Mneme is a local, encrypted memory server for AI agents. Facts, decisions, and context persist in SQLite on your machine: searchable by keyword and meaning, decaying by confidence, never leaving your control.
FTS5 keyword search and vector similarity in one query: exact matches and semantic neighbors, ranked together.
AES-256-GCM on a local SQLite store. Your memory never leaves the machine: no cloud, no account, no telemetry.
Memories age. Stale facts lose confidence over time and fresh ones surface first. Recall stays relevant, not cluttered.
A full Model Context Protocol surface (store, recall, link, forget, and reflect), usable by any MCP-aware assistant.
Every memory is written to an encrypted SQLite file with an embedding and a confidence score. On recall, Mneme blends keyword and vector hits, weights them by freshness, and returns a compact, ranked set: the same shape your assistant already reads.
▸ staging = eu-west-2/stg-a conf 0.98 ▸ deploy via ./scripts/deploy.sh conf 0.95 ▸ helm chart v0.4.1 conf 0.91 ▸ old: helm v0.3 conf 0.22 (decayed) ✓ 3 memories · 7ms · FTS5+vector
Point Perseus at your Mneme store and your assistant opens with both live environment state and durable memory, resolved into one byte-stable file. Two single-binary tools, one build-pair.
Explore Perseus ↗Install the server, point your assistant at it. Everything stays encrypted on your machine.