Memory refinement

README.md

@@ -186,7 +186,7 @@ console.log(chunks);
 
 // Manually compile history into memories at end of conversation
 // Happens automatically when coverstaions are compressed
-await ai.language.updateMemory(history, memory);
+await ai.language.memorize(history, memory);
 
 // Summarize text
 const summary = await ai.language.summarize(longText, 200);

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "1.4.1",
+	"version": "1.4.5",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",

src/antrhopic.ts

@@ -24,49 +24,29 @@ export class Anthropic extends LLMProvider {
 		return client;
 	}
 
-	private toStandard(history: any[]): LLMMessage[] {
+	/** Convert standard history -> Anthropic wire format */
-		const timestamp = Date.now();
+	private toWire(history: LLMMessage[]): any[] {
-		const messages: LLMMessage[] = [];
+		const wire: any[] = [];
-		for(let h of history) {
+		for(const h of history) {
-			if(typeof h.content == 'string') {
+			if(h.role === 'tool') {
-				messages.push(<any>{timestamp, ...h});
+				wire.push(
-			} else {
-				const textContent = h.content?.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
-				if(textContent) messages.push({role: h.role, content: textContent, timestamp: timestamp, duration: h.duration, tps: h.tps});
-				h.content.forEach((c: any) => {
-					if(c.type == 'tool_use') {
-						messages.push({role: 'tool', id: c.id, name: c.name, args: c.input, timestamp: h.timestamp, content: undefined, duration: h.duration, tps: h.tps});
-					} 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;
-					}
-				});
-			}
-		}
-		return messages;
-	}
-
-	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,
 					{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}]}
+					{role: 'user', content: [{type: 'tool_result', tool_use_id: h.id, is_error: !!h.error, content: h.error || h.content || ''}]}
-				)
+				);
-				i++;
+			} else {
+				wire.push({role: h.role, content: h.content});
 			}
 		}
-		return history;
+		return wire;
 	}
 
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<string | any> {
 		const controller = new AbortController();
-		return Object.assign(new Promise<any>(async (res) => {
+		return Object.assign(new Promise<any>(async (res, rej) => {
-			let history = this.fromStandard([
+			if(!options.history) options.history = [];
-				...(options.history || []).filter(h => h.role !== 'system'),
+			const history = options.history;
-				{role: 'user', content: message, timestamp: Date.now()}
+			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,
@@ -80,54 +60,43 @@ export class Anthropic extends LLMProvider {
 						type: 'object',
 						properties: t.args ? objectMap(t.args, (key, value) => ({...value, required: undefined})) : {},
 						required: t.args ? Object.entries(t.args).filter(t => t[1].required).map(t => t[0]) : []
-					},
+					}
-					fn: undefined
 				})),
-				messages: history,
 				stream: !!options.stream,
 			};
 
-			// Add structured output support
 			if(options.schema) {
-				requestParams.output_config = {
+				requestParams.output_config = {format: {type: 'json_schema', schema: convertSchema(options.schema)}};
-					format: {
-						type: 'json_schema',
-						schema: convertSchema(options.schema)
-					}
-				};
 			}
 
-			let resp: any, terminal = false, duration = 0, tps = 0;
+			try {
+				let terminal = false;
 				do {
-				requestParams.messages = history.map(({timestamp, ...m}) => m);
+					requestParams.messages = this.toWire(history.filter(h => h.role !== 'system'));
+
 					const callStart = Date.now();
-				resp = await this.tokenPool.run(token => this.getClient(token).messages.create(requestParams)).catch(err => {
+					const resp: any = await this.tokenPool.run(token => this.getClient(token).messages.create(requestParams)).catch(err => {
-					err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
+						err.message += `\n\nMessages:\n${JSON.stringify(requestParams.messages, null, 2)}`;
 						throw err;
 					});
 
-				let usage: any;
+					let usage: any, content: any[] = [];
 					if(options.stream) {
-					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') {
+								if(chunk.content_block.type === 'text') content.push({type: 'text', text: ''});
-								resp.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.content_block.type === 'tool_use') {
-								resp.content.push({type: 'tool_use', id: chunk.content_block.id, name: chunk.content_block.name, input: <any>''});
-							}
 							} else if(chunk.type === 'content_block_delta') {
 								if(chunk.delta.type === 'text_delta') {
-								const text = chunk.delta.text;
+									content.at(-1).text += chunk.delta.text;
-								resp.content.at(-1).text += text;
+									options.stream({text: chunk.delta.text});
-								options.stream({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);
+								const last = content.at(-1);
-							if(last?.input != null) last.input = last.input ? JSONAttemptParse(last.input, {}) : {};
+								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') {
@@ -136,49 +105,52 @@ export class Anthropic extends LLMProvider {
 						}
 					} else {
 						usage = resp.usage;
+						content = resp.content;
 					}
-				duration = Date.now() - callStart;
+					const duration = Date.now() - callStart;
-				tps = usage?.output_tokens && duration > 0 ? usage.output_tokens / (duration / 1000) : 0;
+					const tps = usage?.output_tokens && duration > 0 ? usage.output_tokens / (duration / 1000) : 0;
 
-				const toolCalls = resp.content.filter((c: any) => c.type === 'tool_use');
+					const toolCalls = content.filter((c: any) => c.type === 'tool_use');
 					if(toolCalls.length && !controller.signal.aborted) {
-					history.push({role: 'assistant', content: resp.content, timestamp: Date.now(), duration, tps});
+						const text = content.filter((c: any) => c.type === 'text').map((c: any) => c.text).join('\n\n').trim();
-					const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
+						if(text) history.push({role: 'assistant', content: text, timestamp: Date.now(), duration, tps});
-						const tool = tools.find(findByProp('name', toolCall.name));
+
-						if(options.stream) options.stream({tool: toolCall.name});
+						const entries = toolCalls.map((tc: any) => {
-						if(!tool) return {tool_use_id: toolCall.id, is_error: true, content: 'Tool not found'};
+							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 toolStream = options.stream && ((chunk: any) => {
 									if(chunk.done) { terminal = true; return; }
 									options.stream!(chunk);
 								});
-							const result = await tool.fn(toolCall.input, toolStream, this.ai, toolCall.id);
+								const result = await tool.fn(entry.args, toolStream, this.ai, tc.id);
-							return {type: 'tool_result', tool_use_id: toolCall.id, content: typeof result == 'object' ? JSONSanitize(result) : result};
+								entry.content = typeof result === 'object' ? JSONSanitize(result) : result;
-						} catch (err: any) {
+							} 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, timestamp: Date.now()});
+					} else {
-					requestParams.messages = history;
+						terminal = true;
+						const text = content.filter((c: any) => c.type === 'text').map((c: any) => c.text).join('\n\n').trim();
+						if(text) history.push({role: 'assistant', content: text, timestamp: Date.now(), duration, tps});
 					}
-			} while (!terminal && !controller.signal.aborted && resp.content.some((c: any) => c.type === 'tool_use'));
+				} while(!terminal && !controller.signal.aborted);
 
-			if(!terminal) {
-				const textContent = resp.content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
-				history.push({role: 'assistant', content: textContent.trim(), timestamp: Date.now(), duration, tps});
-			}
-
-			history = this.toStandard(history);
-			if(options.history) options.history.splice(0, options.history.length, ...history);
 				if(options.stream) options.stream({done: true});
 
 				const turnStart = history.map(h => h.role).lastIndexOf('user');
-			const finalContent = history.slice(turnStart + 1).reduce((str, h) => {
+				const finalContent = history.slice(turnStart + 1).reduce((str, h) => h.role === 'assistant' ? str + (h.content || '') : str, '').trim();
-				if(h.role === 'assistant') return str + (h.content || '');
-				return str;
-			}, '').trim();
-
 				res(options.schema ? JSONAttemptParse(finalContent, finalContent) : finalContent);
+			} catch(err) {
+				rej(err);
+			}
 		}), {abort: () => controller.abort()});
 	}
 }

src/helpers.ts

@@ -7,10 +7,24 @@ export type MemoryNode = {
 	backlinks: string[];
 }
 
-export function buildMemoryGraph(memories: Memory[] | MemoryCache): MemoryNode[] {
+export function extractLinks(content: string): string[] {
+	if (!content) return [];
+	const matches = content.matchAll(/\[\[([^\]|]+)(?:\|[^\]]*)?\]\]/g);
+	return [...new Set([...matches].map(m => m[1].trim()))];
+}
+
+export function rebuildGraph(memories: Memory[] | MemoryCache): MemoryNode[] {
 	const mems = memories instanceof MemoryCache ? memories.memories : memories;
 	const nameSet = new Set(mems.map(m => m.name));
-	const ghosts = new Set<string>();
+
+	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,
@@ -19,6 +33,7 @@ export function buildMemoryGraph(memories: Memory[] | MemoryCache): MemoryNode[]
 		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);
@@ -31,30 +46,28 @@ export function buildMemoryGraph(memories: Memory[] | MemoryCache): MemoryNode[]
 			name,
 			missing: true,
 			links: [],
-			backlinks: nodes
+			backlinks: nodes.filter(n => n.links.includes(name)).map(n => n.name),
-				.filter(n => n.links.includes(name))
+		})),
-				.map(n => n.name),
-		}))
 	];
 }
 
-export function renderMemoryGraph(nodes) {
+export function renderMemoryGraph(nodes: MemoryNode[]): string {
 	if (!nodes.length) return 'No memories yet.';
 
-	const groups = new Map();
+	const groups = new Map<string, (MemoryNode & {label: string})[]>();
 	for (const node of nodes) {
 		const [prefix, ...rest] = node.name.split('/');
 		const group = rest.length ? prefix : 'Root';
 		const label = rest.length ? rest.join('/') : node.name;
 		if (!groups.has(group)) groups.set(group, []);
-		groups.get(group).push({...node, label});
+		groups.get(group)!.push({...node, label});
 	}
 
 	const ghostCount = nodes.filter(n => n.missing).length;
 	const lines = [`Memory Graph (${nodes.length} nodes, ${ghostCount} ghost${ghostCount === 1 ? '' : 's'})`, ''];
 
 	for (const group of [...groups.keys()].sort()) {
-		const items = groups.get(group).sort((a, b) => a.label.localeCompare(b.label));
+		const items = groups.get(group)!.sort((a, b) => a.label.localeCompare(b.label));
 		lines.push(`${group}/`);
 		items.forEach((n, i) => {
 			const last = i === items.length - 1;

src/kd-tree.ts

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

src/llm.ts

@@ -1,4 +1,4 @@
-import {snakeCase} from '@ztimson/utils';
+import {clean, snakeCase} from '@ztimson/utils';
 import {AbortablePromise, Ai} from './ai.ts';
 import {Anthropic} from './antrhopic.ts';
 import {OpenAi} from './open-ai.ts';
@@ -34,6 +34,10 @@ export type LLMMessage = {
 	content: string | any;
 	/** Timestamp */
 	timestamp?: number;
+	/** Response duration in ms */
+	duration?: number;
+	/** Tokens per second */
+	tps?: number;
 } | {
 	/** Tool call */
 	role: 'tool';
@@ -128,21 +132,23 @@ class LLM {
 			return {
 				name: toolName,
 				description: `${a.delegate ? 'Delegate to ' : ''}Subagent: ${a.description || a.name}`,
-				args: <any>(a.delegate ? {} : {
+				args: clean<any>({
-					context: {type: 'string', description: 'Summary of related messages, samples, files, etc...', required: true},
+					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';
 
-					// Opt-in only, self always excluded regardless of whitelist
 					const nested = (a.agents || [])
 						.map(name => allAgents.find(x => x.name === name))
 						.filter((x): x is Agent => !!x && x.name !== a.name);
 
-					const request = this.ask(a.delegate ? '' : `${args.instructions}${args.context ? `\n\n<context>${args.context}</context>` : ''}`, {
+					const q = a.delegate ? '' : `${args.instructions}${args.context ? `\n\n<context>${args.context}</context>` : ''}`;
-						system: `You are a specialized subagent. ${a.delegate ? 'Your output streams directly to the user for the remainder of this turn. You are mid conversation - dispense with greetings.' : 'You are wrapped in a tool call that will be analysis by an LLM - dispense with conversation'}
+
-As a subagent, focus on executing your task completely using available tools and returning only the final result - no commentary, questions, or dialogue.
+					const request = this.ask(q, {
+						system: `You are a specialized subagent being called from an orchestrator
+${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,
@@ -210,7 +216,7 @@ ${a.system}`,
 		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: `You have access to the following skill documents, whenever there is overlap between a question and a skill file, use \`skill_read\` to get instructions and background knowledge:\n${list}`,
 			tools: [{
 				name: 'skill_read',
 				description: 'Read the full content of a skill/knowledge document',
@@ -266,6 +272,7 @@ ${a.system}`,
 			let tools: AiTool[] = options.tools || this.ai.options.llm?.tools || [];
 			const prompts: string[] = [];
 			let history = options.history || [];
+			if(message) history.push({role: 'user', content: message, timestamp: Date.now()});
 
 			// MCP
 			const mcp = options.mcp || this.ai.options?.llm?.mcp;
@@ -309,18 +316,30 @@ ${a.system}`,
 							} else listed.push(r);
 						}
 
-						prompts.unshift(`You have access to the following memory files:
+						prompts.unshift(`You have a background memory process which has prefetched relevant information${mem.update ? ' and will create new memories from this conversation' : ''} for you
-${mems.map(m => `- ${m.name}: ${m.description}`).join('\n')}
+Assume it is perfect and never mention this process to anyone ever
+Always use your memories to craft a personalized response, they contain links / [[wiki links]] which you use navigate between them
+${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 unique queries before asking` : ''}
+
 ${preloaded.length ? `
-Relevant memories have been preloaded:
+Prefetched Memories (Most relevant first):
-${preloaded.map(r => `
+
-**${r.name}**
+${preloaded.map(r => `Memory: ${r.name}
-${r.description}
+Description: ${r.description}
+Linked: ${[r.links, ...r.backlinks].join(', ')}
+\`\`\`
 ${r.content}
-`).join('\n---\n')}
+\`\`\``).join('\n\n')}` : ''}
-` : ''}${listed.length ? `
+
-Also relevant but not preloaded (use \`memory_recall\`): ${listed.map(r => r.name).join(', ')}
+${mem.tool && listed.length ? listed.map(r => `Memory: ${r.name}
-` : ''}`.trim());
+Description: ${r.description}
+Linked: ${[r.links, ...r.backlinks].join(', ')}
+<!-- Truncated -->`).join('\n\n') : ''}
+
+${mem.tool ? `Full memory list:
+${mems.map(m => `- ${m.name}: ${m.description}`).join('\n')}` : ''}`.trim())
 					}
 					if(mem.tool) tools.push(this.memoryManager.tools.read(mem.memory));
 				}
@@ -334,7 +353,7 @@ Also relevant but not preloaded (use \`memory_recall\`): ${listed.map(r => r.nam
 			if(aborted) throw Object.assign(new Error('Aborted'), {name: 'AbortError'});
 
 			prompts.unshift(options.system || this.ai.options.llm?.system || '');
-			request = this.models[m].ask(message, {...options, tools, system: prompts.filter(Boolean).join('\n\n')});
+			request = this.models[m].ask('', {...options, tools, system: prompts.filter(Boolean).join('\n\n')});
 			let resp = await request;
 
 			// Capture meta (duration / tps)
@@ -364,14 +383,6 @@ Also relevant but not preloaded (use \`memory_recall\`): ${listed.map(r => r.nam
 		return Object.assign(promise, {abort});
 	}
 
-	/**
-	 * Digest full conversation history into memory documents.
-	 * Call on session end to persist the conversation.
-	 */
-	async updateMemory(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest = {}): Promise<Memory[]> {
-		return this.memoryManager.memorize(history, memories, {model: this.defaultModel, ...options});
-	}
-
 	/**
 	 * Compress chat history to reduce context size
 	 * @param {LLMMessage[]} history Chatlog that will be compressed
@@ -551,6 +562,14 @@ Also relevant but not preloaded (use \`memory_recall\`): ${listed.map(r => r.nam
 		};
 	}
 
+	/**
+	 * Digest full conversation history into memory documents.
+	 * Call on session end to persist the conversation.
+	 */
+	async memorize(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest = {}): Promise<Memory[]> {
+		return this.memoryManager.memorize(history, memories, {model: this.defaultModel, ...options});
+	}
+
 	/**
 	 * Create a summary of some text
 	 * @param {string} text Text to summarize

src/memory.ts

@@ -1,3 +1,4 @@
+import {MemoryNode, rebuildGraph} from './helpers.ts';
 import {LLMRequest, LLMMessage} from './llm.ts';
 import {AiTool} from './tools.ts';
 import {KDPoint, KDTree} from './kd-tree.ts';
@@ -12,76 +13,6 @@ const GENERIC_TEMPLATE = `# {{Title}}
 
 ## Related`;
 
-export class MemoryCache {
-	private tree: KDTree<MemoryRef>;
-	public memories: Memory[];
-
-	get length() { return this.memories.length; }
-
-	constructor(memories: Memory[]) {
-		this.memories = memories;
-		this.tree = this.buildTree();
-	}
-
-	private buildTree(): KDTree<MemoryRef> {
-		const embedded = this.memories.filter(m => m.embedding?.length);
-		if (!embedded.length) return new KDTree<MemoryRef>(0);
-
-		const dims = embedded[0].embedding.length;
-		const points: KDPoint<MemoryRef>[] = embedded.map(m => ({
-			vector: m.embedding,
-			payload: {name: m.name, description: m.description},
-		}));
-
-		return new KDTree<MemoryRef>(dims, 'cosine', points);
-	}
-
-	search(query: number[], limit: number): MemoryRef[] {
-		const results = this.tree.knn(query, limit);
-		return results.map(r => r.point.payload);
-	}
-
-	add(memory: Memory): void {
-		this.memories.push(memory);
-		this.rebuild();
-	}
-
-	update(memory: Memory): void {
-		const idx = this.memories.findIndex(m => m.name === memory.name);
-		if (idx !== -1) {
-			this.memories[idx] = memory;
-		} else {
-			this.memories.push(memory);
-		}
-		this.rebuild();
-	}
-
-	remove(name: string): void {
-		const idx = this.memories.findIndex(m => m.name === name);
-		if (idx !== -1) {
-			this.memories.splice(idx, 1);
-			this.rebuild();
-		}
-	}
-
-	rebuild(): void {
-		this.tree = this.buildTree();
-	}
-}
-
-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 type Memory = {
 	name: string;
 	description: string;
@@ -101,23 +32,6 @@ type FactBucket = {
 	facts: string[];
 }
 
-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[]): void {
-	for (const m of memories) m.links = extractLinks(m.content).filter(l => l !== m.name);
-	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 dedupeFacts(facts: string[]): string[] {
 	const seen = new Map<string, string>();
 	for (const f of facts) {
@@ -138,12 +52,132 @@ function cosineDistance(a: number[], b: number[]): number {
 	return denom === 0 ? 1 : 1 - dot / denom;
 }
 
-function getWeekMonday(date: Date = new Date()): string {
+function cosineSearch(query: number[], memories: Memory[], limit: number): MemoryRef[] {
-	const d = new Date(Date.UTC(date.getFullYear(), date.getMonth(), date.getDate()));
+	return memories
-	const day = d.getUTCDay();
+		.filter(m => m.embedding?.length)
-	const diff = day === 0 ? -6 : 1 - day;
+		.map(m => ({ref: {name: m.name, description: m.description}, distance: cosineDistance(query, m.embedding)}))
-	d.setUTCDate(d.getUTCDate() + diff);
+		.sort((a, b) => a.distance - b.distance)
-	return d.toISOString().slice(0, 10);
+		.slice(0, limit)
+		.map(s => s.ref);
+}
+
+export class MemoryCache {
+	private tree!: KDTree<MemoryRef>;
+	public memories: Memory[];
+	public nodes: MemoryNode[] = [];
+
+	get length() { return this.memories.length; }
+
+	constructor(memories: Memory[]) {
+		this.memories = memories;
+		this.rebuild();
+	}
+
+	private buildTree(): KDTree<MemoryRef> {
+		const embedded = this.memories.filter(m => m.embedding?.length);
+		if (!embedded.length) return new KDTree<MemoryRef>(0);
+
+		const dims = embedded[0].embedding.length;
+		const points: KDPoint<MemoryRef>[] = embedded.map(m => ({
+			vector: m.embedding,
+			payload: {name: m.name, description: m.description},
+		}));
+
+		return new KDTree<MemoryRef>(dims, 'cosine', points);
+	}
+
+	search(query: number[], limit: number): MemoryRef[] {
+		if (!this.tree || this.tree.dims === 0) return [];
+		return this.tree.knn(query, limit).map(r => r.point.payload);
+	}
+
+	add(memory: Memory): void {
+		this.memories.push(memory);
+		this.rebuild();
+	}
+
+	update(memory: Memory): void {
+		const idx = this.memories.findIndex(m => m.name === memory.name);
+		if (idx !== -1) this.memories[idx] = memory;
+		else this.memories.push(memory);
+		this.rebuild();
+	}
+
+	remove(name: string): void {
+		const idx = this.memories.findIndex(m => m.name === name);
+		if (idx !== -1) {
+			this.memories.splice(idx, 1);
+			this.rebuild();
+		}
+	}
+
+	rebuild(): void {
+		this.nodes = rebuildGraph(this.memories);
+		this.tree = this.buildTree();
+	}
+}
+
+class MemoryAccessor {
+	readonly list: Memory[];
+	private readonly cache: MemoryCache | null;
+
+	constructor(memories: Memory[] | MemoryCache) {
+		this.cache = memories instanceof MemoryCache ? memories : null;
+		this.list = this.cache ? this.cache.memories : <Memory[]>memories;
+	}
+
+	find(name: string): Memory | undefined {
+		return this.list.find(m => m.name === name);
+	}
+
+	commit(): MemoryNode[] {
+		if (this.cache) {
+			this.cache.rebuild();
+			return this.cache.nodes;
+		}
+		return rebuildGraph(this.list);
+	}
+
+	ghosts(): string[] {
+		const nodes = this.cache ? this.cache.nodes : rebuildGraph(this.list);
+		return nodes.filter(n => n.missing).map(n => n.name);
+	}
+
+	search(vector: number[], limit: number): MemoryRef[] {
+		return this.cache ? this.cache.search(vector, limit) : cosineSearch(vector, this.list, limit);
+	}
+
+	forget(name: string): boolean {
+		const idx = this.list.findIndex(m => m.name === name);
+		if (idx === -1) return false;
+		this.list.splice(idx, 1);
+		this.commit();
+		return true;
+	}
+
+	async backfillEmbeddings(llm: any): Promise<number> {
+		const missing = this.list.filter(m => !m.embedding?.length);
+		if (!missing.length) return 0;
+		await Promise.all(missing.map(async node => {
+			const [e] = await llm.embedding(node.content);
+			if (e) node.embedding = e.embedding;
+		}));
+		this.commit();
+		return missing.length;
+	}
+}
+
+export type MemoryOptions = {
+	/** Memory object */
+	memory: Memory[] | MemoryCache;
+	/** Inject N memories into the system prompt */
+	inject?: boolean;
+	/** expose recall tool to LLM */
+	tool?: boolean;
+	/** Update memory on compression */
+	update?: boolean;
+	/** Max context size of memories to inject to each call (removed immediately after use) */
+	maxTokens?: number;
 }
 
 export class MemoryManager {
@@ -156,21 +190,6 @@ export class MemoryManager {
 	}>();
 
 	tools = {
-		read: (memories: Memory[] | MemoryCache): AiTool => ({
-			name: 'memory_recall',
-			description: 'Read the full content of a memory document',
-			args: {
-				name: {type: 'string', description: 'Exact memory name', required: true},
-			},
-			fn: (args: any) => {
-				const mems = this.unwrap(memories);
-				const mem = mems.find(m => m.name === args.name);
-				if (!mem) return 'Document not found';
-				this.touch(mem.name);
-				return mem.content;
-			},
-		}),
-
 		forget: (memories: Memory[] | MemoryCache): AiTool => ({
 			name: 'memory_forget',
 			description: 'Permanently delete a memory document and clean up all references to it',
@@ -182,57 +201,50 @@ export class MemoryManager {
 				return result ? `Forgotten: ${args.name}` : `Not found: ${args.name}`;
 			},
 		}),
+
+		read: (memories: Memory[] | MemoryCache): AiTool => ({
+			name: 'memory_recall',
+			description: 'Read the full content of a memory document',
+			args: {
+				name: {type: 'string', description: 'Exact memory name', required: true},
+			},
+			fn: (args: any) => {
+				const mem = this.access(memories).find(args.name);
+				if (!mem) return 'Document not found';
+				this.touch(mem.name);
+				return mem.content;
+			},
+		}),
+
+		search: (memories: Memory[] | MemoryCache): AiTool => ({
+			name: 'memory_search',
+			description: 'Use embeddings to find the MOST relevant memories, even if NOT relevant',
+			args: {
+				query: {type: 'string', description: 'What to look for in the memories', required: true},
+				limit: {type: 'number', description: 'Number of memories to return', default: 1},
+			},
+			fn: async ({query, limit}) => {
+				const mem = await this.recollect(query, memories, limit)
+				return mem.map(m => `Memory: ${m.name}
+Description: ${m.description}
+Links: ${[...m.links, ...m.backlinks].join(', ')}
+\`\`\`
+${m.content}
+\`\`\``).join('\n\n');
+			},
+		}),
 	};
 
 	constructor(private llm: any) {}
 
 	static normalize(m?: Memory[] | MemoryCache | MemoryOptions) {
-		if(!m) return null;
+		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 unwrap(memories: Memory[] | MemoryCache): Memory[] {
+	private access(memories: Memory[] | MemoryCache): MemoryAccessor {
-		return memories instanceof MemoryCache ? memories.memories : memories;
+		return new MemoryAccessor(memories);
-	}
-
-	private sync(memories: Memory[] | MemoryCache): void {
-		if (memories instanceof MemoryCache) memories.rebuild();
-	}
-
-	private parseFrontmatter(content: string): {fm: Map<string, string>, body: string} {
-		const match = content.match(/^---\n([\s\S]*?)\n---\n?([\s\S]*)$/);
-		if (!match) return {fm: new Map(), body: content};
-		const fm = new Map<string, string>();
-		for (const line of match[1].split('\n')) {
-			const i = line.indexOf(':');
-			if (i === -1) continue;
-			fm.set(line.slice(0, i).trim(), line.slice(i + 1).trim());
-		}
-		return {fm, body: match[2]};
-	}
-
-	private writeFrontmatter(fm: Map<string, string>, body: string): string {
-		const lines = [...fm.entries()].map(([k, v]) => `${k}: ${v}`);
-		return `---\n${lines.join('\n')}\n---\n\n${body.trimStart()}`;
-	}
-
-	private stripHeader(content: string): string {
-		return content.replace(/^---[\s\S]*?\n---\n?/, '').trimStart();
-	}
-
-	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 ensureDoc(node: Memory): void {
-		if (node.content) return;
-		const title = node.name.split('/').pop() ?? node.name;
-		node.content = this.touchHeader(node, `# ${title}\n`);
 	}
 
 	private appendFacts(node: Memory, facts: string[]): void {
@@ -246,137 +258,78 @@ export class MemoryManager {
 		node.content = this.touchHeader(node, newBody);
 	}
 
-	decay() {
+	private ensureDoc(node: Memory): void {
-		for(const [name, ttl] of this.recentlyTouched) {
+		if (node.content) return;
-			if(ttl <= 1) this.recentlyTouched.delete(name);
+		const title = node.name.split('/').pop() ?? node.name;
-			else this.recentlyTouched.set(name, ttl - 1);
+		node.content = this.touchHeader(node, `# ${title}\n`);
-		}
 	}
 
-	touch(name: string, ttl = 2) {
+	private async factAgent(conversation: string, store: MemoryAccessor, options: LLMRequest, weekKey: string): Promise<FactBucket[]> {
-		this.recentlyTouched.set(name, ttl);
+		const ghosts = store.ghosts();
+
+		const response = await this.llm.ask(conversation, {
+			model: options.model,
+			temperature: 0.2,
+			system: `You are a fact extractor to build obsidian knowledge vaults.
+Analyze this conversation and extract facts worth remembering long-term.
+
+Rules:
+- Always extract facts that the user explicitly told you to remember
+- ONLY extract current facts the USER explicitly stated about themselves, their work, projects or decisions that were MADE during this conversation
+- DO NOT extract greetings, pleasantries, or generic exchanges
+- DO NOT extract deltas or changes in facts; ONLY the end fact
+- DO NOT extract anything the AI/assistant itself said
+- If nothing worth remembering was said, return an empty buckets array
+
+When extracting facts, you MUST also decide the exact destination path:
+- Reuse node names (including ghost) as much as possible IF the facts belongs there
+- All information primarily about the user should go under "People/User"
+- When required, create a new path following collection/subject format (e.g., People/Sarah, Projects/Oxide) — you are not limited to any fixed list of collections, use whatever fits
+- For journal entries, use "Journal"
+
+Available nodes:
+- Journal
+${this.listNodes(store.list).filter(n => !n.name.includes('Journal')).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}
+${ghosts.length ? `${ghosts.map(g => `- ${g}: (Ghost)`).join('\n')}` : ''}`,
+			schema: {
+				buckets: {type: 'array', description: 'Groups of facts to remember, each assigned to a different node. Return an empty array if there is nothing worth storing in an obsidian vault', items: {
+					type: 'object', items: {
+							subject: {type: 'string', description: 'Exact existing node name OR new path (e.g. "People/Sarah", "Projects/Oxide"), or "Journal"', required: true},
+							facts: {
+								type: 'array',
+								description: 'Facts to store at this destination',
+								items: {type: 'string', description: 'A single fact'},
+							},
+						},
+					},
+				},
+			},
+		});
+
+		const buckets = new Map<string, string[]>();
+		for(const bucket of response.buckets ?? []) {
+			const subject = bucket.subject.trim().toLowerCase() === 'journal'
+				? `Journal/${weekKey}` : bucket.subject.trim();
+			const facts = buckets.get(subject) ?? [];
+			facts.push(...dedupeFacts(bucket.facts));
+			buckets.set(subject, facts);
 		}
 
-	getTouched(): string[] {
+		return buckets.entries().toArray().map(([subject, facts]) => ({subject, facts}));
-		return [...this.recentlyTouched.keys()];
 	}
 
-	forget(name: string, memories: Memory[] | MemoryCache): boolean {
+	private getWeekMonday(date: Date = new Date()): string {
-		const mem = this.unwrap(memories);
+		const d = new Date(Date.UTC(date.getFullYear(), date.getMonth(), date.getDate()));
-		const idx = mem.findIndex(m => m.name === name);
+		const day = d.getUTCDay();
-		if (idx === -1) return false;
+		const diff = day === 0 ? -6 : 1 - day;
-
+		d.setUTCDate(d.getUTCDate() + diff);
-		mem.splice(idx, 1);
+		return d.toISOString().slice(0, 10);
-		rebuildGraph(mem);
-		this.sync(memories);
-		return true;
-	}
-
-	async recollect(query: string, memories: Memory[] | MemoryCache, limit = 5, graphDepth = 1): Promise<Memory[]> {
-		const mem = this.unwrap(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);
-	}
-
-	private cosineSearch(query: number[], memories: Memory[], limit: number): MemoryRef[] {
-		const scored = memories
-			.filter(m => m.embedding?.length)
-			.map(m => ({
-				ref: {name: m.name, description: m.description},
-				distance: cosineDistance(query, m.embedding),
-			}))
-			.sort((a, b) => a.distance - b.distance)
-			.slice(0, limit);
-		return scored.map(s => s.ref);
 	}
 
 	private listNodes(memories: Memory[]): MemoryRef[] {
 		return memories.map(m => ({name: m.name, description: m.description}));
 	}
 
-	async memorize(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest): Promise<Memory[]> {
-		const conversation = history
-			.filter(h => h.role === 'user' || h.role === 'assistant')
-			.map(h => `[${h.role}]: ${h.content}`).join('\n\n').trim();
-		if (!conversation) return [];
-
-		const uid = `${Date.now()}_${Math.random().toString(36).slice(2)}`;
-		// NOTE: adjust field names below (id/tool_call_id/name) to match your LLMMessage/tool-call schema.
-		const pending = {role: 'tool', name: 'memory_process', id: uid, content: 'Processing…'} as unknown as LLMMessage;
-		history.push(pending);
-
-		const mem = this.unwrap(memories);
-		const buckets = await this.factAgent(conversation, mem, options, getWeekMonday());
-		const touched: Memory[] = [];
-
-		for (const {subject, facts} of buckets) {
-			let node = mem.find(m => m.name === subject);
-			if (!node) {
-				node = {name: subject, description: '', content: '', embedding: [], links: [], backlinks: []};
-				mem.push(node);
-			}
-			this.appendFacts(node, facts);
-			const [e] = await this.llm.embedding(node.content);
-			if (e) node.embedding = e.embedding;
-			this.touch(node.name);
-			touched.push(node);
-		}
-
-		if (touched.length) {
-			rebuildGraph(mem);
-			this.sync(memories);
-			(pending as any).content = `Saved to ${touched.map(n => `[[${n.name}]]`).join(', ')}`;
-			for (const node of touched) this.reconcile(node, memories, options).catch(() => {});
-		} else {
-			(pending as any).content = 'Nothing worth remembering.';
-		}
-
-		return touched;
-	}
-
-	/** Manual/cron entry point. scope 'touched' only reconciles docs with a pending Facts inbox. */
-	async reconcileVault(memories: Memory[] | MemoryCache, options: LLMRequest, scope: 'touched' | 'all' = 'touched'): Promise<void> {
-		const mem = this.unwrap(memories);
-		const targets = scope === 'all' ? mem : mem.filter(m => m.content.includes(FACTS_HEADING));
-		await Promise.all(targets.map(node => this.reconcile(node, memories, options)));
-		this.sync(memories);
-	}
-
-	/**
-	 * Coalescing queue: if a doc is already reconciling, mark it dirty and abort the in-flight
-	 * request. The loop below always re-reads node.content fresh, so nothing is ever dropped.
-	 */
 	private reconcile(node: Memory, memories: Memory[] | MemoryCache, options: LLMRequest): Promise<void> {
 		const key = node.name;
 		const existing = this.queues.get(key);
@@ -388,21 +341,20 @@ export class MemoryManager {
 
 		const entry = {dirty: false, request: null, task: Promise.resolve()};
 		this.queues.set(key, entry);
-		const mem = this.unwrap(memories);
+		const store = this.access(memories);
 		entry.task = (async () => {
 			do {
 				entry.dirty = false;
-				await this.reconcileDoc(node, mem, options, entry);
+				await this.docAgent(node, store.list, options, entry);
 			} while (entry.dirty);
 		})().finally(() => {
 			this.queues.delete(key);
-			rebuildGraph(mem);
+			store.commit();
-			this.sync(memories);
 		});
 		return entry.task;
 	}
 
-	private async reconcileDoc(node: Memory, memories: Memory[], options: LLMRequest, entry: {request: {abort?: () => void} | null}): Promise<void> {
+	private async docAgent(node: Memory, memories: Memory[], options: LLMRequest, entry: {request: {abort?: () => void} | null}): Promise<void> {
 		const currentBody = this.stripHeader(node.content);
 		let update;
 		try {
@@ -456,47 +408,128 @@ ${currentBody}
 		if (e) node.embedding = e.embedding;
 	}
 
-	private async factAgent(conversation: string, memories: Memory[], options: LLMRequest, weekKey: string): Promise<FactBucket[]> {
+	private parseFrontmatter(content: string): {fm: Map<string, string>, body: string} {
-		const buckets = new Map<string, string[]>();
+		const match = content.match(/^---\n([\s\S]*?)\n---\n?([\s\S]*)$/);
-		await this.llm.ask(conversation, {
+		if (!match) return {fm: new Map(), body: content};
-			model: options.model,
+		const fm = new Map<string, string>();
-			temperature: 0.2,
+		for (const line of match[1].split('\n')) {
-			system: `You are a fact extractor. Analyze this conversation and extract facts worth remembering long-term.
+			const i = line.indexOf(':');
+			if (i === -1) continue;
+			fm.set(line.slice(0, i).trim(), line.slice(i + 1).trim());
+		}
+		return {fm, body: match[2]};
+	}
 
-Rules:
+	private stripHeader(content: string): string {
-- ONLY extract current facts the USER explicitly stated about themselves, their work, or their projects
+		return content.replace(/^---[\s\S]*?\n---\n?/, '').trimStart();
-- ONLY extract decisions that were MADE during this conversation
+	}
-- DO NOT extract anything the AI said, its capabilities, or meta-conversation about the AI
-- DO NOT extract greetings, pleasantries, or generic exchanges
-- DO NOT extract deltas or changes in facts; ONLY the end fact
-- If nothing worth remembering was said, do not call any tools
 
-When extracting facts, you MUST also decide the exact destination path:
+	private touchHeader(node: Memory, body: string): string {
-- Use an existing node name if the facts clearly belong there
+		const {fm} = this.parseFrontmatter(node.content);
-- All information primarily about the user should go under "People/User"
+		fm.set('name', node.name);
-- When required, create a new path following collection/subject format (e.g., People/Sarah, Projects/Oxide) — you are not limited to any fixed list of collections, use whatever fits
+		fm.set('description', node.description || '');
-- For journal entries, use "Journal"
+		fm.set('modified', new Date().toISOString());
+		return this.writeFrontmatter(fm, body);
+	}
 
-Available nodes:
+	private writeFrontmatter(fm: Map<string, string>, body: string): string {
-- Journal
+		const lines = [...fm.entries()].map(([k, v]) => `${k}: ${v}`);
-${this.listNodes(memories).filter(n => !n.name.includes('Journal')).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}`,
+		return `---\n${lines.join('\n')}\n---\n\n${body.trimStart()}`;
-			tools: [{
+	}
-				name: 'facts_extract',
+
-				description: 'Submit facts with their destination',
+	decay() {
-				args: {
+		for (const [name, ttl] of this.recentlyTouched) {
-					destination: {type: 'string', description: 'Exact existing node name OR new path (e.g. "People/Sarah", "Projects/Oxide")', required: true},
+			if (ttl <= 1) this.recentlyTouched.delete(name);
-					facts: {type: 'string', description: 'Comma-separated facts', required: true},
+			else this.recentlyTouched.set(name, ttl - 1);
-				},
+		}
-				fn: (args: any) => {
+	}
-					const subject = args.destination.trim().toLowerCase() === 'journal'
+
-						? `Journal/${weekKey}` : args.destination.trim();
+	touch(name: string, ttl = 2) {
-					const facts = buckets.get(subject) ?? [];
+		this.recentlyTouched.set(name, ttl);
-					facts.push(...dedupeFacts(String(args.facts).split(',')));
+	}
-					buckets.set(subject, facts);
+
-					return 'Recorded';
+	forget(name: string, memories: Memory[] | MemoryCache): boolean {
-				},
+		return this.access(memories).forget(name);
-			}],
+	}
-		});
+
-		return buckets.entries().toArray().map(([subject, facts]) => ({subject, facts}));
+	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);
+		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;
+			}
+		}
+
+		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}]: ${h.content}`).join('\n\n').trim();
+		if (!conversation) return [];
+
+		const uid = `${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 = await this.factAgent(conversation, store, options, this.getWeekMonday());
+		const touched: Memory[] = [];
+
+		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.appendFacts(node, facts);
+			const [e] = await this.llm.embedding(node.content);
+			if (e) node.embedding = e.embedding;
+			this.touch(node.name);
+			touched.push(node);
+		}
+
+		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 reconcileVault(memories: Memory[] | MemoryCache, options: LLMRequest, scope: 'touched' | 'all' = 'touched'): Promise<void> {
+		const store = this.access(memories);
+		const targets = scope === 'all' ? store.list : store.list.filter(m => m.content.includes(FACTS_HEADING));
+		await Promise.all(targets.map(node => this.reconcile(node, memories, options)));
+		store.commit();
 	}
 }

src/open-ai.ts

@@ -25,73 +25,38 @@ export class OpenAi extends LLMProvider {
 		return client;
 	}
 
-	private toStandard(history: any[]): LLMMessage[] {
+	/** Convert standard history -> OpenAI wire format */
-		for(let i = 0; i < history.length; i++) {
+	private toWire(history: LLMMessage[], system?: string): any[] {
-			const h = history[i];
+		const wire: any[] = [];
-			if(h.role === 'assistant' && h.tool_calls) {
+		if(system) wire.push({role: 'system', content: system});
-				const items: any[] = [];
+		for(const h of history) {
-				if(h.content) items.push({role: 'assistant', content: h.content, timestamp: h.timestamp, duration: h.duration, tps: h.tps});
-				items.push(...h.tool_calls.map((tc: any) => ({
-					role: 'tool',
-					id: tc.id,
-					name: tc.function.name,
-					args: JSONAttemptParse(tc.function.arguments, {}),
-					timestamp: h.timestamp,
-					duration: h.duration,
-					tps: h.tps
-				})));
-				history.splice(i, 1, ...items);
-				i += items.length - 1;
-			} else if(h.role === 'tool') {
-				const record = history.find(h2 => h.tool_call_id == h2.id);
-				if(record) {
-					if(h.content?.includes('"error":')) record.error = h.content;
-					else record.content = h.content || '';
-				}
-				history.splice(i, 1);
-				i--;
-			}
-			if(!history[i]?.timestamp) history[i].timestamp = Date.now();
-		}
-		return history;
-	}
-
-	private fromStandard(history: LLMMessage[]): any[] {
-		return history.reduce((result, h) => {
 			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) } }],
+					tool_calls: [{id: h.id, type: 'function', function: {name: h.name, arguments: JSON.stringify(h.args)}}],
-					refusal: null,
-					annotations: [],
-					timestamp: h.timestamp,
 				}, {
 					role: 'tool',
 					tool_call_id: h.id,
-					content: h.error || h.content,
+					content: h.error || h.content || '',
-					timestamp: h.timestamp,
 				});
 			} else {
-				result.push(h);
+				wire.push({role: h.role, content: h.content});
 			}
-			return result;
+		}
-		}, [] as any[]);
+		return wire;
 	}
 
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<string | any> {
 		const controller = new AbortController();
 		return Object.assign(new Promise<any>(async (res, rej) => {
-			const base = (options.history || []).filter(h => h.role !== 'system');
+			if(!options.history) options.history = [];
-			let history = this.fromStandard([
+			const history = options.history;
-				...(options.system ? [{role: <any>'system', content: options.system, timestamp: Date.now()}] : []),
+			if(message) history.push({role: 'user', content: message, timestamp: Date.now()});
-				...base,
+
-				{role: 'user', content: message, timestamp: Date.now()}
-			]);
 			const tools = options.tools || this.ai.options.llm?.tools || [];
 			const requestParams: any = {
 				model: options.model || this.model,
-				messages: history,
 				stream: !!options.stream,
 				max_completion_tokens: options.max_tokens || this.ai.options.llm?.max_tokens || undefined,
 				temperature: options.temperature || this.ai.options.llm?.temperature || undefined,
@@ -111,56 +76,42 @@ export class OpenAi extends LLMProvider {
 
 			if(options.schema) {
 				const schema = convertSchema(options.schema);
-				requestParams.response_format = {
+				requestParams.response_format = {type: 'json_schema', json_schema: {name: 'response', strict: true, schema}};
-					type: 'json_schema',
-					json_schema: {
-						name: 'response',
-						strict: true,
-						schema
 			}
-				};
-			}
-
 			if(options.stream) requestParams.stream_options = {include_usage: true};
-			let resp: any, terminal = false, duration = 0, tps = 0;
+
+			try {
+				let terminal = false;
 				do {
-				requestParams.messages = history.map(({timestamp, ...m}) => m);
+					requestParams.messages = this.toWire(history.filter(h => h.role !== 'system'), options.system);
+
 					const callStart = Date.now();
-				resp = await this.tokenPool.run(token => this.getClient(token).chat.completions.create(requestParams)).catch(err => {
+					const resp: any = await this.tokenPool.run(token => this.getClient(token).chat.completions.create(requestParams)).catch(err => {
-					err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
+						err.message += `\n\nMessages:\n${JSON.stringify(requestParams.messages, null, 2)}`;
 						throw err;
 					});
 
-				let usage: any;
+					let usage: any, msg: any = {content: '', tool_calls: []};
 					if(options.stream) {
-					resp.choices = [{message: {role: 'assistant', content: '', tool_calls: [], timestamp: Date.now()}}];
 						for await (const chunk of resp) {
 							if(controller.signal.aborted) break;
 							if(chunk.usage) usage = chunk.usage;
 							if(chunk.choices[0]?.delta?.content) {
-							resp.choices[0].message.content += 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) {
 								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?.name) existing.function.name = deltaTC.function.name;
-									if(deltaTC.function) {
+										if(deltaTC.function?.arguments) existing.function.arguments += deltaTC.function.arguments;
-										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;
-									}
 									} 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 || ''
-										}
 										});
 									}
 								}
@@ -168,51 +119,51 @@ export class OpenAi extends LLMProvider {
 						}
 					} else {
 						usage = resp.usage;
+						msg = resp.choices[0].message;
 					}
-				duration = Date.now() - callStart;
+					const duration = Date.now() - callStart;
-				tps = usage?.completion_tokens && duration > 0 ? usage.completion_tokens / (duration / 1000) : 0;
+					const tps = usage?.completion_tokens && duration > 0 ? usage.completion_tokens / (duration / 1000) : 0;
 
-				if(resp.error) throw new Error(resp.error);
+					const toolCalls = msg.tool_calls || [];
-				const toolCalls = resp.choices[0].message.tool_calls || [];
 					if(toolCalls.length && !controller.signal.aborted) {
-					history.push({...resp.choices[0].message, duration, tps});
+						if(msg.content?.trim()) history.push({role: 'assistant', content: msg.content.trim(), timestamp: Date.now(), duration, tps});
-					const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
+
-						const tool = tools?.find(findByProp('name', toolCall.function.name));
+						const entries = toolCalls.map((tc: any) => {
-						if(options.stream) options.stream({tool: toolCall.function.name});
+							const entry: any = {role: 'tool', id: tc.id, name: tc.function.name, args: JSONAttemptParse(tc.function.arguments, {}), content: undefined, timestamp: Date.now()};
-						if(!tool) return {role: 'tool', tool_call_id: toolCall.id, content: '{"error": "Tool not found"}', 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 toolStream = options.stream && ((chunk: any) => {
 									if(chunk.done) { terminal = true; return; }
 									options.stream!(chunk);
 								});
-							const result = await tool.fn(args, toolStream, this.ai, toolCall.id);
+								const result = await tool.fn(entry.args, toolStream, this.ai, tc.id);
-							return {role: 'tool', tool_call_id: toolCall.id, content: typeof result == 'object' ? JSONSanitize(result) : result, timestamp: Date.now()};
+								entry.content = typeof result === 'object' ? JSONSanitize(result) : result;
-						} catch (err: any) {
+							} catch(err: any) {
-							return {role: 'tool', tool_call_id: toolCall.id, content: JSONSanitize({error: err?.message || err?.toString() || 'Unknown'}), timestamp: Date.now()};
+								entry.error = err?.message || err?.toString() || 'Unknown';
 							}
 						}));
-					history.push(...results);
+					} else {
-					requestParams.messages = history;
+						terminal = true;
+						const text = (msg.content || '').trim();
+						if(text) history.push({role: 'assistant', content: text, timestamp: Date.now(), duration, tps});
 					}
-			} while (!terminal && !controller.signal.aborted && resp.choices?.[0]?.message?.tool_calls?.length);
+				} while(!terminal && !controller.signal.aborted);
 
-			if(!terminal) {
-				const textContent = resp.choices[0].message.content || '';
-				history.push({role: 'assistant', content: textContent.trim(), timestamp: Date.now(), duration, tps});
-			}
-
-			history = this.toStandard(history);
-			if(options.history) options.history.splice(0, options.history.length, ...history.filter(h => h.role !== 'system'));
 				if(options.stream) options.stream({done: true});
 
 				const turnStart = history.map(h => h.role).lastIndexOf('user');
-			const finalContent = history.slice(turnStart + 1).reduce((str, h) => {
+				const finalContent = history.slice(turnStart + 1).reduce((str, h) => h.role === 'assistant' ? str + (h.content || '') : str, '').trim();
-				if(h.role === 'assistant') return str + (h.content || '');
-				return str;
-			}, '').trim();
-
 				res(options.schema ? JSONAttemptParse(finalContent, finalContent) : finalContent);
+			} catch(err) {
+				rej(err);
+			}
 		}), {abort: () => controller.abort()});
 	}
 }

tests/llm.spec.ts

@@ -1,167 +0,0 @@
-
-import {describe, it, expect, vi, beforeEach} from 'vitest';
-import LLM from '../src/llm';
-
-const {FakeProvider, providerLog} = vi.hoisted(() => {
-	const providerLog: any[] = [];
-	class FakeProvider {
-		model: string;
-		constructor(...args: any[]) { this.model = args[args.length - 1]; }
-		ask(message: string, opts: any) {
-			let aborted = false;
-			const p = (async () => {
-				const script = (globalThis as any).__scripts?.[this.model];
-				const plan = script ? script(message, opts) : {text: ''};
-				providerLog.push({model: this.model, message, system: opts.system, tools: (opts.tools || []).map((t: any) => t.name)});
-				for (const c of plan.calls || []) {
-					if (aborted) break;
-					const tool = (opts.tools || []).find((t: any) => t.name === c.tool);
-					const id = c.id || `${c.tool}_${Math.random()}`;
-					const content = await tool.fn(c.args, opts.stream, null, id);
-					opts.history.push({role: 'tool', id, name: c.tool, args: c.args, content, timestamp: Date.now()});
-				}
-				const text = plan.text ?? '';
-				if (opts.stream && text) opts.stream({text, done: true});
-				opts.history.push({role: 'assistant', content: text, timestamp: Date.now(), duration: 10, tps: 5});
-				return text;
-			})();
-			return Object.assign(p, {abort: () => { aborted = true; }});
-		}
-	}
-	return {FakeProvider, providerLog};
-});
-
-vi.mock('../src/antrhopic.ts', () => ({Anthropic: FakeProvider}));
-vi.mock('../src/open-ai.ts', () => ({OpenAi: FakeProvider}));
-
-function makeAi(models: any) {
-	return {options: {llm: {models}}} as any;
-}
-
-beforeEach(() => {
-	providerLog.length = 0;
-	(globalThis as any).__scripts = {};
-});
-
-describe('LLM cross-provider interchangeability', () => {
-	it('runs identical tool calls the same way on an anthropic-backed model and an openai-backed model', async () => {
-		const ai = makeAi({
-			claude: {proto: 'anthropic', token: 'x'},
-			gpt: {proto: 'openai', token: 'y', host: 'http://local'},
-		});
-		const llm = new LLM(ai);
-		const calc = {
-			name: 'calc_add',
-			description: 'Add two numbers',
-			args: {a: {type: 'number', required: true}, b: {type: 'number', required: true}},
-			fn: (args: any) => String(args.a + args.b),
-		};
-
-		(globalThis as any).__scripts.claude = () => ({calls: [{tool: 'calc_add', args: {a: 2, b: 3}}], text: 'Result: 5'});
-		(globalThis as any).__scripts.gpt = () => ({calls: [{tool: 'calc_add', args: {a: 2, b: 3}}], text: 'Result: 5'});
-
-		const historyA: any[] = [], historyB: any[] = [];
-		const respA = await llm.ask('add 2 and 3', {model: 'claude', tools: [calc], history: historyA});
-		const respB = await llm.ask('add 2 and 3', {model: 'gpt', tools: [calc], history: historyB});
-
-		expect(respA).toBe('Result: 5');
-		expect(respB).toBe('Result: 5');
-		expect(providerLog.find(l => l.model === 'claude')!.tools).toContain('calc_add');
-		expect(providerLog.find(l => l.model === 'gpt')!.tools).toContain('calc_add');
-
-		// tool timing gets recomputed from real execution regardless of proto
-		for (const h of [historyA.find(h => h.name === 'calc_add'), historyB.find(h => h.name === 'calc_add')]) {
-			expect(h.content).toBe('5');
-			expect(typeof h.duration).toBe('number');
-			expect(typeof h.tps).toBe('number');
-		}
-	});
-
-	it('lets the same shared history flow across model + proto swaps with different system prompts', async () => {
-		const ai = makeAi({
-			claude: {proto: 'anthropic', token: 'x'},
-			gpt: {proto: 'openai', token: 'y', host: 'http://local'},
-		});
-		const llm = new LLM(ai);
-		const history: any[] = [];
-
-		(globalThis as any).__scripts.claude = () => ({text: 'Hi from claude'});
-		(globalThis as any).__scripts.gpt = () => ({text: 'Hi from gpt'});
-
-		const r1 = await llm.ask('hello', {model: 'claude', system: 'You are terse.', history});
-		const r2 = await llm.ask('follow up', {model: 'gpt', system: 'You are verbose.', history});
-
-		expect(r1).toBe('Hi from claude');
-		expect(r2).toBe('Hi from gpt');
-		expect(history.filter(h => h.role === 'assistant').map(h => h.content)).toEqual(['Hi from claude', 'Hi from gpt']);
-		expect(providerLog[0].system).toContain('You are terse.');
-		expect(providerLog[1].system).toContain('You are verbose.');
-	});
-
-	it('exposes MCP tools the same way no matter which proto backs the model', async () => {
-		const ai = makeAi({claude: {proto: 'anthropic', token: 'x'}, gpt: {proto: 'openai', token: 'y', host: 'http://local'}});
-		const llm = new LLM(ai);
-		const mcp = [{name: 'weather', host: 'http://mcp.local'}];
-
-		global.fetch = vi.fn(async (url: string, opts?: any) => {
-			if (url.endsWith('/tools')) {
-				return {json: async () => ({tools: [{name: 'lookup', description: 'Look up weather', inputSchema: {properties: {city: {type: 'string'}}, required: ['city']}}]})} as any;
-			}
-			const body = JSON.parse(opts.body);
-			return {json: async () => ({content: [{text: `Sunny in ${body.arguments.city}`}]})} as any;
-		}) as any;
-
-		for (const model of ['claude', 'gpt']) {
-			(globalThis as any).__scripts[model] = () => ({calls: [{tool: 'weather_lookup', args: {city: 'Rome'}}], text: 'done'});
-			const history: any[] = [];
-			await llm.ask('weather?', {model, mcp, history});
-			expect(history.find(h => h.name === 'weather_lookup')?.content).toBe('Sunny in Rome');
-		}
-	});
-
-	it('exposes and resolves skill documents identically across protos', async () => {
-		const ai = makeAi({claude: {proto: 'anthropic', token: 'x'}, gpt: {proto: 'openai', token: 'y', host: 'http://local'}});
-		const llm = new LLM(ai);
-		const skills = [{name: 'Onboarding', description: 'How to onboard a user', content: 'Step 1...'}];
-
-		for (const model of ['claude', 'gpt']) {
-			(globalThis as any).__scripts[model] = () => ({calls: [{tool: 'skill_read', args: {name: 'Onboarding'}}], text: 'done'});
-			const history: any[] = [];
-			await llm.ask('onboard me', {model, skills, history});
-			expect(history.find(h => h.name === 'skill_read')?.content).toContain('Step 1...');
-		}
-	});
-
-	it('delegate agent mutates the shared history directly and backfills the orchestrator response, across protos', async () => {
-		const ai = makeAi({claude: {proto: 'anthropic', token: 'x'}, gpt: {proto: 'openai', token: 'y', host: 'http://local'}});
-		const llm = new LLM(ai);
-		const history: any[] = [{role: 'user', content: 'research quantum computing'}];
-		const researcher = {name: 'researcher', system: 'You research topics.', delegate: true, model: 'gpt'};
-
-		(globalThis as any).__scripts.claude = () => ({calls: [{tool: 'agent_researcher', args: {}}], text: ''});
-		(globalThis as any).__scripts.gpt = () => ({text: 'Quantum computers use qubits.'});
-
-		const resp = await llm.ask('go', {model: 'claude', agents: [researcher], history});
-
-		expect(resp).toBe('Quantum computers use qubits.');
-		expect(history.some(h => h.role === 'assistant' && h.content === 'Quantum computers use qubits.')).toBe(true);
-		expect(history.find(h => h.name === 'agent_researcher')?.content).toBe('');
-	});
-
-	it('regular (non-delegate) subagent keeps its own isolated history separate from the parent, across protos', async () => {
-		const ai = makeAi({claude: {proto: 'anthropic', token: 'x'}, gpt: {proto: 'openai', token: 'y', host: 'http://local'}});
-		const llm = new LLM(ai);
-		const history: any[] = [];
-		const summarizer = {name: 'summarizer', system: 'You summarize text.', model: 'gpt'};
-
-		(globalThis as any).__scripts.claude = () => ({calls: [{tool: 'subagent_summarizer', args: {context: 'a long article', instructions: 'summarize it'}}], text: 'Summary: short version'});
-		(globalThis as any).__scripts.gpt = () => ({text: 'short version'});
-
-		const resp = await llm.ask('summarize this', {model: 'claude', agents: [summarizer], history});
-
-		expect(resp).toBe('Summary: short version');
-		expect(history.find(h => h.name === 'subagent_summarizer')?.content).toBe('short version');
-		// isolated history - subagent's own assistant turn never leaks into the parent
-		expect(history.some(h => h.role === 'assistant' && h.content === 'short version')).toBe(false);
-	});
-});

tests/memory.spec.ts

@@ -1,256 +0,0 @@
-import {describe, it, expect, vi, beforeEach} from 'vitest';
-import {MemoryManager, MemoryCache, rebuildGraph, Memory} from '../src/memory';
-
-function makeMemory(overrides: Partial<Memory> = {}): Memory {
-	return {
-		name: 'Test/Doc',
-		description: '',
-		content: '',
-		embedding: [],
-		links: [],
-		backlinks: [],
-		...overrides,
-	};
-}
-
-function makeLLM() {
-	return {
-		embedding: vi.fn(async (_text: string) => [{embedding: [1, 0, 0]}]),
-		ask: vi.fn(async () => undefined),
-	};
-}
-
-describe('rebuildGraph', () => {
-	it('extracts [[WikiLinks]] from content, excluding self-links', () => {
-		const a = makeMemory({name: 'A', content: '[[B]] and [[A]] and [[C]]'});
-		const b = makeMemory({name: 'B', content: 'no links here'});
-		const mem = [a, b];
-
-		rebuildGraph(mem);
-
-		expect(a.links).toEqual(['B', 'C']);
-		expect(b.links).toEqual([]);
-	});
-
-	it('computes backlinks only for links that resolve to a real node', () => {
-		const a = makeMemory({name: 'A', content: '[[B]] [[Missing]]'});
-		const b = makeMemory({name: 'B', content: ''});
-		const mem = [a, b];
-
-		rebuildGraph(mem);
-
-		expect(b.backlinks).toEqual(['A']);
-		expect(mem.find(m => m.name === 'Missing')).toBeUndefined();
-	});
-
-	it('resets stale backlinks on every rebuild (no leftover from a removed link)', () => {
-		const a = makeMemory({name: 'A', content: '[[B]]'});
-		const b = makeMemory({name: 'B', content: ''});
-		const mem = [a, b];
-		rebuildGraph(mem);
-		expect(b.backlinks).toEqual(['A']);
-
-		a.content = 'no more links';
-		rebuildGraph(mem);
-		expect(b.backlinks).toEqual([]);
-	});
-});
-
-describe('MemoryCache', () => {
-	it('finds nearest neighbor by embedding via KD-tree search', () => {
-		const close = makeMemory({name: 'Close', embedding: [1, 0, 0]});
-		const far = makeMemory({name: 'Far', embedding: [0, 0, 1]});
-		const cache = new MemoryCache([close, far]);
-
-		const results = cache.search([1, 0, 0], 1);
-
-		expect(results[0].name).toBe('Close');
-	});
-
-	it('rebuilds the tree on add/update/remove', () => {
-		const cache = new MemoryCache([makeMemory({name: 'A', embedding: [1, 0, 0]})]);
-		cache.add(makeMemory({name: 'B', embedding: [0, 1, 0]}));
-		expect(cache.search([0, 1, 0], 1)[0].name).toBe('B');
-
-		cache.remove('B');
-		expect(cache.search([0, 1, 0], 1)[0]?.name).not.toBe('B');
-	});
-});
-
-describe('MemoryManager.forget', () => {
-	it('removes the node and recomputes backlinks for the rest of the graph', () => {
-		const llm = makeLLM();
-		const mgr = new MemoryManager(llm);
-		const a = makeMemory({name: 'A', content: '[[B]]'});
-		const b = makeMemory({name: 'B', content: '[[C]]'});
-		const c = makeMemory({name: 'C', content: ''});
-		const mem = [a, b, c];
-		rebuildGraph(mem);
-		expect(c.backlinks).toEqual(['B']);
-
-		const ok = mgr.forget('B', mem);
-
-		expect(ok).toBe(true);
-		expect(mem.find(m => m.name === 'B')).toBeUndefined();
-		expect(a.links).toEqual(['B']);
-		expect(c.backlinks).toEqual([]);
-	});
-
-	it('returns false for an unknown name', () => {
-		const mgr = new MemoryManager(makeLLM());
-		expect(mgr.forget('Nope', [makeMemory({name: 'A'})])).toBe(false);
-	});
-});
-
-describe('MemoryManager.recollect', () => {
-	it('orders vector matches first, then expands one hop via links', async () => {
-		const llm = makeLLM();
-		llm.embedding.mockResolvedValue([{embedding: [1, 0, 0]}]);
-		const mgr = new MemoryManager(llm);
-
-		const near = makeMemory({name: 'Near', embedding: [1, 0, 0], content: '[[Linked]]'});
-		const linked = makeMemory({name: 'Linked', embedding: [0, 0, 1], content: ''});
-		const far = makeMemory({name: 'Far', embedding: [0, 1, 0], content: ''});
-		const mem = [near, linked, far];
-		rebuildGraph(mem);
-
-		const result = await mgr.recollect('query', mem, 1, 1);
-
-		expect(result.map(r => r.name)).toEqual(['Near', 'Linked']);
-	});
-
-	it('returns [] when there are no memories', async () => {
-		const mgr = new MemoryManager(makeLLM());
-		expect(await mgr.recollect('q', [])).toEqual([]);
-	});
-});
-
-describe('MemoryManager.memorize (fast path)', () => {
-	let llm: ReturnType<typeof makeLLM>;
-	let mgr: MemoryManager;
-
-	beforeEach(() => {
-		llm = makeLLM();
-		mgr = new MemoryManager(llm);
-	});
-
-	it('pushes a pending tool message, then resolves it to links once facts land', async () => {
-		llm.ask.mockImplementation(async (_prompt: string, opts: any) => {
-			if (opts.tools) {
-				opts.tools[0].fn({destination: 'Projects/Oxide', facts: 'Uses a hybrid memory system'});
-				return undefined;
-			}
-			return {description: 'd', content: '# doc'};
-		});
-
-		const history: any[] = [{role: 'user', content: 'we use a hybrid memory system'}];
-		const touched = await mgr.memorize(history, [], {model: 'test'} as any);
-
-		const pending = history.find(h => h.name === 'memory_process');
-		expect(pending).toBeDefined();
-		expect(pending.content).toContain('[[Projects/Oxide]]');
-		expect(touched.map(t => t.name)).toEqual(['Projects/Oxide']);
-	});
-
-	it('creates a new node and appends facts under "## Facts" without calling the doc LLM', async () => {
-		llm.ask.mockImplementation(async (_prompt: string, opts: any) => {
-			if (opts.tools) opts.tools[0].fn({destination: 'People/Sarah', facts: 'Works at Acme, Likes hiking'});
-			return undefined;
-		});
-
-		const mem: Memory[] = [];
-		await mgr.memorize([{role: 'user', content: 'Sarah works at Acme and likes hiking'}] as any, mem, {model: 'test'} as any);
-
-		const node = mem.find(m => m.name === 'People/Sarah')!;
-		expect(node).toBeDefined();
-		expect(node.content).toContain('## Facts');
-		expect(node.content).toContain('- Works at Acme');
-		expect(node.content).toContain('- Likes hiking');
-		// doc reconciler LLM (schema call) should NOT have been awaited synchronously in this fast path assertion
-	});
-
-	it('routes "journal" destination to Journal/{weekMonday}', async () => {
-		llm.ask.mockImplementation(async (_prompt: string, opts: any) => {
-			if (opts.tools) opts.tools[0].fn({destination: 'journal', facts: 'Shipped v1'});
-			return undefined;
-		});
-
-		const mem: Memory[] = [];
-		const touched = await mgr.memorize([{role: 'user', content: 'shipped v1 today'}] as any, mem, {model: 'test'} as any);
-
-		expect(touched[0].name).toMatch(/^Journal\/\d{4}-\d{2}-\d{2}$/);
-	});
-
-	it('reports nothing to remember when no facts are extracted', async () => {
-		llm.ask.mockResolvedValue(undefined); // tools present but fn never called
-
-		const history: any[] = [{role: 'user', content: 'hey'}];
-		const touched = await mgr.memorize(history, [], {model: 'test'} as any);
-
-		expect(touched).toEqual([]);
-		expect(history.find(h => h.name === 'memory_process').content).toBe('Nothing worth remembering.');
-	});
-
-	it('returns [] and does nothing for an empty conversation', async () => {
-		const touched = await mgr.memorize([], [], {model: 'test'} as any);
-		expect(touched).toEqual([]);
-		expect(llm.ask).not.toHaveBeenCalled();
-	});
-});
-
-describe('MemoryManager reconcileVault', () => {
-	it('integrates the "## Facts" section via the doc LLM and removes it', async () => {
-		const llm = makeLLM();
-		llm.ask.mockResolvedValue({description: 'Tidy summary', content: '# Doc\n\nIntegrated fact.'});
-		const mgr = new MemoryManager(llm);
-
-		const node = makeMemory({
-			name: 'Projects/Oxide',
-			content: '---\nname: Projects/Oxide\n---\n\n# Doc\n\n## Facts\n- some raw fact\n',
-		});
-		const mem = [node];
-
-		await mgr.reconcileVault(mem, {model: 'test'} as any, 'all');
-
-		expect(node.content).not.toContain('## Facts');
-		expect(node.content).toContain('Integrated fact.');
-		expect(node.description).toBe('Tidy summary');
-	});
-
-	it('only targets docs with a pending Facts inbox when scope is "touched"', async () => {
-		const llm = makeLLM();
-		llm.ask.mockResolvedValue({description: 'd', content: '# clean'});
-		const mgr = new MemoryManager(llm);
-
-		const dirty = makeMemory({name: 'A', content: '## Facts\n- x'});
-		const clean = makeMemory({name: 'B', content: '# already tidy'});
-		await mgr.reconcileVault([dirty, clean], {model: 'test'} as any, 'touched');
-
-		expect(dirty.content).toContain('# clean'); // rewritten (frontmatter now wraps it)
-		expect(clean.content).toBe('# already tidy'); // untouched, never queued
-	});
-});
-
-describe('MemoryManager reconcile coalescing', () => {
-	it('coalesces a second call while one is in-flight: marks dirty, aborts, reuses the same task promise', () => {
-		const llm = makeLLM();
-		const abort = vi.fn();
-		let calls = 0;
-		llm.ask.mockImplementation(() => {
-			calls++;
-			const pending: any = new Promise(() => {}); // never resolves in this test
-			pending.abort = abort;
-			return pending;
-		});
-		const mgr: any = new MemoryManager(llm);
-		const node = makeMemory({name: 'Q', content: '# Q\n\n## Facts\n- f'});
-		const mem = [node];
-
-		const p1 = mgr.reconcile(node, mem, {model: 'test'});
-		const p2 = mgr.reconcile(node, mem, {model: 'test'});
-
-		expect(p2).toBe(p1); // same in-flight task, not a new queue entry
-		expect(abort).toHaveBeenCalledTimes(1); // second call aborted the in-flight request
-		expect(calls).toBe(1); // no second ask() fired synchronously — it'll rerun via the dirty loop
-	});
-});