Agent memory you can open in a text editor.
BizerBrain is a markdown knowledge base an AI agent and a person share. Notes are plain .md files in a folder you host, the agent reads and writes them over HTTP, and the web UI shows the same files changing live.
A memory kept in someone else's database is not yours.
Most agent memory is an opaque store you query through the agent that wrote it. BizerBrain keeps every note as a markdown file on a volume you control. You can read it, grep it, back it up, edit it by hand, and take it somewhere else, with or without the agent that produced it.
Every page carries what is true now and how it came to be known.
A topical page has two halves. Above the separator is compiled truth, which the agent rewrites as understanding changes. Below it is an append-only timeline, where each entry is dated and links back to the source it came from. Current understanding and audit trail sit in the same file, so a claim can always be traced.
## Timeline
A web UI, an HTTP API, and an MCP server over one folder.
Markdown editor
A React editor with a file tree, rendered preview, backlinks, and wiki-link navigation. Changes the agent makes appear without a reload.
File API
A small service exposing the folder over HTTP: read, write, move, delete, search, tree, and a watch stream. The agent never needs the filesystem.
MCP server and skill
Four tools, list_notes, search_notes, read_note, and write_note, shipped both as an MCP server and as an agentskills.io skill.
One compose file, one volume, no public port.
The published image runs the UI and the file API together. Your notes live on the host at a path you choose, so removing the container never removes the brain.
# Your notes live here, on the host sudo mkdir -p /srv/bizerbrain/brain # Start it docker compose up -d # Reach the UI from your laptop ssh -L 8080:localhost:8080 user@your-vm
The host port binds to loopback only, so nothing is exposed publicly and the UI is reached through an SSH tunnel. Agents in other containers connect over the internal Docker network instead.
Any MCP-capable agent, connected with one setting.
The agent talks to BizerBrain over HTTP, so it can run in its own container, on the same host, or on a different machine entirely. No shared volume mount is required. Set one environment variable and install either the MCP server or the skill.
| Where the agent runs | BIZERBRAIN_API_URL |
|---|---|
| Its own container, same host | http://bizerbrain:8080, after attaching the container to the bizerbrain network. This is the built-in default, so the common case usually needs no configuration at all. |
| Directly on the host | http://localhost:8080 |
| A different machine | http://your-bizerbrain-host:8080 |
Alpha, and open source under Apache-2.0.
Clone it, run it, tell us what breaks.
The repository has the compose file, the deployment steps, and the skill the agent follows, all in one README.