mnemo

Local memory for AI coding agents. On‑demand. Strictly offline. Built for 10+ agents on a 16 GB machine.

What it is

A persistent memory layer for agents (Claude Code, Cursor, Windsurf, any MCP client) that remembers decisions, bugs, progress, and rules across sessions. What sets it apart:

• strictly local — zero cloud calls; embeddings and LLM run only on your machine;
• on‑demand — nothing runs in the background; the service spins up under load and shuts down after a grace period;
• no Docker daemon — embedded storage inside a single process (LanceDB);
• lightweight — ~1 GB RAM while active, ~0 when idle; a small model (Qwen3‑4B / Gemma 4) runs only during background consolidation;
• concurrent — one shared service serves 10+ agents; writes are cheap (embed + insert, no LLM on the hot path);
• simple — 3 core MCP verbs (recall / remember / search); soft project scoping with first‑class cross‑project search.

Core principles

1. No LLM on the write path. A write = local embedding + upsert. The LLM runs only in the background, in batches.
2. One shared process, started on demand. Not 10 stdio processes hitting one file — one service + thin shims.
3. Heavy things are transient. The generative model is loaded only for a consolidation window, then unloaded.
4. Typed memory, soft scoping. decision / debug / progress / rule / ... across project / global / (optional) session scopes — projects organize memory but never wall off search; cross‑project search is first‑class.
5. Tiny, obvious API. Three verbs the agent learns in seconds; type/scope are parameters, not extra tools. Ops commands live in the CLI.
6. Extra models are opt‑in. Default = embedder + one optional background generator; more small models only when justified.