Pass deligate subagents full history, improved memory managment

This commit is contained in:
2026-08-04 12:24:23 -04:00
parent 7fbb42c26a
commit 9c04e58c63
4 changed files with 631 additions and 309 deletions

View File

@@ -2,6 +2,16 @@ import {LLMRequest, LLMMessage} from './llm.ts';
import {AiTool} from './tools.ts';
import {KDPoint, KDTree} from './kd-tree.ts';
const FACTS_HEADING = '## Facts';
const GENERIC_TEMPLATE = `# {{Title}}
## Summary
## Details
## Related`;
export class MemoryCache {
private tree: KDTree<MemoryRef>;
public memories: Memory[];
@@ -77,46 +87,35 @@ export type Memory = {
description: string;
content: string;
embedding: number[];
}
export type MemoryRef = {
name: string;
description: string;
}
export type FactBucket = {
subject: string;
facts: string[];
}
export type MemoryNode = {
name: string;
missing: boolean;
links: string[];
backlinks: string[];
}
type MemoryRef = {
name: string;
description: string;
}
type FactBucket = {
subject: string;
facts: string[];
}
function extractLinks(content: string): string[] {
if(!content) return [];
if (!content) return [];
const matches = content.matchAll(/\[\[([^\]]+)\]\]/g);
return [...new Set([...matches].map(m => m[1].trim()))];
}
export function extractMetadata(content: string): {links: string[], backlinks: string[]} {
const match = content.match(/^---\n([\s\S]*?)\n---/);
if (!match) return {links: [], backlinks: []};
const fm = match[1];
const getList = (key: string): string[] => {
const m = fm.match(new RegExp(`^${key}:\\s*\\[(.*)\\]$`, 'm'));
if (!m || !m[1].trim()) return [];
return m[1].split(',').map(s => s.trim().replace(/^"|"$/g, '')).filter(Boolean);
};
return {
links: getList('links'),
backlinks: getList('backlinks'),
};
export function rebuildGraph(memories: Memory[]): void {
for (const m of memories) m.links = extractLinks(m.content).filter(l => l !== m.name);
for (const m of memories) m.backlinks = [];
for (const m of memories) {
for (const link of m.links) {
const target = memories.find(t => t.name === link);
if (target) target.backlinks.push(m.name);
}
}
}
function dedupeFacts(facts: string[]): string[] {
@@ -147,23 +146,11 @@ function getWeekMonday(date: Date = new Date()): string {
return d.toISOString().slice(0, 10);
}
function getWeekSunday(monday: string): string {
const d = new Date(`${monday}T00:00:00Z`);
d.setUTCDate(d.getUTCDate() + 6);
return d.toISOString().slice(0, 10);
}
export class MemoryManager {
private recentlyTouched = new Map<string, number>();
private pendingMemorizations = new Map<string, {
memories: Memory[] | MemoryCache,
tempMemoryName: string,
timestamp: number,
}>();
private queues = new Map<string, {
pending: string[],
dirty: boolean,
request: {abort?: () => void} | null,
task: Promise<void>,
}>();
@@ -176,7 +163,7 @@ export class MemoryManager {
name: {type: 'string', description: 'Exact memory name', required: true},
},
fn: (args: any) => {
const mems = memories instanceof MemoryCache ? memories.memories : memories;
const mems = this.unwrap(memories);
const mem = mems.find(m => m.name === args.name);
if (!mem) return 'Document not found';
this.touch(mem.name);
@@ -205,110 +192,58 @@ export class MemoryManager {
return raw ? {memory: <Memory[] | MemoryCache>m, inject: true, tool: true, update: true} : {inject: true, tool: true, update: true, ...m};
}
private async createTempMemory(conversation: string): Promise<Memory> {
const timestamp = Date.now();
const content = `---
name: _temp_${timestamp}
description: Temporary memory - processing in background
tags: [_temporary]
links: []
backlinks: []
modified: ${new Date().toISOString()}
---
# Recent Conversation (Processing)
${conversation}`;
const [e] = await this.llm.embedding(content);
return {
name: `_temp_${timestamp}`,
description: 'Temporary memory - processing in background',
content,
embedding: e?.embedding || [],
};
private unwrap(memories: Memory[] | MemoryCache): Memory[] {
return memories instanceof MemoryCache ? memories.memories : memories;
}
private applyHeader(content: string, header: string): string {
return `${header}\n\n${this.stripHeader(content)}`;
private sync(memories: Memory[] | MemoryCache): void {
if (memories instanceof MemoryCache) memories.rebuild();
}
private async backgroundMemorization(conversation: string, memories: Memory[] | MemoryCache, options: LLMRequest, tempName: string): Promise<void> {
const mem = memories instanceof MemoryCache ? memories.memories : memories;
const monday = getWeekMonday();
const sunday = getWeekSunday(monday);
const buckets = await this.factAgent(conversation, mem, options, monday);
if(!buckets.length) return;
const jobs = [...buckets].map(({subject, facts}) => {
let node = mem.find(m => m.name === subject);
if(!node) {
node = {name: subject, description: '', content: '', embedding: [],};
mem.push(node);
}
const week = subject.startsWith('Journal/') ? {monday, sunday} : undefined;
return this.enqueue(node, facts, mem, options, tempName, week);
});
await Promise.all(jobs);
}
private buildHeader(node: Memory, week?: {monday: string, sunday: string}, links: string[] = [], backlinks: string[] = []): string {
const tags = node.name.split('/')[0]?.toLowerCase();
const lines = [
'---',
`name: ${node.name}`,
`description: ${node.description || ''}`,
tags ? `tags: [${tags}]` : '',
links.length ? `links: [${links.map(l => `"${l}"`).join(', ')}]` : 'links: []',
backlinks.length ? `backlinks: [${backlinks.map(l => `"${l}"`).join(', ')}]` : 'backlinks: []',
week ? `week: ${week.monday} ${week.sunday}` : '',
`modified: ${new Date().toISOString()}`,
'---',
].filter(Boolean);
return lines.join('\n');
}
private cosineSearch(query: number[], memories: Memory[], limit: number): MemoryRef[] {
const scored = memories
.filter(m => m.embedding?.length)
.map(m => ({
ref: {name: m.name, description: m.description},
distance: cosineDistance(query, m.embedding),
}))
.sort((a, b) => a.distance - b.distance)
.slice(0, limit);
return scored.map(s => s.ref);
}
/**
* Coalescing queue: if a doc is already compiling, abort the in-flight run, merge its
* facts with the new ones and restart. Never blocks a pending update, never drops facts.
*/
private enqueue(node: Memory, facts: string[], memories: Memory[] | MemoryCache, options: LLMRequest, tempName: string, week?: {monday: string, sunday: string}): Promise<void> {
const key = node.name;
const existing = this.queues.get(key);
if (existing) {
existing.pending.push(...facts);
existing.request?.abort?.();
return existing.task;
private parseFrontmatter(content: string): {fm: Map<string, string>, body: string} {
const match = content.match(/^---\n([\s\S]*?)\n---\n?([\s\S]*)$/);
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());
}
const entry: {pending: string[], request: {abort?: () => void} | null, task: Promise<void>} = {pending: [...facts], request: null, task: Promise.resolve()};
this.queues.set(key, entry);
const m = memories instanceof MemoryCache ? memories.memories : memories;
entry.task = (async () => {
while (entry.pending.length) {
const batch = dedupeFacts(entry.pending.splice(0, entry.pending.length));
const written = await this.docAgent(node, batch, m, options, tempName, week, entry);
if (!written) entry.pending.unshift(...batch);
}
})().finally(() => {
this.queues.delete(key);
if(!this.queues.size && memories instanceof MemoryCache) memories.rebuild();
});
return entry.task;
return {fm, body: match[2]};
}
private listNodes(memories: Memory[]): MemoryRef[] {
return memories.map(m => ({name: m.name, description: m.description}));
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 {
this.ensureDoc(node);
const body = this.stripHeader(node.content);
const bullets = facts.map(f => `- ${f}`).join('\n');
const idx = body.indexOf(FACTS_HEADING);
const newBody = idx === -1
? `${body.trimEnd()}\n\n${FACTS_HEADING}\n${bullets}\n`
: `${body.slice(0, idx + FACTS_HEADING.length)}\n${bullets}${body.slice(idx + FACTS_HEADING.length)}`;
node.content = this.touchHeader(node, newBody);
}
decay() {
@@ -318,71 +253,27 @@ ${conversation}`;
}
}
forget(name: string, memories: Memory[] | MemoryCache): boolean {
const mem = memories instanceof MemoryCache ? memories.memories : memories;
const idx = mem.findIndex(m => m.name === name);
if (idx === -1) return false;
for (const node of mem) {
const {links, backlinks} = extractMetadata(node.content);
const newBacklinks = backlinks.filter(b => b !== name);
const newLinks = links.filter(l => l !== name);
if (newBacklinks.length !== backlinks.length || newLinks.length !== links.length) {
node.content = this.updateFrontmatter(node.content, {
links: newLinks,
backlinks: newBacklinks,
});
}
}
mem.splice(idx, 1);
if (memories instanceof MemoryCache) memories.rebuild();
return true;
touch(name: string, ttl = 2) {
this.recentlyTouched.set(name, ttl);
}
getTouched(): string[] {
return [...this.recentlyTouched.keys()];
}
async memorize(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest): Promise<Memory[]> {
const conversation = history
.filter(h => h.role === 'user' || h.role === 'assistant')
.map(h => `[${h.role}]: ${h.content}`).join('\n\n').trim();
if(!conversation) return [];
forget(name: string, memories: Memory[] | MemoryCache): boolean {
const mem = this.unwrap(memories);
const idx = mem.findIndex(m => m.name === name);
if (idx === -1) return false;
const trackingId = `${Date.now()}_${Math.random()}`;
const tempMemory = await this.createTempMemory(conversation);
const mem = memories instanceof MemoryCache ? memories.memories : memories;
mem.push(tempMemory);
if (memories instanceof MemoryCache) memories.rebuild();
this.pendingMemorizations.set(trackingId, {
memories,
tempMemoryName: tempMemory.name,
timestamp: Date.now(),
});
try {
await this.backgroundMemorization(conversation, memories, options, tempMemory.name);
const finalMem = memories instanceof MemoryCache ? memories.memories : memories;
return finalMem.filter(m => !m.name.startsWith('_temp_'));
} finally {
const pending = this.pendingMemorizations.get(trackingId);
if (pending) {
const cleanMem = pending.memories instanceof MemoryCache
? pending.memories.memories
: pending.memories;
const idx = cleanMem.findIndex(m => m.name === pending.tempMemoryName);
if (idx !== -1) cleanMem.splice(idx, 1);
if (pending.memories instanceof MemoryCache) pending.memories.rebuild();
}
this.pendingMemorizations.delete(trackingId);
}
mem.splice(idx, 1);
rebuildGraph(mem);
this.sync(memories);
return true;
}
async recollect(query: string, memories: Memory[] | MemoryCache, limit = 5, graphDepth = 1): Promise<Memory[]> {
const mem: Memory[] = memories instanceof MemoryCache ? memories.memories : memories;
const mem = this.unwrap(memories);
if (!mem.length) return [];
const [e] = await this.llm.embedding(query);
@@ -400,8 +291,7 @@ ${conversation}`;
for (const name of frontier) {
const node = mem.find(m => m.name === name);
if (!node) continue;
const {links} = extractMetadata(node.content);
for (const link of links) {
for (const link of node.links) {
if (!found.has(link) && mem.find(m => m.name === link)) {
found.add(link);
next.push(link);
@@ -419,114 +309,151 @@ ${conversation}`;
return ordered.map(n => mem.find(m => m.name === n)!).filter(Boolean);
}
touch(name: string, ttl = 2) {
this.recentlyTouched.set(name, ttl);
private cosineSearch(query: number[], memories: Memory[], limit: number): MemoryRef[] {
const scored = memories
.filter(m => m.embedding?.length)
.map(m => ({
ref: {name: m.name, description: m.description},
distance: cosineDistance(query, m.embedding),
}))
.sort((a, b) => a.distance - b.distance)
.slice(0, limit);
return scored.map(s => s.ref);
}
private updateFrontmatter(content: string, updates: {links?: string[], backlinks?: string[]}): string {
const match = content.match(/^---\n([\s\S]*?)\n---\n\n?([\s\S]*)$/);
if (!match) return content;
private listNodes(memories: Memory[]): MemoryRef[] {
return memories.map(m => ({name: m.name, description: m.description}));
}
const [, fm, body] = match;
let newFm = fm;
async memorize(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest): Promise<Memory[]> {
const conversation = history
.filter(h => h.role === 'user' || h.role === 'assistant')
.map(h => `[${h.role}]: ${h.content}`).join('\n\n').trim();
if (!conversation) return [];
if (updates.links !== undefined) {
const linksList = updates.links.length ? `[${updates.links.map(l => `"${l}"`).join(', ')}]` : '[]';
newFm = newFm.replace(/^links:.*$/m, `links: ${linksList}`);
const uid = `${Date.now()}_${Math.random().toString(36).slice(2)}`;
// NOTE: adjust field names below (id/tool_call_id/name) to match your LLMMessage/tool-call schema.
const pending = {role: 'tool', name: 'memory_process', id: uid, content: 'Processing…'} as unknown as LLMMessage;
history.push(pending);
const mem = this.unwrap(memories);
const buckets = await this.factAgent(conversation, mem, options, getWeekMonday());
const touched: Memory[] = [];
for (const {subject, facts} of buckets) {
let node = mem.find(m => m.name === subject);
if (!node) {
node = {name: subject, description: '', content: '', embedding: [], links: [], backlinks: []};
mem.push(node);
}
this.appendFacts(node, facts);
const [e] = await this.llm.embedding(node.content);
if (e) node.embedding = e.embedding;
this.touch(node.name);
touched.push(node);
}
if (updates.backlinks !== undefined) {
const backlinksList = updates.backlinks.length ? `[${updates.backlinks.map(l => `"${l}"`).join(', ')}]` : '[]';
newFm = newFm.replace(/^backlinks:.*$/m, `backlinks: ${backlinksList}`);
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.';
}
newFm = newFm.replace(/^modified:.*$/m, `modified: ${new Date().toISOString()}`);
return `---\n${newFm}\n---\n\n${body}`;
return touched;
}
private stripHeader(content: string): string {
return content.replace(/^---[\s\S]*?\n---\n?/, '').trimStart();
/** 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);
}
private async docAgent(node: Memory, facts: string[], memories: Memory[], options: LLMRequest, tempName: string, week: {monday: string, sunday: string} | undefined, entry: {request: {abort?: () => void} | null}): Promise<boolean> {
const {links: oldLinks} = extractMetadata(node.content);
/**
* 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);
if (existing) {
existing.dirty = true;
existing.request?.abort?.();
return existing.task;
}
const entry = {dirty: false, request: null, task: Promise.resolve()};
this.queues.set(key, entry);
const mem = this.unwrap(memories);
entry.task = (async () => {
do {
entry.dirty = false;
await this.reconcileDoc(node, mem, options, entry);
} while (entry.dirty);
})().finally(() => {
this.queues.delete(key);
rebuildGraph(mem);
this.sync(memories);
});
return entry.task;
}
private async reconcileDoc(node: Memory, memories: Memory[], options: LLMRequest, entry: {request: {abort?: () => void} | null}): Promise<void> {
const currentBody = this.stripHeader(node.content);
let update;
try {
for(let i = 0; i < 3 && !update?.content; i++) {
const request = this.llm.ask(`New Facts:\n${facts.map(f => `- ${f}`).join('\n')}`, {
for (let i = 0; i < 2 && !update?.content; i++) {
const request = this.llm.ask(currentBody, {
model: options.model,
temperature: 0.3,
schema: {
description: {type: 'string', description: 'One-line description of what this document covers, no formatting or emojis', required: true},
content: {type: 'string', description: 'Rewritten document in markdown, without the frontmatter block', required: true},
content: {type: 'string', description: 'Rewritten document body in markdown, without the frontmatter block', required: true},
},
system: `You are a knowledge base editor. Rewrite the current document below so it incorporates the new facts.
system: `You are a knowledge base editor maintaining one document in an Obsidian-style vault.
If the document has a "${FACTS_HEADING}" section, integrate every bullet under it into the appropriate part of the document, then remove the "${FACTS_HEADING}" section entirely. If there is no such section, just tidy the document per the rules below.
Structure: follow this generic shape loosely, adapting section names/order to what the content actually needs (e.g. journal-style docs may want a timeline instead of "Details"):
\`\`\`markdown
${GENERIC_TEMPLATE}
\`\`\`
Formatting rules:
- Use Obsidian-style markdown: # headings, **bold** to add emphasis, __italics__ for titles, terms, etc, bullet & numbered lists for grouped 1D data and tables for 2D data
- 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 (quantum mechanics, entropy) but skip generics (car, red, dog)
- 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: the new facts always win — delete the outdated statement entirely, never keep both
- Later facts in the list override earlier ones
- Resolve contradictions: newer facts always win — delete the outdated statement entirely, never keep both
- Do not add frontmatter blocks, filler, preamble, or AI commentary
${week ? '- This is a weekly journal entry.\n' : ''}
All nodes:
${this.listNodes(memories).map(n => n.name).join(', ') || 'none'}
Other nodes in the vault (link to these instead of duplicating their content):
${this.listNodes(memories).filter(n => n.name !== node.name).map(n => n.name).join(', ') || 'none'}
Current document:
\`\`\`markdown
${currentBody}
\`\`\``}
);
\`\`\``,
});
entry.request = request;
update = await request;
}
} catch (err: any) {
if (err?.name === 'AbortError') return false;
if (err?.name === 'AbortError') return;
throw err;
} finally {
entry.request = null;
}
if(!update?.content) return false;
const newLinks = extractLinks(update.content).filter(l => l !== node.name && l !== tempName);
const newLinkSet = new Set(newLinks);
const oldLinkSet = new Set(oldLinks);
for (const added of newLinkSet) {
if (!oldLinkSet.has(added)) {
const target = memories.find(m => m.name === added);
if (target) {
const {backlinks} = extractMetadata(target.content);
if (!backlinks.includes(node.name)) {
target.content = this.updateFrontmatter(target.content, {
backlinks: [...backlinks, node.name],
});
}
}
}
}
for (const removed of oldLinkSet) {
if (!newLinkSet.has(removed)) {
const target = memories.find(m => m.name === removed);
if (target) {
const {backlinks} = extractMetadata(target.content);
target.content = this.updateFrontmatter(target.content, {
backlinks: backlinks.filter(b => b !== node.name),
});
}
}
}
const {backlinks} = extractMetadata(node.content);
node.description = node.name !== 'Person/User' ? update.description : 'All information about the current user';
node.content = this.applyHeader(update.content, this.buildHeader(node, week, newLinks, backlinks));
if (!update?.content) return;
node.description = node.name !== 'People/User' ? update.description : 'All information about the current user';
node.content = this.touchHeader(node, update.content);
const [e] = await this.llm.embedding(node.content);
if(e) node.embedding = e.embedding;
return true;
if (e) node.embedding = e.embedding;
}
private async factAgent(conversation: string, memories: Memory[], options: LLMRequest, weekKey: string): Promise<FactBucket[]> {
@@ -546,13 +473,13 @@ Rules:
When extracting facts, you MUST also decide the exact destination path:
- Use an existing node name if the facts clearly belong there
- All information primary about the user should go under "People/User"
- When required, create a new path following collection/subject format (e.g., People/Sarah, Projects/Oxide)
- 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(memories).filter(n => !n.name.includes('_temp_') && !n.name.includes('Journal')).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}`,
${this.listNodes(memories).filter(n => !n.name.includes('Journal')).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}`,
tools: [{
name: 'facts_extract',
description: 'Submit facts with their destination',