Local-first memory & context framework

Sovereign memory for AI agents.

Lummenna organizes agent memory as a memory palace, retrieves with hybrid search — BM25, dense vectors, and graph traversal fused with reciprocal rank fusion — and manages the lifecycle with three-layer forgetting. Embeddings run on-device. Storage is one SQLite file. No cloud, no phone-home. Free for developers; Pro runs it as infrastructure.

Free & open source under AGPL-3.0-or-later · Python 3.10+ · macOS, Linux, Windows

Why Lummenna

Memory your agents own — on hardware you own.

Five properties, engineered in rather than promised.

Sovereign

Your data. Under your control. Verification, retrieval, and storage all happen on your machine — air-gapped deployments work.

Private

Local-first, always. No telemetry, no license server, no per-query API calls. The dashboard itself runs fully offline.

Persistent

Memory that endures. Context that remains. A single SQLite file holds the whole palace — trivial to back up, move, or inspect.

Intelligent

AI that remembers what matters. Hybrid retrieval fuses lexical, semantic, and graph signals with reciprocal rank fusion.

Attentive

Signal over noise. Purpose over clutter. Three-layer forgetting — time decay, interference, budget eviction — keeps memory sharp.

Integrated

MCP server for ten coding agents and a LangChain memory adapter — free. Pro adds the FastAPI REST server + dashboard, LangGraph checkpointing, and P2P sync.

How it works

A palace, a searchlight, a gardener.

01 — ORGANIZE

Memory palace

Memories live as chunks inside loci inside rooms — the classical method of loci, made queryable. Structure carries meaning that flat vector stores lose.

02 — RETRIEVE

Hybrid search

BM25 full-text, dense vectors, and graph traversal run together and fuse with reciprocal rank fusion. Token-budgeted context assembly returns exactly what fits.

03 — FORGET

Three-layer lifecycle

Time-based decay, similarity interference, and budget eviction degrade memories gracefully — FP32 → FP16 → INT8 → BINARY → released — instead of hoarding everything forever.

Works where your agents work

One memory across every tool.

Ship-ready MCP configs are included for ten clients, plus framework adapters and a REST API.

Claude Desktop · MCP OpenAI Codex CLI · MCP GitHub Copilot · MCP Cline · MCP Zed · MCP opencode · MCP Continue · MCP Goose · MCP Aider · MCP Devin · MCP LangChain · adapter LangGraph · checkpoint saver · Pro REST + OpenAPI · FastAPI · Pro

Get started in minutes

The Community Edition is free. Forever.

The complete framework under AGPL-3.0-or-later. Pro adds the commercial license and enterprise controls.