Loading page
Two ways in, once you have access. Either works — pick whichever matches how you already work.
The repository is private, because it holds the engineering records Scar was distilled from and some of those are client work. What you are reading on this site — the rules, the lessons, the architecture — is the part that transfers, and the part meant to be public. If you want to try the system yourself, get in touch and I’ll grant access — then the steps below get you running.
Your coding agent starts getting stopped before it repeats work you have already paid for. It begins with what ships in the repository — today 9 operating rules and 325 lessons, each carrying the failures that produced it — and grows from your own work without you writing any of it: the agent records what surprised it, and records sharing a mechanism become a lesson, through gates it does not control.
At the start of a session the agent loads a ~2.2K-token digest, then asks Scar what it knows before a non-trivial task. You don’t run anything — after setup it happens on its own.
Already installed? Skip to the commands you reuse ↓
Paste this into Claude Code, or any coding agent that can run commands. It clones the repository, reads the setup instructions written for it, runs them, and checks its own work.
Set up Scar for me.
Clone https://github.com/lumen202/unified-brain.git, read its SETUP.md,
and follow it. When you're done, tell me the kernel size and confirm
you can query Scar.The instructions it follows are SETUP.md in the repository — written to be executed rather than read, with a verification step after each action so a half-finished setup reports itself instead of looking fine.
Needs Node 20 or newer. One command; safe to run again at any time.
git clone https://github.com/lumen202/unified-brain.git
cd unified-brain
./install.shAfter setup, using Scar is a handful of things you say in chat, and one thing you watch for. Nothing here needs a terminal.
The agent writes one import line into that project’s AGENTS.md/CLAUDE.md, with the right absolute path, then installs the gates with a single command. From then on every session in that project loads the operating loop by itself. Without the line Scar stays silent, and nothing will tell you — which is also true of the gates: each one fails open by design, so a missing gate and a quiet gate look identical. That is why installing them is one command rather than a checklist, and why hooks-coverage exists to say which projects are actually running them.
The agent consults Scar on its own: a ~2.2K-token digest once per session, then ~2.0K–4.4K tokens of recalled lessons before a non-trivial task, and it reports the outcome afterwards so ranking improves. That’s the whole standing cost — less than one screenful of code per task.
This line is the proof it’s working. Tool calls are invisible, so a Scar that is consulted silently and one that isn’t wired up look identical from your side. If the agent never mentions Scar, don’t assume it’s thinking quietly — jump to the fix below.
The agent runs a recall for you and shows the ranked lessons with what matched. Any task, phrased the way you’d say it to a colleague.
The agent writes the sanitized, transferable lesson from what the session just learned — and refuses if it can’t be stated without confidential terms. Only for surprises: a fix you’d have predicted teaches Scar nothing.
The one upkeep command. It checks whether records have piled up unread, compresses the ones sharing a cause into single lessons, rebuilds what agents load, and runs the gate. The rewriting is done by the session you’re already in, so there is no API key and no per-token bill. Say sync only if you just want your edits to take effect.
Two causes, in order of likelihood. The project is missing its adoption line — add it (or run /scar-adopt there), substituting your clone location:
@<path-to-unified-brain>/brain/core/AGENTS.mdOr the agent says it can’t reach scar_recall — the tool server isn’t registered in that environment:
claude mcp add scar -- node <path-to-unified-brain>/brain/tools/mcp-server/index.mjsEverything above works without an agent too. The two worth knowing: ask Scar directly, and rebuild the generated index after records change (your agent normally does both for you).
cd brain
npm run recall -- "adding a modal that must work on narrow viewports"npm run distill:cluster
npm run recall:build
npm run verify