tanstack-ai-memory-redis

Create a Redis client with ioredis, pass it to redis() from @tanstack/ai-memory/redis with a key prefix, then wire the adapter into memoryMiddleware with your scope. For node-redis v4+, wrap the client with fromNodeRedis first.

ioredis works directly with the redis() adapter since it exposes the lowercase RedisLike API. node-redis v4+ uses camelCase methods like sAdd, so it must be wrapped with fromNodeRedis or you get a client.sadd is not a function error.

No, the adapter works with plain Redis and no vector index. Ranking happens client-side using lexical matching, optional cosine similarity via an embedder, recency, and importance signals.

Records are invisible across processes when each process uses a different REDIS_URL or key prefix. Ensure every process connects to the same Redis instance and passes the identical prefix option to redis().

Client-side ranking loads every record for a scope into Node.js and scores it, which works well up to roughly 10k records per scope. Beyond that, write a vector-index-aware adapter against the same recall and save contract.

A record whose JSON fails to parse is skipped during recall and left in place, never deleted automatically. The adapter emits a one-time console.warn per bad id, and you must fix or delete the offending record key manually.