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, toServerSentEventsResponse } from '@tanstack/ai'
import { openaiText } from '@tanstack/ai-openai'
import { memoryMiddleware } from '@tanstack/ai-memory'
import { inMemory } from '@tanstack/ai-memory/in-memory'
import { requireSession } from './auth'

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

export async function POST(request: Request) {
  const { messages } = await request.json()
  // Resolved by your auth layer from cookies/headers — never from the request body.
  const session = await requireSession(request)

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

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

The contract

import type {
  MemoryFact,
  MemoryScope,
  MemorySnapshot,
  MemoryTurn,
  RecallResult,
  SaveReceipt,
} from '@tanstack/ai-memory'

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

• recall decides relevance and renders a systemPrompt; it may also return tools + toolGuidance to hand the model direct control of memory (hindsight does this).
• save owns extraction — turning the raw turn into whatever gets persisted.

Scope security

MemoryScope is an alias of the shared Scope type from @tanstack/ai: { threadId, userId?, tenantId?, namespace? }. It is the isolation boundary. Never trust a client-supplied userId/threadId. Resolve scope server-side from session/auth and pass the validated session through chat({ context: { session } }). If you accept a thread id from the request body, validate it belongs to the session user BEFORE using it.

Adapters

• inMemory() / redis() — exact match on threadId + optional userId/tenantId (namespace ignored). Redis index keys include all three segments.
• hindsight() — bank {tenant|_}__{user}__{threadId}.
• mem0() — user_id + run_id (threadId); no tenantId.
• honcho() — session {tenant|_}__{threadId}; peer tenant-prefixed when set.
• Custom — implement recall/save and run runMemoryAdapterContract from @tanstack/ai-memory/testkit.

Failure modes

Memory failures are non-fatal: a throwing recall or save emits memory:error and the run continues with degraded memory. Streaming is never blocked; a failed save never fails the turn.

Devtools

Five events on aiEventClient (from @tanstack/ai-event-client): memory:retrieve:started / :completed, memory:persist:started / :completed, memory:error (phase: 'recall' | 'save'). Payloads carry the adapter id and fragment/receipt counts, not full memory text. Error events include scope only when it was already resolved; if the resolver threw, scope is omitted.