Been going down the rabbit hole of agentic memory setups lately, and I’m curious what people are using.
Right now, my stack is Redis for agent memory and LangGraph checkpointers for state recovery
It’s been solid from a latency and flexibility standpoint, I get fast retrieval and control over eviction / TTL / session memory vs persistent memory
But I keep seeing a lot of different approaches floating around Weaviate / Pinecone-style vector DB as “memory”, MongoDB, Postgres + pgvector setups, purpose-built agent memory frameworks popping up recently (mem0, cognee)
So checking a few things with folks building real systems:
1. What are you using today for agent memory?
(and is it actually holding up under multi-step agents?)
2. Biggest pain so far?
3. If you switched stacks, what made you do it?
My current take (could be wrong):
Vector DBs alone don’t feel like enough for true agentic systems. You need something that behaves more like a real-time memory layer + state store, not just embeddings lookup. That’s where Redis has been working for me, it kind of sits between cache, database, and memory fabric.
But I’m curious what’s working for others.