$ man bellamente
Questions, answered.What Bellamente is, how it differs from the other memory tools, and where your data actually lives.
How is Bellamente different from GBrain, Mem0, Zep, and other agent-memory tools?
They share plumbing — an embedded Postgres + pgvector store is a commodity now — but they solve different problems. GBrain is an authored markdown "second brain": you write a git repo of notes and it builds a typed knowledge graph. Hosted tools like Mem0 and Zep keep your memory on their servers. Bellamente is runtime agent memory that lives entirely on your machine: the agent writes and recalls at runtime, every recall leaves a durable trace you can read, corrections are versioned and forgetting is reversible, and an OpenAI-compatible proxy can ground any chat. It deliberately has no knowledge graph and sends zero telemetry. Same engine room, different ship.
Where does my data live? Is any of it sent anywhere?
Everything stays on your machine — one embedded database file in your app directory. The server binds to 127.0.0.1 by default and sends no telemetry. There is no cloud account and nothing phones home; even the bug reporter just prints a prefilled GitHub link for you to submit yourself.
Which agents and clients work with it?
It is MCP-native — bella mcp plugs into Claude Code, Cursor, Codex, and Cline. Any OpenAI-compatible client works through the proxy by changing one base URL. And any agent can call the plain HTTP API directly.
Do I need Docker, a database server, or a model server?
No. It is one binary with embedded Postgres (PGlite + pgvector) and local embedding models built in. Run npm install -g bellamente (or pipx install bellamente), then bella serve. First boot creates the database and downloads the model automatically.
What platforms are supported?
Windows x64 and Linux x64 today. macOS binaries are not published yet — it is on the roadmap. You can also run it on any platform from source with Bun.
How do I trust what it remembered — or catch a bad memory?
That is the whole point. Every recall is logged with the query, the candidates, and their scores; every correction is a new version and the old one stays; forgetting is reversible and audited. You can inspect all of it in a local dashboard or query it with plain SQL.
Can it leak my API keys or secrets into memory?
It is built to avoid that. The agent is instructed not to store credentials, and a deterministic gate redacts API keys, tokens, and private keys from memory content and metadata before storage. We are honest about the limit: it catches real credential formats, not every conceivable secret.
Is it free and open source?
Yes — MIT licensed and free. The source is on GitHub and contributions are welcome.
Is it production-ready?
It is version 0.1.0 — an early but real, tested MVP. It is usable today; the public roadmap is honest about what is still ahead (macOS builds, provider-specific proxy shapes, retrieval evals, and more).