Bellamente docs
Bellamente is a local-first memory tool for AI agents. It stores durable facts and documents,
recalls them semantically, and sits in front of your local LLM as a drop-in
/v1/chat/completions proxy that injects relevant memory — with a durable trace of exactly what
it did. One binary. No Docker, no account, no cloud, no telemetry.
Early release. v0.1.0 is usable and tested — and not complete. The roadmap is the honest backlog.
Install
Use npm:
npm install -g bellamente
bella doctor
bella
Or Python tooling:
pipx install bellamente
bella doctor
bella
For a one-shot Python run without installing a persistent bella command:
uvx bellamente doctor
The npm and PyPI packages install a tiny launcher. When no verified cache exists on supported
Windows x64 and Linux x64 machines, it downloads the matching GitHub release binary, verifies it
against SHA256SUMS.txt, and runs it.
You can also download a supported Windows x64 or Linux x64 binary from the latest release:
chmod +x bella-linux-x64 # Linux only; skip on Windows
./bella-linux-x64 doctor # verifies DB, embedding model, ports, disk
./bella-linux-x64 # serves on 127.0.0.1:8080
Zero config, really
- No
.env, no API key on localhost. The first boot creates the embedded database (Postgres + pgvector compiled into the binary) and downloads the embedding model once. - Set
BELLA_API_KEYif you want a bearer key required anyway. - Bind beyond localhost (
BELLA_HOST=0.0.0.0) and a key is auto-generated, stored in the data dir (apikeyfile), and enforced. Exposure without auth is never the default.
First memory in 30 seconds
curl -s localhost:8080/memories -H 'content-type: application/json' \
-d '{"memories":[{"content":"Jeff prefers dark mode"}]}'
curl -s localhost:8080/search -H 'content-type: application/json' \
-d '{"q":"what theme does Jeff like?"}'
Every response carries an x-bella-trace-id header — open the dashboard at
http://127.0.0.1:8080/ to see exactly what was searched, what matched, at what score.