Memava for agent builders
Give your agents a memory humans can read.
Everyone building agents hits the same wall: the agent is stateless, so you bolt on a memory. A vector store remembers approximately. A JSON file remembers privately. What you usually want is simpler and harder: durable notes a human can open, correct, and trust, that several agents and several people share under rules.
Why plain files win for agent memory
When memory is markdown in a vault, debugging an agent's beliefs is reading a file. Correcting them is editing it. Reviewing what changed after a run is a diff with the agent's name on it. Memory stops being a black box inside your stack and becomes part of the team's shared world, which is where it was always supposed to live.
What the MCP surface gives you
- The twenty-three note tools, from list, read and search to history, proposals and batches with one undo. Plain JSON-RPC, easy to reason about.
- Per-agent identity and permissions: each agent gets its own token and its own grants. An agent can never gain access from a wildcard rule; every grant is explicit.
- Attribution and reversibility: every write lands in git history under the agent's name; any note rewinds.
- A read audit: the log answers what the agent looked at, which matters the day something goes wrong.
Multi-agent, multiplayer
Two agents writing the same note merge instead of clobbering, and a human watching the note sees the words arrive live.
That one property, merge instead of overwrite, removes a whole class of agent-memory corruption bugs.
Honest limits
Search today is text search, not embeddings; semantic search is on the roadmap and we say so. If your use case is retrieval over a million chunks, this is not that. If it is a working memory your agents and your team maintain together, it is exactly that.
The developer page has the connection details.
Keep reading: Memava for family offices · Memava for founders · Memava for product teams mixing AI tools