As you work, Claude keeps one short dossier of the project: every decision with its reason, the key numbers, what was agreed, the open questions and the next steps. It answers "what did we decide?" from it and carries it into a new chat or a Claude Project in one step.
| # | Decision | Why |
|---|---|---|
| 1 | — | |
| 2 | $12/mo, annual 20% off | replaces #1: costs came in higher |
| 3 | Stripe Billing | already used for invoices |
Forty messages in, nobody remembers why you picked Postgres or which price you settled on. Start a new chat and it's worse: everything gets retold, and a long chat re-sends its whole history with every message.
When a real decision lands in an ongoing project, Claude offers once to keep a Brain. Or just say "keep track of this".
Each decision goes in with its reason and date. You get one line, 🧠 Brain: +1 decision, not the document again.
"Why did we pick Stripe?" gets the decision number, the reason and the date, or a plain "it isn't in the Brain".
Only what was said or confirmed. Claude's own inferences are marked (assumed).
Decisions are never silently rewritten. A changed one is struck through; the new one says what it replaces.
Short. Under about 800 words; closed items move to an archive, one line each.
No secrets. Passwords, keys and card numbers become (kept elsewhere).
claude.ai, desktop, mobile: one artifact, updated in place.
Cowork: PROJECT_BRAIN.md in your folder.
Claude Code: PROJECT_BRAIN.md at the repo root, read when needed, so it doesn't cost tokens on every request.
"Keep a project brain for this." · "What did we decide about the database?" · "Show the brain." · "Let's continue in a new chat."
The Brain already holds what matters, so the next place starts from it instead of from the whole history.
One block to paste: the Brain plus a line telling Claude to keep it updated. With the saving, e.g. 60K → 1.5K tokens.
The Brain in the Project's knowledge and instructions that keep it current in every new chat there.
Rewritten for someone who wasn't there, with who to ask about each open question.
Saved as PROJECT_BRAIN.md, with one line to start the next session.
# Directory → search "Project Brain" → Add # or upload the skills: Settings → Capabilities → Skills
claude plugin marketplace add egorfedorov/project-brain claude plugin install project-brain@project-brain
Instructions for Claude, nothing more: no scripts, no hooks, no network calls, no settings.