@weaveintel/memory

Gives an AI agent a memory — the running conversation, durable facts, the people and things it has met, and how they connect.

Why it exists

A fresh AI call is an amnesiac: it knows only what you paste into this one prompt. Real assistants need to remember — that you already introduced yourself, that "the Munich office" is the same place you mentioned last week, that a correction you made should stick. Think of it as the difference between a stranger and a colleague: the colleague carries context forward. This package is that carried-forward context — short-term chat history, long-term facts, the entities involved, and a small knowledge graph of how they relate (it absorbed the old standalone graph package, so entities and relationships live here too).

When to reach for it

Reach for it whenever an agent should recall anything across turns or sessions: conversation history, learned facts, or a graph of people/orgs/topics. Pick a backend to match your scale — in-memory for tests, SQLite/Postgres/pgvector/Redis/Mongo/DynamoDB for production. If all you need is to look up passages from a document corpus at query time (RAG over files), use @weaveintel/retrieval instead; memory is about what the agent remembers, not what it can fetch on demand.

How to use it

import { weaveConversationMemory } from '@weaveintel/memory';
import { weaveContext } from '@weaveintel/core';

const memory = weaveConversationMemory({ maxHistory: 50 });
const ctx = weaveContext();

await memory.addMessage(ctx, { role: 'user', content: 'I work in Munich.' });
await memory.addMessage(ctx, { role: 'assistant', content: 'Noted!' });

const recent = await memory.getMessages(ctx, 10); // last 10 turns, ready to re-prompt

What's in the box

• Conversation — weaveConversationMemory, createConfiguredConversationMemory.
• Stores (pick a backend) — weaveMemoryStore, weaveRuntimeMemoryStore, weaveSqliteMemoryStore, weavePostgresMemoryStore, weavePgVectorMemoryStore, weaveRedisMemoryStore, weaveMongoDbMemoryStore, weaveCloudNoSqlMemoryStore, plus createConfiguredMemoryStore / …Async.
  • Sharing one Postgres pool: the Postgres stores take either a connection string or an existing pool — weavePostgresMemoryStore({ url }) or weavePostgresMemoryStore({ pool }) (same for weavePgVectorMemoryStore). Pass a shared pool (e.g. from weaveSharedPostgres in @weaveintel/persistence) so your whole app runs on one connection instead of each store opening its own. When you inject a pool, the store leaves it open on close() — you own its lifecycle.
  • One implementation for SQLite + Postgres: the plain weaveSqliteMemoryStore and weavePostgresMemoryStore now share a single query implementation (built with Drizzle), so the two databases can't drift apart. memoryStoreContract runs the same battery against both to prove it. (The pgvector store is separate — it's Postgres-only by nature.)
• Semantic & entity — weaveSemanticMemory, weaveEntityMemory, fusedMemorySearch (recency × importance × relevance fusion).
• Working memory & context — weaveWorkingMemory, createContextAssembler, compressor registry.
• Lifecycle — deduplicateExact/deduplicateByKey, recordCorrection/supersede, enforceRetention/forgetUser/forgetSession, weaveMemoryConsolidator.
• Extraction & procedural — runHybridMemoryExtraction, extractEntitiesByRegexRules, createProceduralEntry/runProceduralCurator.
• Knowledge graph — createGraphMemoryStore, createEntityNode, createRelationshipEdge, createEntityLinker, createTimelineGraph, createGraphRetriever.
• Governance & provenance — weaveGovernancePolicy, setProvenance/filterByConfidence.

License

MIT.