tanstack-ai-memory

TanStack AI Memory Middleware

Use this when adding server-side memory to a chat() call. Everything lives in @tanstack/ai-memory. A memory adapter is a single contract with two verbs — recall and save — and the middleware is thin: it recalls into the system prompt before the model runs and defers save after the turn finishes.

When to reach for it

• A user expects "remember what I told you last time."
• Per-user or per-thread context that must survive across sessions.
• A hosted memory service (mem0, Honcho, Hindsight).

Do NOT use this just to keep recent messages — that's the messages array on chat(). Memory is for cross-turn / cross-session recall, not within-turn history.

Wire it up

import { chat } from '@tanstack/ai'
import { openaiText } from '@tanstack/ai-openai'
import { memoryMiddleware } from '@tanstack/ai-memory'
import { inMemory } from '@tanstack/ai-memory/in-memory'

const memory = inMemory() // dev/tests only — see the in-memory skill

const stream = chat({
  adapter: openaiText('gpt-5.5'),
  messages,
  context: { session }, // attached by your auth middleware
  middleware: [
    memoryMiddleware({
      adapter: memory,
      // Derive scope server-side from trusted session state.
      scope: (ctx) => {
        const session = getSession(ctx)
        return { threadId: session.threadId, userId: session.userId }
      },
    }),
  ],
})

memoryMiddleware options: adapter, scope (static or a function of ctx), role ('recall+save' default, or 'save-only'), and onRecall / onSave telemetry callbacks.

The contract

interface MemoryAdapter {
  id: string
  recall(scope, query): Promise<RecallResult> // { systemPrompt, fragments?, tools?, toolGuidance? }
  save(scope, turn): Promise<Array<SaveReceipt>> // turn = { user, assistant }; extraction lives HERE
  inspect?(scope): Promise<MemorySnapshot> // optional (devtools)
  listFacts?(scope): Promise<Array<MemoryFact>> // optional (devtools)
}