Refined memory management prompts

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "1.6.1",
+	"version": "1.6.4",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",

src/llm.ts

@@ -243,10 +243,8 @@ class LLM {
 			} else if(isText) {
 				text = (await this.loadBuffer(file, true)).toString('utf-8');
 			} else {
-				text = `Unsupported file type: ${ext || mime}`;
+				text = typeof file.content === 'string' ? file.content : `[Binary file, unable to extract: ${name}]`;
 			}
-
-			// Cache result, skip re-extraction on future turns of the same conversation
 			file.content = text;
 			file.extracted = true;
 			delete file.path;
@@ -413,7 +411,8 @@ ${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()});
+			const files = options.files || [];
+			if(message || files.length) history.push({role: 'user', content: message || '', timestamp: Date.now()});
 
 			// MCP
 			const mcp = options.mcp || this.ai.options?.llm?.mcp;
@@ -484,15 +483,16 @@ Linked: ${makeUnique([...r.links, ...r.backlinks]).join(', ')}
 
 			if(aborted) throw Object.assign(new Error('Aborted'), {name: 'AbortError'});
 
-			// Files
-			const files = options.files || [];
 			const lastMsg = history[history.length - 1];
-			const originalContent = lastMsg?.content;
-			if(files.length && lastMsg?.role === 'user') {
-				lastMsg.files = files;
-				const {text, images} = await this.resolveFiles(files);
-				const merged = text ? `${originalContent}\n\n${text}` : originalContent;
-				lastMsg.content = images.length
+			if(files.length && lastMsg?.role === 'user') lastMsg.files = files;
+			const restores: {msg: LLMMessage, content: any}[] = [];
+			for(const msg of history) {
+				if(msg.role !== 'user' || !msg.files?.length) continue;
+				const {text, images} = await this.resolveFiles(msg.files);
+				if(!text && !images.length) continue;
+				restores.push({msg, content: msg.content});
+				const merged = text ? [msg.content, text].filter(Boolean).join('\n\n') : msg.content;
+				msg.content = images.length
 					? [...images.map(i => ({type: 'image', mime: i.mime, data: i.data})), {type: 'text', text: merged}]
 					: merged;
 			}
@@ -506,7 +506,8 @@ Linked: ${makeUnique([...r.links, ...r.backlinks]).join(', ')}
 			request = this.models[m].ask('', {...options, tools, system: prompts.filter(Boolean).join('\n\n')});
 			let resp = await request;
 
-			if(files.length && lastMsg?.role === 'user') lastMsg.content = originalContent;
+			// Strip the file injection shim
+			restores.forEach(({msg, content}) => msg.content = content);
 
 			// Capture meta (duration / tps)
 			for(const h of history) {

src/memory.ts

@@ -2,9 +2,10 @@ import {MemoryNode, rebuildGraph} from './helpers.ts';
 import {LLMRequest, LLMMessage} from './llm.ts';
 import {AiTool} from './tools.ts';
 import {KDPoint, KDTree} from './kd-tree.ts';
+import {escapeRegex} from '@ztimson/utils';
 
-const FACTS_HEADING = '## Facts';
-
+const MERGE_THRESHOLD = 0.88;
+const PENDING_HEADING = '## Pending';
 const GENERIC_TEMPLATE = `# {{Title}}
 
 ## Summary
@@ -32,6 +33,11 @@ type FactBucket = {
 	facts: string[];
 }
 
+type FactAgentResult = {
+	buckets: FactBucket[];
+	journal: string;
+}
+
 function dedupeFacts(facts: string[]): string[] {
 	const seen = new Map<string, string>();
 	for (const f of facts) {
@@ -101,8 +107,8 @@ export class MemoryCache {
 	}
 
 	update(memory: Memory): void {
-		const idx = this.memories.findIndex(m => m.name === memory.name);
-		if (idx !== -1) this.memories[idx] = memory;
+		const existing = this.memories.find(m => m.name === memory.name);
+		if (existing) Object.assign(existing, memory);
 		else this.memories.push(memory);
 		this.rebuild();
 	}
@@ -163,7 +169,7 @@ class MemoryAccessor {
 		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);
+			const [e] = await llm.embedding(`${node.description}\n\n${stripHeader(node.content)}`.trim());
 			if (e) node.embedding = e.embedding;
 		}));
 		this.commit();
@@ -251,14 +257,13 @@ ${m.content}
 		return new MemoryAccessor(memories);
 	}
 
-	private appendFacts(node: Memory, facts: string[]): void {
+	private stage(node: Memory, block: string): void {
 		this.ensureDoc(node);
 		const body = stripHeader(node.content);
-		const bullets = facts.map(f => `- ${f}`).join('\n');
-		const idx = body.indexOf(FACTS_HEADING);
+		const idx = body.indexOf(PENDING_HEADING);
 		const newBody = idx === -1
-			? `${body.trimEnd()}\n\n${FACTS_HEADING}\n${bullets}\n`
-			: `${body.slice(0, idx + FACTS_HEADING.length)}\n${bullets}${body.slice(idx + FACTS_HEADING.length)}`;
+			? `${body.trimEnd()}\n\n${PENDING_HEADING}\n${block}\n`
+			: `${body.slice(0, idx + PENDING_HEADING.length)}\n${block}${body.slice(idx + PENDING_HEADING.length)}`;
 		node.content = this.touchHeader(node, newBody);
 	}
 
@@ -268,42 +273,52 @@ ${m.content}
 		node.content = this.touchHeader(node, `# ${title}\n`);
 	}
 
-	private async factAgent(conversation: string, store: MemoryAccessor, options: LLMRequest, weekKey: string): Promise<FactBucket[]> {
+	private sanitizeDescription(text: string): string {
+		return (text ?? '').replace(/\s+/g, ' ').trim().slice(0, 240);
+	}
+
+	private relink(memories: Memory[], from: string, to: string): void {
+		const pattern = new RegExp(`\\[\\[${escapeRegex(from)}\\]\\]`, 'g');
+		for (const m of memories) if (pattern.test(m.content)) m.content = m.content.replace(pattern, `[[${to}]]`);
+	}
+
+	private async factAgent(conversation: string, store: MemoryAccessor, options: LLMRequest): Promise<FactAgentResult> {
 		const ghosts = store.ghosts();
 
 		const response = await this.llm.ask(conversation, {
 			model: options.model,
 			temperature: 0.2,
-			system: `You are a fact extractor to build obsidian knowledge vaults.
-Analyze this conversation and extract facts worth remembering long-term.
+			system: `You are a fact extractor for Obsidian-style knowledge vaults. Analyze the conversation and produce:
 
-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
+1. Journal recap (single paragraph)
+   - "Captains Log" style record keeping
+   - What was discussed/worked on, decisions, user's events/state/mood, general context
+   - Leave empty only for trivial/empty exchanges/small talk
 
-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"
+2. Fact buckets
+   - ONLY facts the USER explicitly stated about themselves, their work, projects, or decisions made during this conversation
+   - NEVER extract greetings, pleasantries, or anything the assistant itself said
+   - Extract the final/end state, not deltas
+
+Path assignment rules:
+   - Reuse existing node names whenever possible
+   - Documents should be grouped and named by the root subject
+     - Person → People/Name
+     - Project → Projects/Name
+     - Concept → Concepts/Name
+   - A bug report, its investigation, should be nested and attached to the same root subject node
+   - Tickets/one-off tasks → file under the project/name/component they belong to
+   - Only create a new top-level node when the fact belongs to a genuinely new subject (person/project/concept)\`
 
 Available nodes:
-- Journal
-${this.listNodes(store.list).filter(n => !n.name.includes('Journal')).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}
+${this.listNodes(store.list).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'},
-							},
+				journal: {type: 'string', description: 'Short day-to-day recap; empty if nothing happened.', required: false},
+				buckets: {type: 'array', description: 'Groups of facts to remember; empty array if nothing worth storing.', items: {
+						type: 'object', items: {
+							subject: {type: 'string', description: 'Exact node name or new path (e.g. "People/Sarah", "Projects/Oxide")', required: true},
+							facts: {type: 'array', description: 'Facts to store here', items: {type: 'string'}},
 						},
 					},
 				},
@@ -311,15 +326,17 @@ ${ghosts.length ? `${ghosts.map(g => `- ${g}: (Ghost)`).join('\n')}` : ''}`,
 		});
 
 		const buckets = new Map<string, string[]>();
-		for(const bucket of response.buckets ?? []) {
-			const subject = bucket.subject.trim().toLowerCase() === 'journal'
-				? `Journal/${weekKey}` : bucket.subject.trim();
+		for (const bucket of response.buckets ?? []) {
+			const subject = bucket.subject.trim();
 			const facts = buckets.get(subject) ?? [];
 			facts.push(...dedupeFacts(bucket.facts));
 			buckets.set(subject, facts);
 		}
 
-		return buckets.entries().toArray().map(([subject, facts]) => ({subject, facts}));
+		return {
+			buckets: buckets.entries().toArray().map(([subject, facts]) => ({subject, facts})),
+			journal: (response.journal ?? '').trim(),
+		};
 	}
 
 	private getWeekMonday(date: Date = new Date()): string {
@@ -334,6 +351,42 @@ ${ghosts.length ? `${ghosts.map(g => `- ${g}: (Ghost)`).join('\n')}` : ''}`,
 		return memories.map(m => ({name: m.name, description: m.description}));
 	}
 
+	/** Find the nearest node above the similarity threshold and fold the smaller/less-connected one into
+	 * the other. Journals are exempt — they're partitioned by date, not topic, and merging across weeks
+	 * would wreck the timeline. Returns 'merged' if `node` absorbed another (caller should re-run the doc
+	 * agent), 'absorbed' if `node` itself got folded away (caller should stop touching it), or null. */
+	private async checkMerge(node: Memory, memories: Memory[] | MemoryCache, options: LLMRequest, threshold = MERGE_THRESHOLD): Promise<Memory | null> {
+		if (!node.embedding?.length || node.name.startsWith('Journal/')) return null;
+		const store = this.access(memories);
+
+		let closest: Memory | null = null, closestDist = Infinity;
+		for (const other of store.list) {
+			if (other.name === node.name || other.name.startsWith('Journal/') || !other.embedding?.length) continue;
+			const d = cosineDistance(node.embedding, other.embedding);
+			if (d < closestDist) { closestDist = d; closest = other; }
+		}
+		if (!closest || closestDist > threshold) return null;
+
+		const result = await this.mergeAgent(node, closest, options);
+		const merged: Memory = {name: result.name, description: this.sanitizeDescription(result.description), content: '', embedding: [], links: [], backlinks: []};
+		merged.content = this.touchHeader(merged, result.content);
+		const [e] = await this.llm.embedding(`${merged.description}\n\n${result.content}`.trim());
+		if (e) merged.embedding = e.embedding;
+
+		this.relink(store.list, node.name, merged.name);
+		this.relink(store.list, closest.name, merged.name);
+
+		this.queues.get(closest.name)?.request?.abort?.();
+		this.queues.delete(closest.name);
+
+		store.forget(node.name);
+		store.forget(closest.name);
+		store.list.push(merged);
+		store.commit();
+
+		return merged;
+	}
+
 	private reconcile(node: Memory, memories: Memory[] | MemoryCache, options: LLMRequest): Promise<void> {
 		const key = node.name;
 		const existing = this.queues.get(key);
@@ -347,9 +400,12 @@ ${ghosts.length ? `${ghosts.map(g => `- ${g}: (Ghost)`).join('\n')}` : ''}`,
 		this.queues.set(key, entry);
 		const store = this.access(memories);
 		entry.task = (async () => {
+			let current = node;
 			do {
 				entry.dirty = false;
-				await this.docAgent(node, store.list, options, entry);
+				await this.docAgent(current, store.list, options, entry);
+				const merged = await this.checkMerge(current, memories, options);
+				if (merged) { current = merged; entry.dirty = true; }
 			} while (entry.dirty);
 		})().finally(() => {
 			this.queues.delete(key);
@@ -359,6 +415,7 @@ ${ghosts.length ? `${ghosts.map(g => `- ${g}: (Ghost)`).join('\n')}` : ''}`,
 	}
 
 	private async docAgent(node: Memory, memories: Memory[], options: LLMRequest, entry: {request: {abort?: () => void} | null}): Promise<void> {
+		if(!memories.includes(node)) return;
 		const currentBody = stripHeader(node.content);
 		let update;
 		try {
@@ -367,27 +424,29 @@ ${ghosts.length ? `${ghosts.map(g => `- ${g}: (Ghost)`).join('\n')}` : ''}`,
 					model: options.model,
 					temperature: 0.3,
 					schema: {
-						description: {type: 'string', description: 'One-line description of what this document covers, no formatting or emojis', required: true},
+						description: {type: 'string', description: 'One factual sentence describing the document\'s ENTIRE SUBJECT MATTER — for use as a search/merge fingerprint', required: true},
 						content: {type: 'string', description: 'Rewritten document body in markdown, without the frontmatter block', required: true},
 					},
-					system: `You are a knowledge base editor maintaining one document in an Obsidian-style vault.
+					system: `You are a knowledge base editor maintaining one Obsidian-style document.
 
-If the document has a "${FACTS_HEADING}" section, integrate every bullet under it into the appropriate part of the document, then remove the "${FACTS_HEADING}" section entirely. If there is no such section, just tidy the document per the rules below.
+If it has a "## Pending" section, fold all new material into the appropriate part, resolve overlap, then remove the section entirely. If no section, just tidy per the rules below.
 
-Structure: follow this generic shape loosely, adapting section names/order to what the content actually needs (e.g. journal-style docs may want a timeline instead of "Details"):
-\`\`\`markdown
-${GENERIC_TEMPLATE}
-\`\`\`
+Use this loose structure, adapting headings to what the content needs:
 
-Formatting rules:
-- Use Obsidian-style markdown: # headings, **bold** for emphasis, bullet & numbered lists for grouped 1D data, tables for 2D data
-- Link related concepts with [[WikiLink]] notation using full paths like [[People/Sarah]] or [[Projects/Website]]
-- Create links for specific entities (person, place, project, program) and abstract concepts, but skip generics (car, red, dog)
-- Keep the document concise, factual, and human-readable
-- Resolve contradictions: newer facts always win — delete the outdated statement entirely, never keep both
-- Do not add frontmatter blocks, filler, preamble, or AI commentary
+# Title
+## Summary
+## Details
+## Related
 
-Other nodes in the vault (link to these instead of duplicating their content):
+Rules:
+- Contradictions: newer facts always win — delete outdated statements entirely
+- Journals (Journal/...): keep entries as a chronological timeline; clean up grammar within entries but never delete history
+- Use Obsidian markdown: # headings, **bold**, bullet/numbered lists, tables for 2D data
+- Link specific entities and concepts with [[WikiLink]] (e.g., [[Projects/KiwixServer]]); skip generics
+- Keep concise, factual, human-readable
+- NO frontmatter, filler, preamble, or AI commentary
+
+Available nodes to link to (don't duplicate their content):
 ${this.listNodes(memories).filter(n => n.name !== node.name).map(n => n.name).join(', ') || 'none'}
 
 Current document:
@@ -406,12 +465,44 @@ ${currentBody}
 		}
 
 		if (!update?.content) return;
-		node.description = node.name !== 'People/User' ? update.description : 'All information about the current user';
+		node.description = node.name !== 'People/User' ? this.sanitizeDescription(update.description) : 'All information about the current user';
 		node.content = this.touchHeader(node, update.content);
-		const [e] = await this.llm.embedding(node.content);
+		const [e] = await this.llm.embedding(`${node.description}\n\n${update.content}`.trim());
 		if (e) node.embedding = e.embedding;
 	}
 
+	private async mergeAgent(a: Memory, b: Memory, options: LLMRequest): Promise<{name: string, description: string, content: string}> {
+		return this.llm.ask('', {
+			model: options.model,
+			temperature: 0.3,
+			schema: {
+				name: {type: 'string', description: 'New path for the merged doc, collection/subject format (e.g. Projects/Oxide) — only reuse an old title if it\'s genuinely the best fit', required: true},
+				description: {type: 'string', description: 'One factual sentence describing the merged document\'s subject matter', required: true},
+				content: {type: 'string', description: 'Fully reconciled body in markdown, without frontmatter', required: true},
+			},
+			system: `You are a knowledge base editor merging two overlapping Obsidian documents into one. Newer facts win on contradiction.
+
+Structure loosely:
+
+# Title
+## Summary
+## Details
+## Related
+
+Combine both documents, resolve duplication and contradictions.
+
+Document A ("${a.name}"):
+\`\`\`markdown
+${stripHeader(a.content)}
+\`\`\`
+
+Document B ("${b.name}"):
+\`\`\`markdown
+${stripHeader(b.content)}
+\`\`\``,
+		});
+	}
+
 	private parseFrontmatter(content: string): {fm: Map<string, string>, body: string} {
 		const match = content.match(/^---\n([\s\S]*?)\n---\n?([\s\S]*)$/);
 		if (!match) return {fm: new Map(), body: content};
@@ -419,7 +510,11 @@ ${currentBody}
 		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());
+			const key = line.slice(0, i).trim();
+			const raw = line.slice(i + 1).trim();
+			let value = raw;
+			try { value = JSON.parse(raw); } catch { /* legacy unquoted value, keep raw */ }
+			fm.set(key, value);
 		}
 		return {fm, body: match[2]};
 	}
@@ -433,7 +528,7 @@ ${currentBody}
 	}
 
 	private writeFrontmatter(fm: Map<string, string>, body: string): string {
-		const lines = [...fm.entries()].map(([k, v]) => `${k}: ${v}`);
+		const lines = [...fm.entries()].map(([k, v]) => `${k}: ${JSON.stringify(String(v).replace(/\s+/g, ' ').trim())}`);
 		return `---\n${lines.join('\n')}\n---\n\n${body.trimStart()}`;
 	}
 
@@ -498,20 +593,34 @@ ${currentBody}
 		history.push(pending);
 
 		const store = this.access(memories);
-		const buckets = await this.factAgent(conversation, store, options, this.getWeekMonday());
+		const {buckets, journal} = await this.factAgent(conversation, store, options);
 		const touched: Memory[] = [];
 
+		if (journal) {
+			const journalName = `Journal/${this.getWeekMonday()}`;
+			let jnode = store.find(journalName);
+			if (!jnode) {
+				jnode = {name: journalName, description: '', content: '', embedding: [], links: [], backlinks: []};
+				store.list.push(jnode);
+			}
+			this.stage(jnode, `### ${new Date().toISOString().slice(0, 10)}\n${journal}`);
+			touched.push(jnode);
+		}
+
 		for (const {subject, facts} of buckets) {
 			let node = store.find(subject);
 			if (!node) {
 				node = {name: subject, description: '', content: '', embedding: [], links: [], backlinks: []};
 				store.list.push(node);
 			}
-			this.appendFacts(node, facts);
-			const [e] = await this.llm.embedding(node.content);
+			this.stage(node, facts.map(f => `- ${f}`).join('\n'));
+			touched.push(node);
+		}
+
+		for (const node of touched) {
+			const [e] = await this.llm.embedding(`${node.description}\n\n${stripHeader(node.content)}`.trim());
 			if (e) node.embedding = e.embedding;
 			this.touch(node.name);
-			touched.push(node);
 		}
 
 		if (touched.length) {
@@ -526,9 +635,9 @@ ${currentBody}
 		return touched;
 	}
 
-	async reconcileVault(memories: Memory[] | MemoryCache, options: LLMRequest, scope: 'touched' | 'all' = 'touched'): Promise<void> {
+	async reconcileAll(memories: Memory[] | MemoryCache, options: LLMRequest, scope: 'touched' | 'all' = 'touched'): Promise<void> {
 		const store = this.access(memories);
-		const targets = scope === 'all' ? store.list : store.list.filter(m => m.content.includes(FACTS_HEADING));
+		const targets = scope === 'all' ? store.list : store.list.filter(m => m.content.includes(PENDING_HEADING));
 		await Promise.all(targets.map(node => this.reconcile(node, memories, options)));
 		store.commit();
 	}