KDTree<MemoryRef>;
+	public memories: Memory[];
+
+	constructor(memories: Memory[]) {
+		this.memories = memories;
+		this.tree = this.buildTree();
+	}
+
+	private buildTree(): KDTree<MemoryRef> {
+		const embedded = this.memories.filter(m => m.embedding?.length);
+		if(!embedded.length) return new KDTree<MemoryRef>(0);
+
+		const dims = embedded[0].embedding.length;
+		const points: KDPoint<MemoryRef>[] = embedded.map(m => ({
+			vector: m.embedding,
+			payload: {name: m.name, description: m.description},
+		}));
+
+		return new KDTree<MemoryRef>(dims, 'cosine', points);
+	}
+
+	search(query: number[], limit: number): MemoryRef[] {
+		const results = this.tree.knn(query, limit);
+		return results.map(r => r.point.payload);
+	}
+
+	add(memory: Memory): void {
+		this.memories.push(memory);
+		this.rebuild();
+	}
+
+	update(memory: Memory): void {
+		const idx = this.memories.findIndex(m => m.name === memory.name);
+		if (idx !== -1) {
+			this.memories[idx] = memory;
+		} else {
+			this.memories.push(memory);
+		}
+		this.rebuild();
+	}
+
+	rebuild(): void {
+		this.tree = this.buildTree();
+	}
+
+	rebuildLinks(): void {
+		rebuildBacklinks(this.memories);
+	}
+}
+
 export class MemoryManager {
 
 	tools = {
-		edit: (memory: Memory): AiTool => ({
+		read: (memories: Memory[] | MemoryCache): 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',
-			args: {
-				content: {type: 'string', description: 'New content', required: true},
-				start: {type: 'number', description: 'First line to replace (0-indexed, inclusive). Omit to append.'},
-				end: {type: 'number', description: 'Last line to replace (0-indexed, inclusive). Omit to replace from start to end of doc.'},
-			},
-			fn: (args: any) => {
-				const lines = memory.content ? memory.content.split('\n') : [];
-				const newLines = args.content.split('\n');
-				if(args.start === undefined) lines.push(...newLines);
-				else if(args.end === undefined) lines.splice(args.start, lines.length - args.start, ...newLines);
-				else lines.splice(args.start, args.end - args.start + 1, ...newLines);
-				memory.content = lines.join('\n');
-				return memory.content;
-			}
-		}),
-		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',
+			description: 'Read the full content of a memory document',
 			args: {
 				name: {type: 'string', description: 'Exact memory name', required: true},
 			},
-			fn: (args: any) => {
+			fn:(args: any) => {
-				const mem = memories.find(m => m.name === args.name);
+				const mems = memories instanceof MemoryCache ? memories.memories : memories;
+				const mem = mems.find(m => m.name === args.name);
 				if(!mem) return 'Document not found';
-				return `Name: ${mem.name}\nDescription: ${mem.description}\n\n${mem.content}`;
+				return this.formatMemory(mem);
 			}
 		}),
+	};
+
+	constructor(private llm: any) {}
+
+	private cosineSearch(query: number[], memories: Memory[], limit: number): MemoryRef[] {
+		const scored = memories
+			.filter(m => m.embedding?.length)
+			.map(m => ({
+				ref: {name: m.name, description: m.description},
+				distance: cosineDistance(query, m.embedding)
+			}))
+			.sort((a, b) => a.distance - b.distance)
+			.slice(0, limit);
+		return scored.map(s => s.ref);
 	}
 
-	constructor(private llm: any, private model?: string) {}
+	private createNode(name: string, memories: Memory[]): Memory {
+		const existing = memories.find(m => m.name === name);
+		if(existing) return existing;
+		return {
+			name,
+			description: '',
+			content: '',
+			embedding: [],
+			links: [],
+			backlinks: [],
+		};
+	}
 
-	/**
+	private formatMemory(mem: Memory): string {
-	 * Extracts facts from conversation and groups them into individual memories
+		return [
-	 * @param {string} conversation Full conversation formatted as [role]: content
+			`# ${mem.name}`,
-	 * @param {Memory[]} memories The user's memory documents
+			mem.description ? `> ${mem.description}` : '',
-	 * @param {LLMRequest} options LLM options
+			mem.links.length ? `**Links:** ${mem.links.map(l => `[[${l}]]`).join(', ')}` : '',
-	 * @returns {Promise<MemoryCollection[]>} Fact pools grouped by target document
+			mem.backlinks.length ? `**Referenced by:** ${mem.backlinks.map(l => `[[${l}]]`).join(', ')}` : '',
-	 */
+			'',
-	private async extract(conversation: string, memories: Memory[], options: LLMRequest): Promise<MemoryCollection[]> {
+			mem.content,
-		const existingDocs = memories.map(m => `Name: ${m.name}\nDescription: ${m.description}`).join('\n\n');
+		].filter(l => l !== undefined).join('\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.
+	private listNodes(memories: Memory[]): MemoryRef[] {
+		return memories.map(m => ({name: m.name, description: m.description}));
+	}
 
-Existing documents:\n${existingDocs || 'None yet.'}`,
+	async recollect(query: string, memories: Memory[] | MemoryCache, limit = 5, graphDepth = 1): Promise<Memory[]> {
-			tools: [this.tools.extract(pools)]
+		const mem: Memory[] = memories instanceof MemoryCache ? memories.memories : memories;
+		if(!mem.length) return [];
+		const [e] = await this.llm.embedding(query);
+		if(!e) return [];
+
+		let vectorResults: MemoryRef[];
+		if(memories instanceof MemoryCache) vectorResults = memories.search(e.embedding, limit);
+		else vectorResults = this.cosineSearch(e.embedding, mem, limit);
+		const found = new Set<string>(vectorResults.map(r => r.name));
+
+		if(graphDepth > 0) {
+			const frontier = [...found];
+			for(let depth = 0; depth < graphDepth; depth++) {
+				const next: string[] = [];
+				for(const name of frontier) {
+					const node = mem.find(m => m.name === name);
+					if(!node) continue;
+					for(const link of node.links) {
+						if(!found.has(link) && mem.find(m => m.name === link)) {
+							found.add(link);
+							next.push(link);
+						}
+					}
+				}
+				frontier.splice(0, frontier.length, ...next);
+				if(!frontier.length) break;
+			}
+		}
+
+		const vectorOrder = vectorResults.map(r => r.name);
+		const graphExpansions = [...found].filter(n => !vectorOrder.includes(n));
+		const ordered = [...vectorOrder, ...graphExpansions];
+		return ordered.map(n => mem.find(m => m.name === n)!).filter(Boolean);
+	}
+
+	async memorize(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest): Promise<void> {
+		const mem = memories instanceof MemoryCache ? memories.memories : memories;
+		const conversation = history
+			.filter(h => h.role === 'user' || h.role === 'assistant')
+			.map(h => `[${h.role}]: ${h.content}`).join('\n\n').trim();
+
+		if(conversation) {
+			const buckets = await this.factAgent(conversation, mem, options);
+			if(buckets.length) {
+				await Promise.all(buckets.map(async bucket => {
+					const node = await this.organizingAgent(bucket, mem, options);
+					if(!mem.find(m => m.name === node.name)) mem.push(node);
+					await this.docAgent(node, bucket, mem, options);
+				}));
+			}
+		}
+
+		// Auto-compress old journals
+		const weekAgo = Date.now() - (7 * 24 * 60 * 60 * 1000);
+		const oldDailies = mem.filter(m => {
+			const journal = /^Journal\/(\d{4}-\d{2}-\d{2}$)/.exec(m.name);
+			return journal && new Date(journal[1]).getTime() < weekAgo;
 		});
-		return pools;
+
+		if(oldDailies.length) {
+			const byMonth = new Map<string, Memory[]>();
+			for(const daily of oldDailies) {
+				const match = daily.name.match(/^Journal\/(\d{4}-\d{2})-\d{2}$/);
+				if(!match) continue;
+				const monthKey = match[1];
+				if(!byMonth.has(monthKey)) byMonth.set(monthKey, []);
+				byMonth.get(monthKey)!.push(daily);
 			}
 
-	/**
+			for(const [monthKey, entries] of byMonth) {
-	 * Bot 2 - Editor: merges a pool of facts into a specific document using surgical line-based edits.
+				const monthlyPath = `Journal/${monthKey}`;
-	 * Receives full document content and uses read + amend tools to make precise edits.
+				let monthly = mem.find(m => m.name === monthlyPath);
-	 * @param {MemoryCollection} newMem The fact pool to merge
+				if(!monthly) {
-	 * @param {Memory[]} memories The user's memory documents
+					monthly = this.createNode(monthlyPath, mem);
-	 * @param {LLMRequest} options LLM options
+					mem.push(monthly);
-	 */
+				}
-	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 => `- ${f}`).join('\n'),
+				const bucket: FactBucket = {
+					subject: monthlyPath,
+					facts: entries.flatMap(e => e.content.split('\n').filter(line => line.trim())),
+				};
+
+				await this.docAgent(monthly, bucket, mem, options);
+				for(const daily of entries) {
+					const idx = mem.indexOf(daily);
+					if(idx !== -1) mem.splice(idx, 1);
+				}
+			}
+		}
+
+		if (memories instanceof MemoryCache) {
+			memories.rebuildLinks();
+			memories.rebuild();
+		} else {
+			rebuildBacklinks(mem);
+		}
+	}
+
+	private async docAgent(node: Memory, bucket: FactBucket, memories: Memory[], options: LLMRequest): Promise<void> {
+		let finalContent = node.content;
+		const isJournalCompression = node.name.match(/^Journal\/\d{4}-\d{2}$/);
+		const systemPrompt = isJournalCompression
+			? `You are a journal compressor. Condense the daily entries below into a monthly summary.
+
+Format:
+# ${node.name}
+
+## Themes
+(Recurring topics, moods, patterns)
+
+## Key Events
+(Important moments, decisions, milestones)
+
+## Notable Conversations
+(Significant discussions or revelations)
+
+Rules:
+- Use [[WikiLinks]] to reference permanent notes using full paths like [[People/Sarah]] or [[Projects/Website]]
+- Keep it concise but preserve emotional/temporal context
+- Discard filler but keep things the user vented about or cared about
+- If a fact belongs in a permanent note, link to it instead of duplicating
+
+Current monthly summary:
+\`\`\`markdown
+${node.content || '(empty — first compression for this month)'}
+\`\`\``
+			: `You are a knowledge base editor. Integrate the provided facts into the document below.
+
+Formatting rules:
+- Use Obsidian-style markdown: # headings, **bold** for key terms, bullet lists for facts
+- Link related concepts with [[WikiLink]] notation using full paths like [[People/Sarah]] or [[Projects/Website]]
+- You may create links to nodes that don't exist yet if the concept is important
+- Keep the document concise, factual, and human-readable
+- Resolve any contradictions between old content and new facts (new facts win)
+- Do not add filler, preamble, or AI commentary — just clean knowledge documents
+
+All nodes:
+${this.listNodes(memories).map(n => n.name).join(', ') || 'none'}
+
+Current document:
+\`\`\`markdown
+${node.content || '(empty — this is a new document)'}
+\`\`\``;
+
+		await this.llm.ask(
+			`New facts to integrate:\n${bucket.facts.map(f => `- ${f}`).join('\n')}`,
 			{
-				model: this.model || options.model,
+				model: options.model,
-				temperature: 0.2,
+				temperature: 0.3,
-				system: `You are a document editor. Merge the users list of facts into the following document using the \`edit_memory\` tool; call it as many times as necessary:
+				system: systemPrompt,
-\`\`\`
+				tools: [{
-${mem.content}
+					name: 'update_document',
-\`\`\``,
+					description: 'Write the complete updated document content',
-				tools: [this.tools.edit(mem)]
+					args: {
+						description: {type: 'string', description: 'One-line description of what this document covers, no formatting or emojis', required: true},
+						content: {type: 'string', description: 'Fully updated document in markdown', required: true},
+					},
+					fn:(args: any) => {
+						node.description = args.description;
+						finalContent = args.content;
+						return 'Saved';
+					}
+				}]
 			}
 		);
 
-		if(isNew || mem.description !== existing?.description) {
+		node.content = finalContent;
-			const e = await this.llm.embedding(mem.description);
+		node.links = extractLinks(finalContent);
-			mem.embedding = e?.[0]?.embedding;
+		const needsEmbed = !node.embedding?.length || node.description !== memories.find(m => m.name === node.name)?.description;
-		}
+		if (needsEmbed) {
-
+			const [e] = await this.llm.embedding(node.description);
-		if(isNew) memories.push(mem);
+			if (e) node.embedding = e.embedding;
-		else {
-			const idx = memories.findIndex(m => m.name === newMem.name);
-			if(idx >= 0) memories[idx] = mem;
 		}
 	}
 
-	/**
+	private async factAgent(conversation: string, memories: Memory[], options: LLMRequest): Promise<FactBucket[]> {
-	 * Find relevant memory documents for a query using description embeddings
+		const buckets: FactBucket[] = [];
-	 * @param {string} query The query to search against
+		const today = new Date().toISOString().split('T')[0];
-	 * @param {Memory[]} memories The user's memory documents
+
-	 * @param {number} limit Max number of results to return
+		await this.llm.ask(conversation, {
-	 * @returns {Promise<Memory[]>} The most relevant memory documents
+			model: options.model,
-	 */
+			temperature: 0.2,
-	async recollect(query: string, memories: Memory[], limit = 5): Promise<Memory[]> {
+			system: `You are a fact extractor. Analyze this conversation and extract facts worth remembering long-term.
-		const [e] = await this.llm.embedding(query);
+
-		return memories
+Rules:
-			.filter(m => m.embedding?.length)
+- ONLY extract facts the USER explicitly stated about themselves, their work, or their projects
-			.map(m => ({...m, score: this.llm.cosineSimilarity(m.embedding, e.embedding)}))
+- ONLY extract decisions that were MADE during this conversation
-			.toSorted((a: any, b: any) => b.score - a.score)
+- DO NOT extract anything the AI said, its capabilities, or meta-conversation about the AI
-			.slice(0, limit);
+- DO NOT extract greetings, pleasantries, or generic exchanges
+- If nothing worth remembering was said, do not call any tools
+
+**Organizational patterns:**
+- Journal entries use paths like: Journal/${today}
+- People use paths like: People/Name
+- Projects use paths like: Projects/Name
+- Personal info uses paths like: Personal/Goals, Personal/Tasks, etc.
+- General knowledge uses paths like: Biology/Topic, History/Topic, etc.
+
+Learn from existing nodes and follow the same pattern when extracting.
+
+Group facts by subject. For each group call \`extract_facts\` once with the FULL PATH.
+
+Known nodes (name: description):
+${this.listNodes(memories).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}`,
+			tools: [{
+				name: 'extract_facts',
+				description: 'Submit a group of related facts for a specific subject',
+				args: {
+					subject: {type: 'string', description: 'Full path for the subject (e.g., "Journal/2025-01-27", "People/Sarah", "Projects/Website")', required: true},
+					facts: {type: 'string', description: 'Comma-separated list of extracted facts', required: true},
+				},
+				fn: (args: any) => {
+					buckets.push({
+						subject: args.subject,
+						facts: args.facts.split(',').map((f: string) => f.trim()).filter(Boolean),
+					});
+					return 'Recorded';
+				}
+			}]
+		});
+
+		return buckets;
 	}
 
-	/**
+	private async organizingAgent(bucket: FactBucket, memories: Memory[], options: LLMRequest): Promise<Memory> {
-	 * Two-stage memory pipeline: classify facts from conversation history then surgically merge them into documents.
+		let candidates = this.listNodes(memories);
-	 * Bot 1 (classify) extracts and groups facts cheaply. Bot 2 (edit) runs per-document in parallel with full content access.
+		let attempts = 0;
-	 * @param {LLMMessage[]} history Full conversation history to digest
+		const maxAttempts = 3;
-	 * @param {Memory[]} memories The user's memory documents — mutated in place
+
-	 * @param {LLMRequest} options LLM options
+		while (attempts++ < maxAttempts) {
-	 */
+			let home = '', mode: string | null = null;
-	async memorize(history: LLMMessage[], memories: Memory[], options: LLMRequest): Promise<void> {
+
-		const conversation = history
+			const resp = await this.llm.ask(`Subject: ${bucket.subject}\n\nFacts:\n${bucket.facts.map(f => `- ${f}`).join('\n')}`, {
-			.filter(h => h.role === 'user' || h.role === 'assistant')
+				model: options.model,
-			.map(h => `[${h.role}]: ${h.content}`)
+				temperature: 0.1,
-			.join('\n\n');
+				system: `You are a knowledge organizer. Your job is to find the correct home for the supplied facts.
-		if(!conversation.trim()) return;
+
-		const pools = await this.extract(conversation, memories, options);
+1. Review the facts and the node list below. Pick the most likely match or decide if a new node is needed.
-		if(!pools.length) return;
+2. If you picked an existing node, use \`read\` to verify it's the right place.
-		await Promise.all(pools.map(pool => this.edit(pool, memories, options)));
+- After reading, call either \`confirm\` (correct node) or \`mismatched\` (wrong node).
+3. If none of the nodes match, call \`create\` to make a new node.
+
+**Organizational patterns:**
+- Journal entries: Journal/YYYY-MM-DD
+- People: People/Name
+- Projects: Projects/Name
+- Personal: Personal/Goals, Personal/Tasks, etc.
+- Knowledge: Biology/Topic, History/Topic, etc.
+
+Available nodes:
+${candidates.map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None — create a new node.'}`,
+				tools: [{
+					name: 'read',
+					description: 'Read a node file to verify it is the right home for these facts',
+					args: {name: {type: 'string', description: 'Exact node name (full path)', required: true}},
+					fn: ({name}) => {
+						const mem = memories.find(m => m.name === name);
+						if (!mem) return 'Node not found';
+						home = name;
+						return this.formatMemory(mem);
+					}
+				}, {
+					name: 'confirm',
+					description: 'Confirm this is the correct node for the facts',
+					args: {},
+					fn: () => {
+						mode = 'success';
+						resp.abort();
+					}
+				}, {
+					name: 'mismatched',
+					description: 'This is not the node you are looking for',
+					args: {},
+					fn: () => {
+						mode = 'failed';
+						resp.abort();
+					}
+				}, {
+					name: 'create',
+					description: 'No existing node fits — create a new one',
+					args: {
+						name: {type: 'string', description: 'Full path for the new node (e.g., "People/Sarah", "Journal/2025-01-27")', required: true}
+					},
+					fn: ({name}) => {
+						home = name;
+						mode = 'create';
+						resp.abort();
+					}
+				}]
+			});
+
+			if(mode === 'create') {
+				return this.createNode(home, memories);
+			} else if (mode === 'failed') {
+				candidates = candidates.filter(c => c.name !== home);
+				if(!candidates.length) return this.createNode(bucket.subject, memories);
+			} else if (mode === 'success') {
+				const existing = memories.find(m => m.name === home);
+				return existing || this.createNode(home, memories);
+			}
+		}
+		return this.createNode(bucket.subject, memories);
 	}
 }

src/open-ai.ts

@@ -8,7 +8,7 @@ import {convertSchema} from './tools.ts';
 export class OpenAi extends LLMProvider {
 	client!: openAI;
 
-	constructor(public readonly ai: Ai, public readonly host: string | null, public readonly token: string, public model: string, public llama?: boolean) {
+	constructor(public readonly ai: Ai, public readonly host: string | null, public readonly token: string, public model: string) {
 		super();
 		this.client = new openAI(clean({
 			baseURL: host,
@@ -96,13 +96,6 @@ export class OpenAi extends LLMProvider {
 
 			if(options.schema) {
 				const schema = convertSchema(options.schema);
-				if(this.llama) {
-					delete requestParams.tools;
-					requestParams.response_format = {
-						type: 'json_schema',
-						json_schema: {name: 'json', schema}
-					}
-				} else {
 				requestParams.response_format = {
 					type: 'json_schema',
 					json_schema: {
@@ -112,7 +105,6 @@ export class OpenAi extends LLMProvider {
 					}
 				};
 			}
-			}
 
 			let resp: any, isFirstMessage = true;
 			do {

src/tools.ts

@@ -100,16 +100,11 @@ export const CliTool: AiTool = {
 
 export const DateTimeTool: AiTool = {
 	name: 'get_datetime',
-	description: 'Get local date / time',
+	description: 'Get local/UTC date/time',
-	args: {},
+	args: {
-	fn: async () => new Date().toString()
+		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 DateTimeUTCTool: AiTool = {
-	name: 'get_datetime_utc',
-	description: 'Get current UTC date / time',
-	args: {},
-	fn: async () => new Date().toUTCString()
 }
 
 export const ExecTool: AiTool = {
@@ -168,7 +163,7 @@ export const JSTool: AiTool = {
 }
 
 export const PythonTool: AiTool = {
-	name: 'exec_javascript',
+	name: 'exec_python',
 	description: 'Execute commonjs javascript',
 	args: {
 		code: {type: 'string', description: 'CommonJS javascript', required: true}
@@ -306,93 +301,90 @@ export const WebSearchTool: AiTool = {
 	}
 }
 
-class WikipediaClient {
+export const WikipediaTool: AiTool = {
-	private async get(url: string): Promise<any> {
+	name: 'wikipedia_search',
+	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'}
+	},
+	fn: async (args: {query: string, mode: 'search' | 'summary' | 'full'}) => {
+		const UA = 'Mozilla/5.0 (Windows NT 10.0; Win64; x64)';
+
+		class WikipediaClient {
+			async get(url: string) {
 				const resp = await fetch(url, {headers: {'User-Agent': UA}});
 				return resp.json();
 			}
 
-	private api(params: Record<string, any>): Promise<any> {
+			api(params: 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 {
+			clean(text: string) {
-		return text.replace(/\n{3,}/g, '\n\n').replace(/ {2,}/g, ' ').replace(/\[\d+\]/g, '').trim();
+				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);
 				}
 
-	private truncate(text: string, max: number): string {
+				return text
-		if(text.length <= max) return text;
+					.replace(/^={4}\s*(.+?)\s*={4}$/gm, '#### $1')
-		const cut = text.slice(0, max);
+					.replace(/^={3}\s*(.+?)\s*={3}$/gm, '### $1')
-		const lastPara = cut.lastIndexOf('\n\n');
+					.replace(/^={2}\s*(.+?)\s*={2}$/gm, '## $1')
-		return lastPara > max * 0.7 ? cut.slice(0, lastPara) : cut;
+					.replace(/\n{3,}/g, '\n\n')
+					.replace(/ {2,}/g, ' ')
+					.replace(/\[\d+\]/g, '')
+					.trim();
 			}
 
-	private async searchTitles(query: string, limit = 6): Promise<any[]> {
+			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 || [];
 			}
 
-	private async fetchExtract(title: string, intro = false): Promise<string> {
+			async fetchExtract(title: string, introOnly = false) {
 				const params: any = {action: 'query', prop: 'extracts', titles: title, explaintext: 1, redirects: 1};
-		if(intro) params.exintro = 1;
+				if(introOnly) params.exintro = 1;
 				const data = await this.api(params);
-		const page = Object.values(data.query?.pages || {})[0] as any;
+				const page: any = Object.values(data.query?.pages || {})[0];
 				return this.clean(page?.extract || '');
 			}
 
-	private pageUrl(title: string): string {
+			pageUrl(title: string) {
 				return `https://en.wikipedia.org/wiki/${encodeURIComponent(title.replace(/ /g, '_'))}`;
 			}
 
-	private stripHtml(text: string): string {
+			stripHtml(text: string) {
 				return text.replace(/<[^>]+>/g, '');
 			}
 
-	async lookup(query: string, detail: 'intro' | 'full' = 'intro'): Promise<string> {
+			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 content = await this.fetchExtract(title, detail === 'intro');
+				const introOnly = detail !== 'full';
-		const text = this.truncate(content, detail === 'intro' ? 2000 : 8000);
+				const content = await this.fetchExtract(title, introOnly);
-		return `## ${title}\n🔗 ${url}\n\n${text}`;
+				return `## ${title}\n🔗 ${url}\n\n${content}`;
 			}
 
-	async search(query: string): Promise<string> {
+			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.truncate(this.stripHtml(r.snippet || ''), 150);
+					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');
 			}
-}
+		}
 
-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');
+		if(args.mode == 'search') return wiki.search(args.query);
-	}
+		return wiki.lookup(args.query, args.mode || 'summary');
-};
-
-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);
 	}
 };