From zero to operational memory. Pick a track below, install it, use it day to day, think decisions through, or push it as far as it goes.
One command sets up everything: a .brain/ folder for state, slash commands, and the MCP config to connect any agent.
From your project root, run init. It creates a .brain/ directory for operational state and installs the slash commands.
Brain OS auto-discovers .brain/ from any subdirectory, no path config needed.
Register Brain OS as an MCP server. Your agent can now read and write operational state.
On Cursor, Zed, Windsurf, or GitHub Copilot? See the next card.
Paste this into your client's MCP config (e.g. mcp.json). Same .brain/, every tool.
Open a fresh agent session and run the scanner. It reads your project and reports what it found.
If /brain already exists in your project, everything installs under the /brain: namespace instead, so nothing collides.
Chat apps like ChatGPT, Claude, and Grok ship their own built-in memory, separate from Brain OS. The connect-time toggle to reference your chat history and memory controls the app's memory, not Brain OS.
Brain OS starts empty, so keep the built-in memory on at first. As you work, say log this to Brain OS for anything worth carrying across tools, that routes it to the shared connector. It compounds: the more you log, the more it can answer. Once it holds enough, you can switch the built-in memory off and make Brain OS the source of truth.
The everyday loop. Open a session, do the work, close it clean, your agent carries the state forward so the next session starts with full context.
/focusAt the start of a session, your agent reads operational state and surfaces what actually matters today, by urgency, momentum, and what's already in motion.
No more "where were we?", the agent already knows.
When you decide something, capture it with the reason and the alternatives you rejected. If it conflicts with a past decision, your agent flags it before you proceed.
This is what lets the agent push back instead of yes-anding.
/wrapEnd the session by capturing what changed across every project: decisions logged, patterns observed, plan steps advanced. A restart-ready handoff for whatever opens next.
Because state lives in .brain/, you can switch tools or machines mid-stream. Log a decision in Claude Code, open Cursor tomorrow, same brain, full context.
Commit .brain/ to git to share the operational state with your whole team.
Some decisions deserve more than a one-liner. Use Brain OS to think them all the way through, and to catch the work that keeps slipping.
/strategyFor a meaty decision, /strategy walks the trade-offs, surfaces related past decisions, and helps you reason to a position you can defend later.
The outcome gets logged with its reasoning, so future-you (and future agents) see why, not just what.
Before you commit a new direction, the agent scans prior decisions and returns conflict, caution, or clear, with the original reasoning attached.
"We considered this in March and rejected it because X." That's the disagreement transcript tools can't give you.
/retroPeriodically, compare what you said you'd do against what actually shipped. /retro surfaces the work that reappears every session but never moves.
/patternsRecurring blockers, avoidance loops, scope creep. /patterns names the things you're too close to notice across sessions.
Once it's part of your loop, these get you the rest of the way: cross-tool continuity, team sharing, semantic recall, and the full command set.
Connect Brain OS in every client you use. The same .brain/ folder is read by Claude Code, Cursor, Zed, Windsurf, and GitHub Copilot, a decision logged in one shows up in the next.
No product changes needed. Commit .brain/ to your repo, or sync it via Dropbox/Drive/NFS. Every tool call reads fresh from disk, so teammates see each other's operational state.
Opt in to recall-by-meaning. Set an embeddings backend, then ask "what did we decide about billing?" and get the answer, not a list of messages.
Use =openai instead if you'd rather use the API. Unset by default, nothing downloads silently.
Eight slash commands cover the whole loop:
/brain /focus /decide /strategy
/wrap /patterns /retro /graph
/graph visualizes how your entities, decisions, and blockers connect. Read the source for the full tool reference.