Mnemra is a memory server for agents.
MCP carries context in. Postgres carries it out. In between: Rust, pgvector, per-tenant isolation, and row-level security. Self-hosted or managed — your choice.
The project exists because agents forget. Every session starts cold. Every tool call loses the prior conversation. Mnemra changes that: a persistent, queryable context layer that plugs directly into the MCP transport your agents already use.
The data flow
Each agent session connects to Mnemra over MCP. Observations — tool results, reasoning traces, user inputs — are written to the store. On the next session, the agent queries the relevant context and resumes where it left off.
The schema is simple: a context entry has a tenant ID, a session ID, a timestamp, a content embedding (pgvector), and the raw text. Queries use cosine similarity. The top-k results come back as MCP resources.
What’s next
The alpha is in development; beta will follow. If you want early access, join the waitlist on the landing page.
The source is Apache-2.0. The repo is at github.com/mnemra.