tanstack-ai-memory-in-memory

Import inMemory from @tanstack/ai-memory/in-memory and pass the result to memoryMiddleware as the adapter along with a scope. No external services or configuration are required for basic usage.

Use inMemory for local development, tests, and single-process demos where persistence is not needed. Use the redis() adapter for production multi-process deployments or anything that must survive restarts, since each process keeps its own Map.

Yes, pass an embedder option with an embed(text) function returning a number array. Both recall and save then embed through it for semantic scoring; without it, recall scores lexically and by recency.

Records vanish on process restart and each worker process has its own Map, so multi-process deployments get inconsistent memory. It scans every record per recall, which is fine up to roughly 100k records before Redis is recommended.

Pass an extract(turn, scope) option that returns derived facts to persist alongside the raw turn, such as calling an LLM to pull out user preferences. Without it, save stores the raw user and assistant messages.