Memory refinement WIP
This commit is contained in:
+94
-723
@@ -1,95 +1,72 @@
|
||||
import {MemoryNode, patchGraph, rebuildGraph} from './helpers.ts';
|
||||
import {LLMRequest, LLMMessage} from './llm.ts';
|
||||
import {AiTool} from './tools.ts';
|
||||
import {KDTree} from './kd-tree.ts';
|
||||
import {KDTree} from './memory/kd-tree.ts';
|
||||
import {Memory, MemoryRef, MemoryStore} from './memory/memory.ts';
|
||||
import {cosineDistance, embedMemoryFields} from './utils.ts';
|
||||
|
||||
const FACT_SIMILARITY_THRESHOLD = 0.62;
|
||||
const PENDING_HEADING = '## Pending';
|
||||
const TODO_HEADING = '## Todo list';
|
||||
const TREE_TOMBSTONE_LIMIT = 0.25;
|
||||
const ALIAS_MATCH_THRESHOLD = 0.55;
|
||||
|
||||
export type Memory = {
|
||||
name: string;
|
||||
description: string;
|
||||
content: string;
|
||||
embedding: number[];
|
||||
titleEmbedding?: number[];
|
||||
bodyEmbeddings?: number[][];
|
||||
links: string[];
|
||||
backlinks: string[];
|
||||
}
|
||||
|
||||
type MemoryRef = {
|
||||
name: string;
|
||||
description: string;
|
||||
distance?: number;
|
||||
}
|
||||
|
||||
type FactBucket = {
|
||||
subject: string;
|
||||
facts: string[];
|
||||
}
|
||||
|
||||
type MemoryTask = {
|
||||
/** Exact node name / new persistent entity path this task belongs to, or '' for a personal task with no entity (goes to the journal) */
|
||||
subject: string;
|
||||
task: string;
|
||||
done: boolean;
|
||||
}
|
||||
|
||||
type FactAgentResult = {
|
||||
buckets: FactBucket[];
|
||||
journal: string;
|
||||
tasks: MemoryTask[];
|
||||
}
|
||||
|
||||
function dedupeFacts(facts: string[]): string[] {
|
||||
const seen = new Map<string, string>();
|
||||
for(const f of facts) {
|
||||
const clean = f.trim();
|
||||
if(clean) seen.set(clean.toLowerCase(), clean);
|
||||
export function memoryStore(memories: MemoryStore): {
|
||||
list: Memory[];
|
||||
cache: MemoryCache | null;
|
||||
find: (name: string) => Memory | undefined;
|
||||
ghosts: () => string[];
|
||||
search: (vector: number[], limit: number) => MemoryRef[];
|
||||
forget: (name: string) => boolean;
|
||||
rebuild: (changed?: Memory[]) => MemoryNode[];
|
||||
backfillEmbeddings: (llm: any) => Promise<number>;
|
||||
} {
|
||||
if(memories instanceof MemoryCache) {
|
||||
return {
|
||||
list: memories.memories,
|
||||
cache: memories,
|
||||
find: name => memories.find(name),
|
||||
ghosts: () => memories.ghosts(),
|
||||
search: (vector, limit) => memories.search(vector, limit),
|
||||
forget: name => memories.remove(name),
|
||||
rebuild: changed => memories.rebuild(changed),
|
||||
backfillEmbeddings: llm => memories.backfillEmbeddings(llm),
|
||||
};
|
||||
}
|
||||
return [...seen.values()];
|
||||
}
|
||||
|
||||
function cosineDistance(a: number[], b: number[]): number {
|
||||
let dot = 0, normA = 0, normB = 0;
|
||||
for(let i = 0; i < a.length; i++) {
|
||||
dot += a[i] * b[i];
|
||||
normA += a[i] * a[i];
|
||||
normB += b[i] * b[i];
|
||||
}
|
||||
const denom = Math.sqrt(normA) * Math.sqrt(normB);
|
||||
return denom === 0 ? 1 : 1 - dot / denom;
|
||||
}
|
||||
return {
|
||||
list: memories,
|
||||
cache: null,
|
||||
find: name => memories.find(m => m.name === name),
|
||||
ghosts: () => rebuildGraph(memories).filter(n => n.missing).map(n => n.name),
|
||||
search: (vector, limit) => memories
|
||||
.filter(m => m.embedding?.length)
|
||||
.map(m => ({
|
||||
name: m.name,
|
||||
description: m.description,
|
||||
distance: cosineDistance(vector, m.embedding),
|
||||
}))
|
||||
.sort((a, b) => a.distance - b.distance)
|
||||
.slice(0, limit),
|
||||
forget: name => {
|
||||
const idx = memories.findIndex(m => m.name === name);
|
||||
if(idx === -1) return false;
|
||||
memories.splice(idx, 1);
|
||||
return true;
|
||||
},
|
||||
rebuild: changed => rebuildGraph(memories),
|
||||
backfillEmbeddings: async llm => {
|
||||
const missing = memories.filter(m => !m.embedding?.length);
|
||||
if(!missing.length) return 0;
|
||||
|
||||
function cosineSearch(query: number[], memories: Memory[], limit: number): MemoryRef[] {
|
||||
return memories
|
||||
.filter(m => m.embedding?.length)
|
||||
.map(m => ({name: m.name, description: m.description, distance: cosineDistance(query, m.embedding)}))
|
||||
.sort((a, b) => a.distance - b.distance)
|
||||
.slice(0, limit);
|
||||
}
|
||||
await Promise.all(missing.map(async node => {
|
||||
const body = node.content.replace(/^---[\s\S]*?\n---\n?/, '').trimStart();
|
||||
const [titleE] = await llm.embedding(node.name.split('/').pop() || node.name);
|
||||
const [descE] = await llm.embedding(node.description || '');
|
||||
const bodyChunks = body ? await llm.embedding(body) : [];
|
||||
|
||||
async function embedMemoryFields(node: Memory, llm: any): Promise<void> {
|
||||
const body = stripHeader(node.content);
|
||||
const [titleE] = await llm.embedding(node.name.split('/').pop() || node.name);
|
||||
const [descE] = await llm.embedding(node.description || '');
|
||||
const bodyChunks = body ? await llm.embedding(body) : [];
|
||||
if(titleE) node.titleEmbedding = titleE.embedding;
|
||||
if(descE) node.embedding = descE.embedding;
|
||||
node.bodyEmbeddings = bodyChunks.map((c: any) => c.embedding).filter(Boolean);
|
||||
}
|
||||
if(titleE) node.titleEmbedding = titleE.embedding;
|
||||
if(descE) node.embedding = descE.embedding;
|
||||
node.bodyEmbeddings = bodyChunks.map((c: any) => c.embedding).filter(Boolean);
|
||||
}));
|
||||
|
||||
export function stripHeader(content: string): string {
|
||||
return content.replace(/^---[\s\S]*?\n---\n?/, '').trimStart();
|
||||
}
|
||||
|
||||
/** True if a task has no persistent entity of its own and belongs in the journal instead. */
|
||||
function isPersonalTask(t: MemoryTask): boolean {
|
||||
const s = (t.subject ?? '').trim().toLowerCase();
|
||||
return !s || s === 'journal' || s.startsWith('journal/');
|
||||
return missing.length;
|
||||
},
|
||||
};
|
||||
}
|
||||
|
||||
export class MemoryCache {
|
||||
@@ -98,7 +75,9 @@ export class MemoryCache {
|
||||
public memories: Memory[];
|
||||
public nodes: MemoryNode[] = [];
|
||||
|
||||
get length() { return this.memories.length; }
|
||||
get length() {
|
||||
return this.memories.length;
|
||||
}
|
||||
|
||||
constructor(memories: Memory[]) {
|
||||
this.memories = memories;
|
||||
@@ -106,6 +85,10 @@ export class MemoryCache {
|
||||
this.rebuild();
|
||||
}
|
||||
|
||||
find(name: string): Memory | undefined {
|
||||
return this.memories.find(m => m.name === name);
|
||||
}
|
||||
|
||||
private syncTree(): void {
|
||||
const current = new Set(this.memories.map(m => m.name));
|
||||
|
||||
@@ -120,8 +103,11 @@ export class MemoryCache {
|
||||
for(const mem of this.memories) {
|
||||
if(!mem.embedding?.length || this.indexed.has(mem.name)) continue;
|
||||
if(this.tree.dims === 0) this.tree = new KDTree<MemoryRef>(mem.embedding.length, 'cosine');
|
||||
if(mem.embedding.length !== this.tree.dims) continue; // guard against embedding model/dim drift
|
||||
this.tree.insert({vector: mem.embedding, payload: {name: mem.name, description: mem.description}});
|
||||
if(mem.embedding.length !== this.tree.dims) continue;
|
||||
this.tree.insert({
|
||||
vector: mem.embedding,
|
||||
payload: {name: mem.name, description: mem.description},
|
||||
});
|
||||
this.indexed.set(mem.name, mem.embedding);
|
||||
}
|
||||
|
||||
@@ -139,656 +125,41 @@ export class MemoryCache {
|
||||
}
|
||||
|
||||
update(memory: Memory): void {
|
||||
const existing = this.memories.find(m => m.name === memory.name);
|
||||
const existing = this.find(memory.name);
|
||||
if(existing) Object.assign(existing, memory);
|
||||
else this.memories.push(memory);
|
||||
this.rebuild([existing ?? memory]);
|
||||
}
|
||||
|
||||
remove(name: string): void {
|
||||
remove(name: string): boolean {
|
||||
const idx = this.memories.findIndex(m => m.name === name);
|
||||
if(idx !== -1) {
|
||||
this.memories.splice(idx, 1);
|
||||
this.rebuild();
|
||||
}
|
||||
}
|
||||
|
||||
rebuild(changed?: Memory[]): void {
|
||||
this.nodes = (changed?.length && this.nodes.length)
|
||||
? patchGraph(this.memories, this.nodes, changed)
|
||||
: rebuildGraph(this.memories);
|
||||
this.syncTree();
|
||||
}
|
||||
}
|
||||
|
||||
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(changed?: Memory[]): MemoryNode[] {
|
||||
if(this.cache) {
|
||||
this.cache.rebuild(changed);
|
||||
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();
|
||||
this.memories.splice(idx, 1);
|
||||
this.rebuild();
|
||||
return true;
|
||||
}
|
||||
|
||||
ghosts(): string[] {
|
||||
return this.nodes.filter(n => n.missing).map(n => n.name);
|
||||
}
|
||||
|
||||
rebuild(changed?: Memory[]): MemoryNode[] {
|
||||
this.nodes = changed?.length && this.nodes.length
|
||||
? patchGraph(this.memories, this.nodes, changed)
|
||||
: rebuildGraph(this.memories);
|
||||
this.syncTree();
|
||||
return this.nodes;
|
||||
}
|
||||
|
||||
commit(changed?: Memory[]): MemoryNode[] {
|
||||
return this.rebuild(changed);
|
||||
}
|
||||
|
||||
async backfillEmbeddings(llm: any): Promise<number> {
|
||||
const missing = this.list.filter(m => !m.embedding?.length);
|
||||
const missing = this.memories.filter(m => !m.embedding?.length);
|
||||
if(!missing.length) return 0;
|
||||
await Promise.all(missing.map(node => embedMemoryFields(node, llm)));
|
||||
this.commit();
|
||||
this.commit(missing);
|
||||
return missing.length;
|
||||
}
|
||||
}
|
||||
|
||||
export type MemoryOptions = {
|
||||
/** Memory object */
|
||||
memory: Memory[] | MemoryCache;
|
||||
/** Inject N memories into the system prompt */
|
||||
inject?: boolean;
|
||||
/** expose recall tool to LLM */
|
||||
tool?: boolean;
|
||||
/** Update memory on compression */
|
||||
update?: boolean;
|
||||
/** Max context size of memories to inject to each call (removed immediately after use) */
|
||||
maxTokens?: number;
|
||||
}
|
||||
|
||||
export class MemoryManager {
|
||||
private mergeLock: Promise<any> = Promise.resolve();
|
||||
private queues = new Map<string, {
|
||||
dirty: boolean,
|
||||
request: {abort?: () => void} | null,
|
||||
task: Promise<void>,
|
||||
}>();
|
||||
private recentlyTouched = new Map<string, number>();
|
||||
|
||||
tools = {
|
||||
forget: (memories: Memory[] | MemoryCache): AiTool => ({
|
||||
name: 'memory_forget',
|
||||
description: 'Permanently delete a memory document and clean up all references to it',
|
||||
args: {
|
||||
name: {type: 'string', description: 'Exact memory name to forget', required: true}
|
||||
},
|
||||
fn: (args: any) => {
|
||||
const result = this.forget(args.name, memories);
|
||||
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 = new MemoryAccessor(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;
|
||||
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 stage(node: Memory, block: string): void {
|
||||
if(!node.content) {
|
||||
const title = node.name.split('/').pop() ?? node.name;
|
||||
node.content = this.touchHeader(node, `# ${title}\n`);
|
||||
}
|
||||
const body = stripHeader(node.content);
|
||||
const idx = body.indexOf(PENDING_HEADING);
|
||||
const newBody = idx === -1
|
||||
? `${body.trimEnd()}\n\n${PENDING_HEADING}\n${block}\n`
|
||||
: `${body.slice(0, idx + PENDING_HEADING.length)}\n${block}${body.slice(idx + PENDING_HEADING.length)}`;
|
||||
node.content = this.touchHeader(node, newBody);
|
||||
}
|
||||
|
||||
private resolveSubject(subject: string, store: MemoryAccessor): string {
|
||||
function normalize(name: string): string {
|
||||
return name.trim().toLowerCase().replace(/\s+/g, ' ');
|
||||
}
|
||||
|
||||
const trimmed = subject.trim();
|
||||
const exact = store.find(trimmed);
|
||||
if(exact) return exact.name;
|
||||
|
||||
const normalized = normalize(trimmed);
|
||||
const caseInsensitive = store.list.find(m => normalize(m.name) === normalized);
|
||||
if(caseInsensitive) return caseInsensitive.name;
|
||||
|
||||
const root = trimmed.split('/')[0];
|
||||
const leaf = trimmed.split('/').slice(1).join('/') || trimmed;
|
||||
const candidates = store.list.filter(m => m.name.split('/')[0] === root && m.name !== trimmed);
|
||||
if(!candidates.length) return trimmed;
|
||||
|
||||
const leaves = candidates.map(m => m.name.split('/').slice(1).join('/') || m.name);
|
||||
const probe = leaves.length > 1 ? leaves : [...leaves, ''];
|
||||
const {max, similarities} = this.llm.fuzzyMatch(leaf, ...probe);
|
||||
if(max >= ALIAS_MATCH_THRESHOLD) return candidates[similarities.indexOf(max)].name;
|
||||
|
||||
return trimmed;
|
||||
}
|
||||
|
||||
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: `Turn this conversation into a persistent memory file by extracting information into organized bullet points
|
||||
|
||||
Think of this like an Obsidian vault with a clear division of responsibility:
|
||||
- The JOURNAL is a timeline. It answers "what happened, and when" and is the only place with a sense of time.
|
||||
- ENTITY DOSSIERS are a wiki. They answer "what is currently true about this subject", with no sense of time — only current state.
|
||||
- Never blur the two: a one-off event, conversation, or debugging session is a journal entry, not an entity, even if it's detailed.
|
||||
|
||||
1. Journal Log
|
||||
- A chronological, skimmable log of what actually happened: real discussions, decisions made, progress on projects, problems worked through
|
||||
- This is NOT a transcript, and it is NOT a step-by-step record, its a compressed log of notable events & developments
|
||||
- One line per development is usually enough: what was worked on and the outcome, not the blow-by-blow of how
|
||||
- Skip small talk and trivial exchanges entirely. Skip anything that's a todo item (goes in Todo Tasks) or a durable fact about a subject (goes in Entity Dossiers)
|
||||
|
||||
2. Todo Tasks
|
||||
- Extract concrete tasks the user says need to be done, should be done, or were completed
|
||||
- Return the task text and whether it is still todo or is done
|
||||
- A completed task should be marked done, not recreated as a new todo
|
||||
- Only extract actionable tasks, not general goals or observations, if none - omit returning a tasks array
|
||||
- Assign each task a subject:
|
||||
- If the task belongs to a persistent entity (a project, a class, etc.), use that entity's exact node name, or a new entity path if it doesn't exist yet
|
||||
- If it's a personal/life task with no entity of its own (reach out to someone, reply to an email, pay a bill, etc.), leave subject as an empty string — it belongs in the journal, not a new document
|
||||
|
||||
3. Entity Dossiers
|
||||
- Detailed dossiers with all factual information regarding a subject
|
||||
- Record the final/end state, not intermediate changes
|
||||
- Ignore assistant claims, guesses, greetings, or temporary details
|
||||
- NEVER create a dossier for something I wouldnt find in a wiki site: temporary information, debugging, guesses, conversations (this is all journal entry stuff!)
|
||||
- identify its HOME ENTITY:
|
||||
- The HOME ENTITY name should always be a [abstract|pro]noun
|
||||
- The grammatical subject/owner of the fact is the strongest clue
|
||||
- Always preference an existing entity over creating a new one
|
||||
- New child entities are appropriate only when they are themselves distinct persistent entities
|
||||
- A document represents a persistent entity, not a topic, feature, bug, event, decision, setting, or conversation fragment
|
||||
- Put project facts under the project they belong to, person facts under the person, etc
|
||||
|
||||
Example Entity Naming Convention:
|
||||
- Projects/[Name]
|
||||
- People/[Name]
|
||||
- History/[Name]
|
||||
- Science/[Name]
|
||||
- [Subject]/[Name]
|
||||
- Class/[Name]/[Chapter]
|
||||
|
||||
Use [[WikiLinks]] to express relationships between entities. NEVER create documents just to hold relationships
|
||||
Keep journal material in the journal; don't turn journal events into entities unless they represent something persistent
|
||||
|
||||
Available nodes:
|
||||
${this.listNodes(store.list).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}
|
||||
${ghosts.length ? `${ghosts.map(g => `- ${g}: (Ghost)`).join('\n')}` : ''}`,
|
||||
schema: {
|
||||
journal: {type: 'string', description: 'Short bullet point recap, omit if nothing notable happened'},
|
||||
tasks: {
|
||||
type: 'array', description: 'Concrete tasks mentioned or completed in the conversation, omit if none', items: {
|
||||
type: 'object', items: {
|
||||
subject: {type: 'string', description: 'Exact node name / new persistent entity path this task belongs to, or an empty string if this is a personal task with no entity of its own (those go in the journal)', required: true},
|
||||
task: {type: 'string', description: 'Concise actionable task', required: true},
|
||||
done: {type: 'boolean', description: 'Whether the task is completed', required: true},
|
||||
},
|
||||
}
|
||||
},
|
||||
buckets: {
|
||||
type: 'array', description: 'Groups of facts to remember; omit if none', items: {
|
||||
type: 'object', items: {
|
||||
subject: {type: 'string', description: 'Exact node name or new persistent entity path', required: true},
|
||||
facts: {type: 'array', description: 'Facts to store here', items: {type: 'string'}},
|
||||
},
|
||||
},
|
||||
},
|
||||
},
|
||||
});
|
||||
|
||||
const buckets = new Map<string, string[]>();
|
||||
for(const bucket of response.buckets ?? []) {
|
||||
const subject = bucket.subject.trim();
|
||||
const facts = buckets.get(subject) ?? [];
|
||||
facts.push(...dedupeFacts(bucket.facts));
|
||||
buckets.set(subject, facts);
|
||||
}
|
||||
|
||||
return {
|
||||
buckets: buckets.entries().toArray().map(([subject, facts]) => ({subject, facts})),
|
||||
journal: (response.journal ?? '').trim(),
|
||||
tasks: response.tasks ?? [],
|
||||
};
|
||||
}
|
||||
|
||||
private getWeekStart(date: Date = new Date()): string {
|
||||
const d = new Date(Date.UTC(date.getFullYear(), date.getMonth(), date.getDate()));
|
||||
const day = d.getUTCDay();
|
||||
const diff = day === 0 ? -6 : 1 - day;
|
||||
d.setUTCDate(d.getUTCDate() + diff);
|
||||
return d.toISOString().slice(0, 10);
|
||||
}
|
||||
|
||||
private journalDescription(journalName?: string): string {
|
||||
const start = journalName?.split('/').pop() || this.getWeekStart();
|
||||
const d = new Date(`${start}T00:00:00Z`);
|
||||
d.setUTCDate(d.getUTCDate() + 6);
|
||||
const end = d.toISOString().slice(0, 10);
|
||||
return `Log from ${start} - ${end}`;
|
||||
}
|
||||
|
||||
private getIncompleteTodos(content: string): string[] {
|
||||
const body = stripHeader(content);
|
||||
const match = body.match(/## Todo list\n([\s\S]*?)(?=\n## |$)/i);
|
||||
if(!match) return [];
|
||||
return match[1].split('\n')
|
||||
.map(line => line.match(/^\s*-\s*\[([ xX])\]\s+(.+?)\s*$/))
|
||||
.filter((m): m is RegExpMatchArray => !!m && m[1].toLowerCase() !== 'x')
|
||||
.map(m => m[2].trim());
|
||||
}
|
||||
|
||||
private listNodes(memories: Memory[]): MemoryRef[] {
|
||||
return memories.map(m => ({name: m.name, description: m.description}));
|
||||
}
|
||||
|
||||
private async mergeAgent(node: Memory, memories: Memory[] | MemoryCache, options: LLMRequest): Promise<Memory | null> {
|
||||
function factSimilarity(a: Memory, b: Memory): number {
|
||||
if(!a.bodyEmbeddings?.length || !b.bodyEmbeddings?.length) return 0;
|
||||
let best = 0;
|
||||
for(const av of a.bodyEmbeddings) {
|
||||
for(const bv of b.bodyEmbeddings) best = Math.max(best, 1 - cosineDistance(av, bv));
|
||||
}
|
||||
return best;
|
||||
}
|
||||
|
||||
if(!node.embedding?.length || node.name.startsWith('Journal/')) return null;
|
||||
const store = new MemoryAccessor(memories);
|
||||
const candidates = store.list
|
||||
.filter(m => m.name !== node.name && !m.name.startsWith('Journal/'))
|
||||
.filter(m => factSimilarity(node, m) >= FACT_SIMILARITY_THRESHOLD);
|
||||
|
||||
if(!candidates.length) return null;
|
||||
const closest = candidates.sort((a, b) => factSimilarity(node, b) - factSimilarity(node, a))[0];
|
||||
const result = await this.llm.ask('', {
|
||||
model: options.model,
|
||||
temperature: 0.3,
|
||||
schema: {
|
||||
aContent: {type: 'string', description: 'Updated document A body in markdown, without frontmatter.', required: true},
|
||||
bContent: {type: 'string', description: 'Updated document B body in markdown, without frontmatter.', required: true},
|
||||
},
|
||||
system: `Maintain these two persistent knowledge-base documents like a wiki.
|
||||
|
||||
Do NOT merge, rename, or delete either document. Both represent entities that should remain independently addressable.
|
||||
|
||||
The documents were selected because their facts may overlap. Your job is to reconcile duplicated information and connect the documents:
|
||||
- Decide which document is the HOME for each duplicated fact.
|
||||
- Keep the authoritative copy in that home document.
|
||||
- In the other document, replace the information with a short preamble and [[WikiLink]] to the home entity explaining the relationship.
|
||||
- If the documents are distinct entities but merely related, keep their distinct facts and add useful [[WikiLinks]] between them.
|
||||
- Do not delete useful entity-specific facts just because they are similar.
|
||||
- Do not invent relationships or facts.
|
||||
- Preserve useful history, technical specifics, structure, and existing [[WikiLinks]].
|
||||
- Most current truth wins when facts conflict.
|
||||
- Keep both documents concise and information-dense.
|
||||
- No frontmatter, preamble, filler, or AI commentary.
|
||||
|
||||
Document A ("${node.name}"):
|
||||
\`\`\`markdown
|
||||
${stripHeader(node.content)}
|
||||
\`\`\`
|
||||
|
||||
Document B ("${closest.name}"):
|
||||
\`\`\`markdown
|
||||
${stripHeader(closest.content)}
|
||||
\`\`\``,
|
||||
});
|
||||
const a = store.find(node.name);
|
||||
const b = store.find(closest.name);
|
||||
if(!a || !b || !result?.aContent || !result?.bContent) return null;
|
||||
a.content = this.touchHeader(a, result.aContent);
|
||||
b.content = this.touchHeader(b, result.bContent);
|
||||
await Promise.all([embedMemoryFields(a, this.llm), embedMemoryFields(b, this.llm)]);
|
||||
return a;
|
||||
}
|
||||
|
||||
private reconcile(node: Memory, memories: Memory[] | MemoryCache, options: LLMRequest): Promise<void> {
|
||||
const key = node.name;
|
||||
const existing = this.queues.get(key);
|
||||
if(existing) {
|
||||
existing.dirty = true;
|
||||
existing.request?.abort?.();
|
||||
return existing.task;
|
||||
}
|
||||
|
||||
const entry = {dirty: false, request: null, task: Promise.resolve()};
|
||||
this.queues.set(key, entry);
|
||||
const store = new MemoryAccessor(memories);
|
||||
entry.task = (async () => {
|
||||
let current = node;
|
||||
try {
|
||||
do {
|
||||
entry.dirty = false;
|
||||
await this.docAgent(current, store.list, options, entry);
|
||||
this.mergeLock = this.mergeLock.then(() => this.mergeAgent(current, memories, options));
|
||||
const result = await this.mergeLock;
|
||||
if(result) current = result;
|
||||
} while(entry.dirty);
|
||||
} finally {
|
||||
store.commit([node]);
|
||||
this.queues.delete(key);
|
||||
}
|
||||
})();
|
||||
return entry.task;
|
||||
}
|
||||
|
||||
private async docAgent(node: Memory, memories: Memory[], options: LLMRequest, entry: {request: {abort?: () => void} | null}): Promise<void> {
|
||||
if(!memories.includes(node)) return;
|
||||
const currentBody = stripHeader(node.content);
|
||||
const journal = node.name.startsWith('Journal/');
|
||||
const system = (journal
|
||||
? `You maintain one persistent journal document
|
||||
|
||||
Rewrite the ENTIRE journal, folding "## Pending" into the existing content removing the heading
|
||||
|
||||
Journal design:
|
||||
- Preserve the chronological daily log
|
||||
- Maintain a single \`## Todo list\` section for this entity: reconcile tasks semantically (merge equivalent tasks, remove duplicates, preserve incomplete tasks, check off completed ones), and keep it distinct from the narrative/fact sections
|
||||
- Group information by day under a date heading
|
||||
- Keep journal entries high level and concise: what was worked on and the outcome, not a step-by-step record of how — that detail lives in conversation history, not here
|
||||
- Use [[WikiLinks]] for persistent entities; don't turn ordinary journal events into entities
|
||||
- No frontmatter, preamble, filler, or AI commentary`
|
||||
: `You maintain one persistent knowledge-base entity document
|
||||
|
||||
Rewrite the ENTIRE document, folding "## Pending" into the existing content. Remove the Pending section when finished.
|
||||
|
||||
Document design:
|
||||
- The document represents one persistent entity. Keep information about that entity together and organized into sections
|
||||
- Merge any pending information in, newest fact wins conflicts; remove redundant content
|
||||
- Maintain a single \`## Todo list\` section for this entity: reconcile tasks semantically (merge equivalent tasks, remove duplicates, preserve incomplete tasks, check off completed ones), and keep it distinct from the narrative/fact sections
|
||||
- Let the structure fit the entity; there is NO fixed template
|
||||
- Add headings only when they meaningfully organize recurring information; don't create headings for one-off facts
|
||||
- Keep the document concise and information-dense without removing useful technical specifics
|
||||
- Current truth wins when facts conflict. Preserve older conflict as context, only when it adds useful meaning
|
||||
- No frontmatter, preamble, filler, or AI commentary`) + `
|
||||
|
||||
Available nodes to link to:
|
||||
${this.listNodes(memories).filter(n => n.name !== node.name).map(n => n.name).join(', ') || 'none'}
|
||||
|
||||
Current document:
|
||||
\`\`\`markdown
|
||||
${currentBody}
|
||||
\`\`\``;
|
||||
let update;
|
||||
try {
|
||||
for(let i = 0; i < 2 && !update?.content; i++) {
|
||||
const request = this.llm.ask(currentBody, {
|
||||
model: options.model,
|
||||
temperature: 0.3,
|
||||
schema: {
|
||||
description: {type: 'string', description: 'One factual sentence describing the document\'s ENTIRE SUBJECT MATTER — for use as a search/merge fingerprint', required: true},
|
||||
content: {type: 'string', description: 'Rewritten document body in markdown, without the frontmatter block', required: true},
|
||||
},
|
||||
system,
|
||||
});
|
||||
entry.request = request;
|
||||
update = await request;
|
||||
}
|
||||
} catch(err: any) {
|
||||
if(err?.name === 'AbortError') return;
|
||||
throw err;
|
||||
} finally {
|
||||
entry.request = null;
|
||||
}
|
||||
|
||||
if(!update?.content) return;
|
||||
node.description = node.name.startsWith('Journal/') ? this.journalDescription(node.name) : node.name !== 'People/User' ? update.description.replaceAll(/[\n:]/g, '') : 'All information about the current user';
|
||||
node.content = this.touchHeader(node, update.content);
|
||||
await embedMemoryFields(node, this.llm);
|
||||
}
|
||||
|
||||
private parseFrontmatter(content: string): {fm: Map<string, string>, body: string} {
|
||||
const match = content.match(/^---\n([\s\S]*?)\n---\n?([\s\S]*)$/);
|
||||
if(!match) return {fm: new Map(), body: content};
|
||||
const fm = new Map<string, string>();
|
||||
for(const line of match[1].split('\n')) {
|
||||
const i = line.indexOf(':');
|
||||
if(i === -1) continue;
|
||||
const key = line.slice(0, i).trim();
|
||||
const raw = line.slice(i + 1).trim();
|
||||
let value = raw;
|
||||
try { value = JSON.parse(raw); } catch { }
|
||||
fm.set(key, value);
|
||||
}
|
||||
return {fm, body: match[2]};
|
||||
}
|
||||
|
||||
private touchHeader(node: Memory, body: string): string {
|
||||
const {fm} = this.parseFrontmatter(node.content);
|
||||
fm.set('name', node.name);
|
||||
fm.set('description', (node.name.startsWith('Journal/') ? this.journalDescription(node.name) : node.description) || 'Persistent memory document');
|
||||
fm.set('modified', new Date().toISOString());
|
||||
return this.writeFrontmatter(fm, stripHeader(body));
|
||||
}
|
||||
|
||||
private writeFrontmatter(fm: Map<string, string>, body: string): string {
|
||||
const lines = [...fm.entries()].map(([k, v]) => `${k}: ${JSON.stringify(String(v).replace(/\s+/g, ' ').trim())}`);
|
||||
return `---\n${lines.join('\n')}\n---\n\n${body.trimStart()}`;
|
||||
}
|
||||
|
||||
decay() {
|
||||
for(const [name, ttl] of this.recentlyTouched) {
|
||||
if(ttl <= 1) this.recentlyTouched.delete(name);
|
||||
else this.recentlyTouched.set(name, ttl - 1);
|
||||
}
|
||||
}
|
||||
|
||||
touch(name: string, ttl = 2) {
|
||||
this.recentlyTouched.set(name, ttl);
|
||||
}
|
||||
|
||||
forget(name: string, memories: Memory[] | MemoryCache): boolean {
|
||||
return new MemoryAccessor(memories).forget(name);
|
||||
}
|
||||
|
||||
async recollect(query: string, memories: Memory[] | MemoryCache, limit = 5, graphDepth = 1): Promise<Memory[]> {
|
||||
function rank(query: number[], candidates: Memory[], limit: number): Memory[] {
|
||||
const scored = candidates.map(m => {
|
||||
const titleSim = m.titleEmbedding?.length ? 1 - cosineDistance(query, m.titleEmbedding) : 0;
|
||||
const descSim = m.embedding?.length ? 1 - cosineDistance(query, m.embedding) : 0;
|
||||
const bodySim = m.bodyEmbeddings?.length
|
||||
? Math.max(...m.bodyEmbeddings.map(b => 1 - cosineDistance(query, b)))
|
||||
: 0;
|
||||
return {memory: m, score: titleSim * 0.5 + descSim * 0.35 + bodySim * 0.15};
|
||||
});
|
||||
return scored.sort((a, b) => b.score - a.score).slice(0, limit).map(s => s.memory);
|
||||
}
|
||||
|
||||
const store = new MemoryAccessor(memories);
|
||||
if(!store.list.length) return [];
|
||||
await store.backfillEmbeddings(this.llm);
|
||||
|
||||
const [e] = await this.llm.embedding(query);
|
||||
if(!e) return [];
|
||||
|
||||
const pool = store.search(e.embedding, Math.max(limit * 3, limit));
|
||||
const poolMemories = pool.map(r => store.find(r.name)).filter((m): m is Memory => !!m);
|
||||
const ranked = rank(e.embedding, poolMemories, limit);
|
||||
const found = new Set<string>(ranked.map(m => m.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 rankedOrder = ranked.map(m => m.name);
|
||||
const graphExpansions = [...found].filter(n => !rankedOrder.includes(n));
|
||||
return [...rankedOrder, ...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()}_${Math.random().toString(36).slice(2)}`;
|
||||
const pending = {role: 'tool', name: 'memory_process', id: uid, content: conversation} as unknown as LLMMessage;
|
||||
history.push(pending);
|
||||
|
||||
const store = new MemoryAccessor(memories);
|
||||
const {buckets, journal, tasks} = await this.factAgent(conversation, store, options);
|
||||
const touched: Memory[] = [];
|
||||
|
||||
const personalTasks = tasks.filter(isPersonalTask);
|
||||
const entityTasks = tasks.filter(t => !isPersonalTask(t));
|
||||
|
||||
if(journal || personalTasks.length) {
|
||||
const journalName = `Journal/${this.getWeekStart()}`;
|
||||
let jnode = store.find(journalName);
|
||||
const isNew = !jnode;
|
||||
if(!jnode) {
|
||||
jnode = {
|
||||
name: journalName,
|
||||
description: this.journalDescription(),
|
||||
content: '',
|
||||
embedding: [],
|
||||
links: [],
|
||||
backlinks: [],
|
||||
};
|
||||
store.list.push(jnode);
|
||||
}
|
||||
|
||||
const blocks: string[] = [];
|
||||
if(journal) blocks.push(`### ${new Date().toISOString().slice(0, 10)}\n${journal}`);
|
||||
if(isNew) {
|
||||
const previousDate = new Date(`${this.getWeekStart()}T00:00:00Z`);
|
||||
previousDate.setUTCDate(previousDate.getUTCDate() - 7);
|
||||
const previous = store.find(`Journal/${previousDate.toISOString().slice(0, 10)}`);
|
||||
if(previous) {
|
||||
const todos = this.getIncompleteTodos(previous.content);
|
||||
if(todos.length) blocks.push(`${TODO_HEADING}\n${todos.map(task => `- [ ] ${task}`).join('\n')}`);
|
||||
}
|
||||
}
|
||||
if(personalTasks.length) blocks.push(`${TODO_HEADING}\n${personalTasks.map(task => `- [${task.done ? 'x' : ' '}] ${task.task}`).join('\n')}`);
|
||||
if(blocks.length) this.stage(jnode, blocks.join('\n\n'));
|
||||
touched.push(jnode);
|
||||
}
|
||||
|
||||
const entityStaging = new Map<string, {facts: string[], tasks: MemoryTask[]}>();
|
||||
for(const {subject, facts} of buckets) {
|
||||
const resolved = this.resolveSubject(subject, store);
|
||||
const entry = entityStaging.get(resolved) ?? {facts: [], tasks: []};
|
||||
entry.facts.push(...facts);
|
||||
entityStaging.set(resolved, entry);
|
||||
}
|
||||
for(const task of entityTasks) {
|
||||
const resolved = this.resolveSubject(task.subject, store);
|
||||
const entry = entityStaging.get(resolved) ?? {facts: [], tasks: []};
|
||||
entry.tasks.push(task);
|
||||
entityStaging.set(resolved, entry);
|
||||
}
|
||||
|
||||
for(const [resolved, {facts, tasks: subjectTasks}] of entityStaging) {
|
||||
let node = store.find(resolved);
|
||||
if(!node) {
|
||||
node = {name: resolved, description: 'Persistent memory document', content: '', embedding: [], links: [], backlinks: []};
|
||||
store.list.push(node);
|
||||
}
|
||||
const blocks: string[] = [];
|
||||
if(facts.length) blocks.push(facts.map(f => `- ${f}`).join('\n'));
|
||||
if(subjectTasks.length) blocks.push(`${TODO_HEADING}\n${subjectTasks.map(t => `- [${t.done ? 'x' : ' '}] ${t.task}`).join('\n')}`);
|
||||
if(blocks.length) this.stage(node, blocks.join('\n\n'));
|
||||
touched.push(node);
|
||||
}
|
||||
|
||||
await Promise.all(touched.map(async node => {
|
||||
await embedMemoryFields(node, this.llm);
|
||||
this.touch(node.name);
|
||||
}));
|
||||
|
||||
if(touched.length) {
|
||||
store.commit(touched);
|
||||
(pending as any).content = `Saved to ${touched.map(n => `[[${n.name}]]`).join(', ')}`;
|
||||
Promise.all(touched.map(node => this.reconcile(node, memories, options).catch(() => {})));
|
||||
} else {
|
||||
(pending as any).content = 'Nothing worth remembering.';
|
||||
}
|
||||
|
||||
(touched as any).uid = uid;
|
||||
return touched;
|
||||
}
|
||||
|
||||
async reconcileAll(memories: Memory[] | MemoryCache, options: LLMRequest, scope: 'touched' | 'all' = 'touched'): Promise<void> {
|
||||
const store = new MemoryAccessor(memories);
|
||||
const targets = scope === 'all' ? store.list : store.list.filter(m => m.content.includes(PENDING_HEADING));
|
||||
await Promise.all(targets.map(node => this.reconcile(node, memories, options)));
|
||||
store.commit();
|
||||
}
|
||||
}
|
||||
|
||||
@@ -0,0 +1,111 @@
|
||||
import {memoryStore} from '../memory.ts';
|
||||
import {Memory, MemoryStore} from './memory.ts';
|
||||
import {LLMRequest} from '../llm.ts';
|
||||
import {embedMemoryFields, journalDescription, stripHeader, touchHeader} from '../utils.ts';
|
||||
|
||||
export const PENDING_HEADING = '## Pending';
|
||||
export const TODO_HEADING = '## Todo List';
|
||||
|
||||
export async function docAgent(llm: any, node: Memory, memories: MemoryStore, options: LLMRequest, entry?: {request: {abort?: () => void} | null},): Promise<void> {
|
||||
const store = memoryStore(memories);
|
||||
if(!store.list.includes(node)) return;
|
||||
|
||||
const currentBody = stripHeader(node.content);
|
||||
const journal = node.name.startsWith('Journal/');
|
||||
|
||||
const system = `You maintain a Markdown ${journal ? 'journal' : 'wiki file'}, produce the final document body
|
||||
|
||||
- Incorporate the \`${PENDING_HEADING}\` into the wiki & remove it. These new facts trump conflicting information
|
||||
- Preserve existing structure, wording and context unless directly affected by \`${PENDING_HEADING}\` information
|
||||
- Only restructure if clearly unorganized, or there is duplicate, stale or misleading information. Do not restructure merely because you dont like it
|
||||
- Let the structure fit the document; do not force a template
|
||||
- Use headings, subheadings, lists, tables, code blocks and other formatting where useful
|
||||
- Create [[WikiLinks]] to existing pages or ghost nodes when the relationship is meaningful but does not exist yet
|
||||
- Remove duplication and obsolete information, preserve useful detail
|
||||
- Do not invent facts or make unnecessary changes
|
||||
- Never add edit commentary` + (journal ? `
|
||||
- This is a JOURNAL document, focus on creating a chronological report of events
|
||||
- Keep events chronological and concise. Do not reorganize events into
|
||||
- subject-based wiki sections. Preserve dates and useful temporal context.
|
||||
|
||||
Rough Template:
|
||||
\`\`\`markdown
|
||||
# Journal
|
||||
|
||||
## ${TODO_HEADING}
|
||||
- [ ] ...
|
||||
|
||||
### YYYY-MM-DD
|
||||
- ...
|
||||
\`\`\`` : `
|
||||
- Use only sections that have content and make sense
|
||||
- Keep related facts together
|
||||
- historical context may be preserved
|
||||
- Use ## Related for meaningful links to other wiki pages.
|
||||
|
||||
Rough Template:
|
||||
# Title
|
||||
|
||||
## ${TODO_HEADING}
|
||||
- [ ] ...
|
||||
|
||||
## Overview
|
||||
...
|
||||
|
||||
## Heading
|
||||
...
|
||||
|
||||
### Subheading
|
||||
...
|
||||
|
||||
## Related
|
||||
- [[wikiLinks]]: relationship description
|
||||
`) + `
|
||||
|
||||
Available pages for wikilinks:
|
||||
${store.list.filter(n => n.name !== node.name).map(n => n.name).join(', ') || 'none'}`;
|
||||
|
||||
let update;
|
||||
|
||||
try {
|
||||
for(let i = 0; i < 2 && !update?.content; i++) {
|
||||
const request = llm.ask(currentBody, {
|
||||
model: options.model,
|
||||
temperature: 0.3,
|
||||
schema: {
|
||||
description: {
|
||||
type: 'string',
|
||||
description: 'One factual sentence describing the document\'s ENTIRE SUBJECT MATTER — for use as a search/merge fingerprint',
|
||||
required: true,
|
||||
},
|
||||
content: {
|
||||
type: 'string',
|
||||
description: 'Rewritten document body in markdown, without the frontmatter block',
|
||||
required: true,
|
||||
},
|
||||
},
|
||||
system,
|
||||
});
|
||||
|
||||
if(entry) entry.request = request;
|
||||
update = await request;
|
||||
}
|
||||
} catch(err: any) {
|
||||
if(err?.name === 'AbortError') return;
|
||||
throw err;
|
||||
} finally {
|
||||
if(entry) entry.request = null;
|
||||
}
|
||||
|
||||
if(!update?.content) return;
|
||||
|
||||
node.description = node.name.startsWith('Journal/')
|
||||
? journalDescription(node.name)
|
||||
: node.name !== 'People/User'
|
||||
? update.description.replaceAll(/[\n:]/g, '')
|
||||
: 'All information about the current user';
|
||||
|
||||
node.content = touchHeader(node, update.content);
|
||||
|
||||
await embedMemoryFields(node, llm);
|
||||
}
|
||||
@@ -0,0 +1,154 @@
|
||||
import {MemoryCache, memoryStore} from '../memory.ts';
|
||||
import {LLMRequest} from '../llm.ts';
|
||||
import {Memory, MemoryStore} from './memory.ts';
|
||||
|
||||
const ALIAS_MATCH_THRESHOLD = 0.55;
|
||||
|
||||
export type FactBucket = {
|
||||
subject: string;
|
||||
facts: string[];
|
||||
}
|
||||
|
||||
export type MemoryTask = {
|
||||
subject: string;
|
||||
task: string;
|
||||
done: boolean;
|
||||
}
|
||||
|
||||
export type FactAgentResult = {
|
||||
buckets: FactBucket[];
|
||||
journal: string;
|
||||
tasks: MemoryTask[];
|
||||
}
|
||||
|
||||
function dedupeFacts(facts: string[]): string[] {
|
||||
const seen = new Map<string, string>();
|
||||
for(const f of facts) {
|
||||
const clean = f.trim();
|
||||
if(clean) seen.set(clean.toLowerCase(), clean);
|
||||
}
|
||||
return [...seen.values()];
|
||||
}
|
||||
|
||||
export function resolveSubject(subject: string, memories: Memory[] | MemoryCache, llm: any,): string {
|
||||
function normalize(name: string): string {
|
||||
return name.trim().toLowerCase().replace(/\s+/g, ' ');
|
||||
}
|
||||
|
||||
const list = memories instanceof MemoryCache ? memories.memories : memories;
|
||||
|
||||
const trimmed = subject.trim();
|
||||
const exact = list.find(m => m.name === trimmed);
|
||||
if(exact) return exact.name;
|
||||
|
||||
const normalized = normalize(trimmed);
|
||||
const caseInsensitive = list.find(m => normalize(m.name) === normalized);
|
||||
if(caseInsensitive) return caseInsensitive.name;
|
||||
|
||||
const root = trimmed.split('/')[0];
|
||||
const leaf = trimmed.split('/').slice(1).join('/') || trimmed;
|
||||
const candidates = list.filter(m => m.name.split('/')[0] === root && m.name !== trimmed);
|
||||
if(!candidates.length) return trimmed;
|
||||
|
||||
const leaves = candidates.map(m => m.name.split('/').slice(1).join('/') || m.name);
|
||||
const probe = leaves.length > 1 ? leaves : [...leaves, ''];
|
||||
const {max, similarities} = llm.fuzzyMatch(leaf, ...probe);
|
||||
|
||||
if(max >= ALIAS_MATCH_THRESHOLD) return candidates[similarities.indexOf(max)].name;
|
||||
|
||||
return trimmed;
|
||||
}
|
||||
|
||||
export async function factAgent(llm: any, conversation: string, memories: MemoryStore, options: LLMRequest,): Promise<FactAgentResult> {
|
||||
const store = memoryStore(memories);
|
||||
const ghosts = store.ghosts();
|
||||
|
||||
const response = await llm.ask(conversation, {
|
||||
model: options.model,
|
||||
temperature: 0.2,
|
||||
system: `Turn this conversation into a persistent memory file by extracting information into organized bullet points
|
||||
|
||||
Think of this like an Obsidian vault with a clear division of responsibility:
|
||||
- The JOURNAL is a chronological log. It answers "what happened, and when" and is the only place with a sense of time.
|
||||
- ENTITY DOSSIERS are a wiki pages. They answer "what is currently true about this subject", with no sense of time — only current state.
|
||||
- Never blur the two: a one-off event, conversation, or debugging session is a journal entry, not an entity, even if it's detailed.
|
||||
- Never extract this assistant's own tools, capabilities, or system behavior — that's not memory, it's spec.
|
||||
- Skip greetings, pleasantries, small talk and generic exchanges entirely.
|
||||
|
||||
1. Journal Log
|
||||
- A chronological, skimmable log of what actually happened: high level discussions, decisions made (including ones reached jointly with the assistant), progress on projects, problems worked through
|
||||
- This is NOT a transcript, and it is NOT a step-by-step record, its a compressed day-to-day log of notable events & developments
|
||||
- One line per development is usually enough: what was worked on and the outcome, not the blow-by-blow of how
|
||||
- Skip anything that's a todo item (goes in Todo Tasks) or a durable fact about a subject (goes in Entity Dossiers)
|
||||
|
||||
2. Todo Tasks
|
||||
- Extract concrete tasks the user says need to be done, should be done, or were completed
|
||||
- Return the task text and whether it is still todo or is done
|
||||
- A completed task should be marked done, not recreated as a new todo
|
||||
- Only extract actionable tasks, not general goals or observations, if none - omit returning a tasks array
|
||||
- Assign each task a HOME ENTITY subject using the same rules as Entity Dossiers below, or an empty string if it belongs in the journal: personal/life task (reach out to someone, pay a bill, etc.)
|
||||
|
||||
3. Entity Dossiers
|
||||
- Detailed dossiers with all factual information regarding a subject
|
||||
- Only extract facts the USER explicitly stated about themselves, their work, projects, or decisions made during this conversation — not assistant claims, guesses, or temporary/debugging details
|
||||
- Record the final/end state, not intermediate deltas
|
||||
- NEVER create a dossier for something I wouldn't find on a wiki: temporary info, debugging, guesses, conversation fragments (that's journal material)
|
||||
|
||||
Entity Dossier Routing — every fact belongs to exactly one HOME ENTITY: the [pro]noun that OWNS the fact, not the initializer
|
||||
- All facts primarily about the user go under People/User
|
||||
- Path format is always Collection/Subject (People/Sarah, Projects/Oxide) — never a bare/rootless name
|
||||
- ALWAYS prefer an existing node (including ghosts) over creating a new one, even under a different alias — route the facts there, and record the unused name as a ghost node
|
||||
- Only use child paths (Collection/Subject/Aspect) when a clear child/parent relationship exists between entities
|
||||
- Use [[WikiLinks]] to express relationships between entities. Ghost links are supported so NEVER create a document just to hold a relationship
|
||||
|
||||
Example Entity Naming Convention:
|
||||
- Projects/[Name]
|
||||
- People/[Name]
|
||||
- History/[Name]
|
||||
- Science/[Name]
|
||||
- [Subject]/[Name]
|
||||
- Class/[Name]/[Chapter]
|
||||
|
||||
Available nodes:
|
||||
${store.list.map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}
|
||||
${ghosts.length ? `${ghosts.map(g => `- ${g}: (Ghost)`).join('\n')}` : ''}`,
|
||||
schema: {
|
||||
journal: {
|
||||
type: 'string',
|
||||
description: 'Short bullet point recap, omit if nothing notable happened',
|
||||
},
|
||||
tasks: {
|
||||
type: 'array',
|
||||
description: 'Concrete tasks mentioned or completed in the conversation, omit if none',
|
||||
items: {type: 'object', items: {
|
||||
subject: {type: 'string', description: 'Exact node name / new persistent entity path this task belongs to, or an empty string if this is a personal task with no entity of its own (those go in the journal)', required: true,},
|
||||
task: {type: 'string', description: 'Concise actionable task', required: true,},
|
||||
done: {type: 'boolean', description: 'Whether the task is completed', required: true,},
|
||||
}},
|
||||
},
|
||||
buckets: {
|
||||
type: 'array',
|
||||
description: 'Groups of facts to remember; omit if none',
|
||||
items: {type: 'object', items: {
|
||||
subject: {type: 'string', description: 'Exact node name or new persistent entity path', required: true,},
|
||||
facts: {type: 'array', description: 'Facts to store here', items: {type: 'string'},},
|
||||
}},
|
||||
},
|
||||
},
|
||||
});
|
||||
|
||||
const buckets = new Map<string, string[]>();
|
||||
|
||||
for(const bucket of response.buckets ?? []) {
|
||||
const subject = bucket.subject.trim();
|
||||
const facts = buckets.get(subject) ?? [];
|
||||
facts.push(...dedupeFacts(bucket.facts));
|
||||
buckets.set(subject, facts);
|
||||
}
|
||||
|
||||
return {
|
||||
buckets: buckets.entries().toArray().map(([subject, facts]) => ({subject, facts})),
|
||||
journal: (response.journal ?? '').trim(),
|
||||
tasks: response.tasks ?? [],
|
||||
};
|
||||
}
|
||||
@@ -1,3 +1,5 @@
|
||||
import {cosineDistance, euclideanDistance} from '../utils.ts';
|
||||
|
||||
export type DistanceMetric = "euclidean" | "cosine";
|
||||
|
||||
export interface KDPoint<T = unknown> {
|
||||
@@ -18,28 +20,6 @@ interface KDNode<T> {
|
||||
deleted?: boolean;
|
||||
}
|
||||
|
||||
// ─── Distance helpers ─────────────────────────────────────────────────────────
|
||||
|
||||
function euclidean(a: number[], b: number[]): number {
|
||||
let sum = 0;
|
||||
for (let i = 0; i < a.length; i++) {
|
||||
const d = a[i] - b[i];
|
||||
sum += d * d;
|
||||
}
|
||||
return Math.sqrt(sum);
|
||||
}
|
||||
|
||||
function cosine(a: number[], b: number[]): number {
|
||||
let dot = 0, normA = 0, normB = 0;
|
||||
for (let i = 0; i < a.length; i++) {
|
||||
dot += a[i] * b[i];
|
||||
normA += a[i] * a[i];
|
||||
normB += b[i] * b[i];
|
||||
}
|
||||
const denom = Math.sqrt(normA) * Math.sqrt(normB);
|
||||
return denom === 0 ? 1 : 1 - dot / denom; // distance = 1 - similarity
|
||||
}
|
||||
|
||||
/**
|
||||
* Keeps the k closest candidates in memory, evicts the furthest when full
|
||||
*/
|
||||
@@ -123,7 +103,7 @@ export class KDTree<T = unknown> {
|
||||
points?: KDPoint<T>[]
|
||||
) {
|
||||
this.dims = dims;
|
||||
this.distanceFn = metric === "cosine" ? cosine : euclidean;
|
||||
this.distanceFn = metric === "cosine" ? cosineDistance : euclideanDistance;
|
||||
|
||||
if (points && points.length > 0) {
|
||||
this.validateAll(points);
|
||||
@@ -300,11 +280,8 @@ export class KDTree<T = unknown> {
|
||||
: [node.right, node.left];
|
||||
|
||||
this.searchKNN(near, query, k, heap, depth + 1);
|
||||
|
||||
// Only explore the far side if it could contain a closer point.
|
||||
// For cosine distance we can't prune by axis gap alone, so always explore.
|
||||
const shouldExplore =
|
||||
this.distanceFn === cosine
|
||||
this.distanceFn === cosineDistance
|
||||
? true
|
||||
: Math.abs(diff) < heap.worstDistance;
|
||||
|
||||
@@ -340,7 +317,7 @@ export class KDTree<T = unknown> {
|
||||
this.searchRadius(near, query, radius, results, depth + 1);
|
||||
|
||||
const shouldExplore =
|
||||
this.distanceFn === cosine ? true : Math.abs(diff) <= radius;
|
||||
this.distanceFn === cosineDistance ? true : Math.abs(diff) <= radius;
|
||||
|
||||
if (shouldExplore) {
|
||||
this.searchRadius(far, query, radius, results, depth + 1);
|
||||
@@ -0,0 +1,79 @@
|
||||
import {memoryStore} from '../memory.ts';
|
||||
import {Memory, MemoryStore} from './memory.ts';
|
||||
import {LLMRequest} from '../llm.ts';
|
||||
|
||||
export type LibrarianAction =
|
||||
| {type: 'merge'; source: string; target: string; instructions?: string}
|
||||
| {type: 'split'; target: string; documents: {name: string; instructions?: string}[]}
|
||||
| {type: 'move'; source: string; target: string}
|
||||
| {type: 'delete'; target: string}
|
||||
| {type: 'link'; source: string; target: string};
|
||||
|
||||
export type LibrarianResult = {
|
||||
actions: LibrarianAction[];
|
||||
};
|
||||
|
||||
export async function librarianAgent(llm: any, node: Memory, memories: MemoryStore, options: LLMRequest, limit = 8,): Promise<LibrarianResult> {
|
||||
if(node.name.startsWith('Journal/')) return {actions: []};
|
||||
const store = memoryStore(memories);
|
||||
const candidates = store.search(node.embedding, limit + 1)
|
||||
.filter(result => result.name !== node.name && !result.name.startsWith('Journal/'))
|
||||
.map(result => store.find(result.name))
|
||||
.filter((memory): memory is Memory => !!memory);
|
||||
|
||||
if(!candidates.length) return {actions: []};
|
||||
const documents = [node, ...candidates];
|
||||
const response = await llm.ask('', {
|
||||
model: options.model,
|
||||
temperature: 0.2,
|
||||
schema: {
|
||||
actions: {
|
||||
type: 'array',
|
||||
description: 'Actions needed to improve the organization of the knowledge base. Omit if no changes are needed.',
|
||||
items: {type: 'object', items: {
|
||||
type: {type: 'string', description: 'One of: merge, split, move, delete, link', required: true,},
|
||||
source: {type: 'string', description: 'Source document name. Required for merge, move, and link.',},
|
||||
target: {type: 'string', description: 'Target document name. Required for merge, move, delete, and link.',},
|
||||
instructions: {type: 'string', description: 'Specific instructions for performing the action.',},
|
||||
documents: {type: 'array', description: 'Documents to create when splitting a document.', items: {type: 'object', items: {
|
||||
name: {type: 'string', required: true},
|
||||
instructions: {type: 'string'},
|
||||
}}}},
|
||||
},
|
||||
},
|
||||
},
|
||||
system: `You are a librarian of a wiki. You keep it organized, coherent, and easy to navigate. You DONT rewrite pages but rather, recommend actions to organize the wiki
|
||||
|
||||
Available actions:
|
||||
- merge: Two pages represent the same entity and should be merged together
|
||||
- split: One page contains significant information on two distinct entities and should become separate pages
|
||||
- move: A document belongs under a different name or path for consistancy
|
||||
- delete: A document is obsolete, accidental or contains nothing redeamable
|
||||
- link: Two distinct entities are meaningfully related and should reference each other
|
||||
|
||||
Rules:
|
||||
- Only act on high confidence
|
||||
- Semantic similarity alone is NOT sufficient reason to merge two distinct entities
|
||||
- Prefer preserving existing entities and paths when possible
|
||||
- ALWAYS use & preserve [[wikilinks]] (including ghost nodes and journal entries) to represent relationships
|
||||
- A document should represent one persistent entity, not a temporal state, feature, bug, event, decision, or conversation
|
||||
- When merging, use the more common name as the destination and use ghost nodes to reference old aliases
|
||||
- When splitting, each resulting document must represent a distinct persistent entity
|
||||
- Return an empty actions array when the documents are already organized correctly
|
||||
|
||||
Documents being reviewed:
|
||||
|
||||
${documents.map(memory => `### ${memory.name}
|
||||
Description: ${memory.description}
|
||||
|
||||
Links: ${[...memory.links, ...memory.backlinks].join(', ') || 'none'}
|
||||
|
||||
\`\`\`markdown
|
||||
${memory.content}
|
||||
\`\`\``).join('\n\n')}`,
|
||||
});
|
||||
|
||||
return {
|
||||
actions: response.actions ?? [],
|
||||
};
|
||||
}
|
||||
@@ -0,0 +1,28 @@
|
||||
import {MemoryCache} from '../memory.ts';
|
||||
|
||||
export type Memory = {
|
||||
name: string;
|
||||
description: string;
|
||||
content: string;
|
||||
embedding: number[];
|
||||
titleEmbedding?: number[];
|
||||
bodyEmbeddings?: number[][];
|
||||
links: string[];
|
||||
backlinks: string[];
|
||||
}
|
||||
|
||||
export type MemoryRef = {
|
||||
name: string;
|
||||
description: string;
|
||||
distance?: number;
|
||||
}
|
||||
|
||||
export type MemoryOptions = {
|
||||
memory: Memory[] | MemoryCache;
|
||||
inject?: boolean;
|
||||
tool?: boolean;
|
||||
update?: boolean;
|
||||
maxTokens?: number;
|
||||
}
|
||||
|
||||
export type MemoryStore = Memory[] | MemoryCache;
|
||||
@@ -0,0 +1,82 @@
|
||||
import {Memory} from './memory/memory.ts';
|
||||
|
||||
export function cosineDistance(a: number[], b: number[]): number {
|
||||
let dot = 0, normA = 0, normB = 0;
|
||||
for(let i = 0; i < a.length; i++) {
|
||||
dot += a[i] * b[i];
|
||||
normA += a[i] * a[i];
|
||||
normB += b[i] * b[i];
|
||||
}
|
||||
const denom = Math.sqrt(normA) * Math.sqrt(normB);
|
||||
return denom === 0 ? 1 : 1 - dot / denom;
|
||||
}
|
||||
|
||||
export async function embedMemoryFields(node: Memory, llm: any): Promise<void> {
|
||||
const body = stripHeader(node.content);
|
||||
const [titleE] = await llm.embedding(node.name.split('/').pop() || node.name);
|
||||
const [descE] = await llm.embedding(node.description || '');
|
||||
const bodyChunks = body ? await llm.embedding(body) : [];
|
||||
if(titleE) node.titleEmbedding = titleE.embedding;
|
||||
if(descE) node.embedding = descE.embedding;
|
||||
node.bodyEmbeddings = bodyChunks.map((c: any) => c.embedding).filter(Boolean);
|
||||
}
|
||||
|
||||
export function euclideanDistance(a: number[], b: number[]): number {
|
||||
let sum = 0;
|
||||
for(let i = 0; i < a.length; i++) {
|
||||
const d = a[i] - b[i];
|
||||
sum += d * d;
|
||||
}
|
||||
return Math.sqrt(sum);
|
||||
}
|
||||
|
||||
function getWeekStart(date: Date = new Date()): string {
|
||||
const d = new Date(Date.UTC(date.getFullYear(), date.getMonth(), date.getDate()));
|
||||
const day = d.getUTCDay();
|
||||
const diff = day === 0 ? -6 : 1 - day;
|
||||
d.setUTCDate(d.getUTCDate() + diff);
|
||||
return d.toISOString().slice(0, 10);
|
||||
}
|
||||
|
||||
export function journalDescription(journalName?: string): string {
|
||||
const start = journalName?.split('/').pop() || getWeekStart();
|
||||
const d = new Date(`${start}T00:00:00Z`);
|
||||
d.setUTCDate(d.getUTCDate() + 6);
|
||||
const end = d.toISOString().slice(0, 10);
|
||||
return `Log from ${start} - ${end}`;
|
||||
}
|
||||
|
||||
function parseFrontmatter(content: string): {fm: Map<string, string>, body: string} {
|
||||
const match = content.match(/^---\n([\s\S]*?)\n---\n?([\s\S]*)$/);
|
||||
if(!match) return {fm: new Map(), body: content};
|
||||
const fm = new Map<string, string>();
|
||||
for(const line of match[1].split('\n')) {
|
||||
const i = line.indexOf(':');
|
||||
if(i === -1) continue;
|
||||
const key = line.slice(0, i).trim();
|
||||
const raw = line.slice(i + 1).trim();
|
||||
let value = raw;
|
||||
try { value = JSON.parse(raw); } catch { }
|
||||
fm.set(key, value);
|
||||
}
|
||||
return {fm, body: match[2]};
|
||||
}
|
||||
|
||||
export function writeFrontmatter(fm: Map<string, string>, body: string): string {
|
||||
const lines = [...fm.entries()].map(([k, v]) =>
|
||||
`${k}: ${JSON.stringify(String(v).replace(/\s+/g, ' ').trim())}`);
|
||||
return `---\n${lines.join('\n')}\n---\n\n${body.trimStart()}`;
|
||||
}
|
||||
|
||||
export function stripHeader(content: string): string {
|
||||
return content.replace(/^---[\s\S]*?\n---\n?/, '').trimStart();
|
||||
}
|
||||
|
||||
export function touchHeader(node: Memory, body: string): string {
|
||||
const {fm} = parseFrontmatter(node.content);
|
||||
fm.set('name', node.name);
|
||||
fm.set('description', (node.name.startsWith('Journal/') ? journalDescription(node.name) : node.description)
|
||||
|| 'Persistent memory document');
|
||||
fm.set('modified', new Date().toISOString());
|
||||
return writeFrontmatter(fm, stripHeader(body));
|
||||
}
|
||||
Reference in New Issue
Block a user