this is a very real take. a lot of teams blindly plugged vector db into their stack without questioning if retrieval is actually learning or just matching numbers.

i’m building an open source project called selfmemory.com exactly around this gap. not just better shelves, but a memory layer that can evolve per agent, per tenant, and across sessions. multi-tenant, mcp with auth 2.1, python sdk, llm chatbot, everything open source. the idea is simple: memory should not be frozen at ingest time. it should adapt, get corrected, and improve over time.

completely agree that cosine similarity alone cant be the brain. retrieval needs feedback loops, structure, and learning on top. that’s what we’re trying to experiment with in selfmemory. would love more people thinking about fixing the foundation instead of just optimizing the storage layer.