memory-hub

Persistent, governed memory for AI agents. MemoryHub gives agents a shared memory layer with semantic search, version history, scoped access control, and audit logging. It works with any agent framework that speaks MCP.

Try it now -- pip install "memoryhub[local]" for a personal edition backed by SQLite (no infrastructure needed), or deploy the cluster edition on OpenShift AI for multi-agent governance, OAuth 2.1, and team-wide shared memory.

Retrieval quality: 84.9% on AMB PersonaMem (2nd on the leaderboard), R@5=0.999 on LongMemEval. Requires Python 3.10+.

How to think about agent memory

The model never remembers anything -- at inference time, a memory is just tokens in context, and it makes no difference whether they came from a markdown file, a vector store, or a graph. Memory is a context-assembly policy problem: how did the right items get selected, who was allowed to see them, what happens when they conflict, and can you reconstruct what an agent knew when it acted?

Two principles drive everything here. First, give the agent 100% of what it needs and 0% of what it doesn't -- no retrieval trick compensates for missing context, and garbage overlap degrades performance even when the right facts are present. Second, work backwards from the forensic investigation: who or what did the thing, what memories were in context, who wrote them, were they in conflict, did storing them violate policy, and which other agents were exposed to them?