Memory

Memory gives an agent persistent conversation threads, semantic recall of past messages, and working memory — all behind a pluggable storage layer.

Setting up memory

memory.ts

import { Memory } from '@redtuma/memory'
import { LibSQLStore } from '@redtuma/store-libsql'

const memory = new Memory({
  storage: new LibSQLStore({ url: 'file:./redtuma.db' }),
  options: {
    lastMessages: 20,                  // recent history window
    semanticRecall: { topK: 5 },       // retrieve relevant past messages
    workingMemory: { enabled: true },  // persistent per-user notes
  },
})

const agent = new Agent({
  id: 'assistant',
  instructions: '...',
  model: 'anthropic/claude-opus-4-8',
  memory,
})

Threads & resources

Scope a conversation with a thread (one conversation) and a resource (the user or entity it belongs to). Pass them per call:

await agent.generate('Remember my name is Sam.', {
  memory: { thread: 'thread-1', resource: 'user-42' },
})

// later, in the same thread, the agent recalls context
await agent.generate('What is my name?', {
  memory: { thread: 'thread-1', resource: 'user-42' },
})

Stores

All stores implement the same Store interface, so you can swap them freely:

InMemoryStore — built into @redtuma/core, great for tests and quick starts
@redtuma/store-libsql — local SQLite file or Turso
@redtuma/store-pg — Postgres
@redtuma/store-do — Cloudflare Durable Object (see Cloudflare)

Same contract everywhere

Every store passes the same behavioral conformance suite, so memory behaves identically whether it's SQLite on your laptop or a Durable Object at the edge.

← Workflows Model routing →