MuninnDB
A long-term memory server for AI agents — persistent, structured recall over MCP.
Named for Muninn, one of Odin's ravens — the one whose job is memory.
Agents forget everything between sessions. MuninnDB is my answer: a persistent
memory store that any MCP-capable agent can read from and write to, so context
survives across sessions, models, and machines.
What it does
- Atomic memories — one fact, one decision, one bug per record, with
concept slugs, tags, and entity links that make recall precise instead of fuzzy. - Recall that ranks — vector similarity blended with recency, trust, and
feedback signals; noisy or stale memories get demoted over time. - Evolution over deletion — memories are updated in place with provenance,
so an agent can see how an understanding changed rather than just its latest state. - Maintenance sweeps — scheduled hygiene passes retire stale nodes and
collapse duplicates so the store stays sharp as it grows.
How it runs
A small fleet: a primary node plus observers replicating from it, managed under
systemd with scheduled backups. Every agent I run — locally and on my VPS —
connects to it over MCP and stores what it learns automatically.