Your agent has amnesia and you're paying the tax every morning.You close the terminal. Tomorrow, Claude Code has no idea what it built, what you decided, or what broke last time. So you explain it again. And if you run more than one project, you explain it again for each one.Agent OS is the layer that fixes this. Not a prompt pack — a set of files, hooks, and enforced rules that make your agent write down what it learned before the session dies, and read it back the moment the next one starts.WHAT YOU ACTUALLY GET• CLAUDE.md master template with a project registry, so the agent knows every project you own from any directory• Three-file continuity system — law, state, and wisdom kept separate, which is the one structural decision that determines whether this survives project three• Two working hooks — one blocks the agent from ever touching .env, one runs your tests the instant code changes and feeds failures straight back• Scaffolding script — full doc kit for a new project in one command, never overwrites what's already there• Registry audit script — walks every project and reports missing or stale docs• knowledge-capture skill — enforces the write-down step• The 12 Rules — each one with the specific failure that produced it• Scaling guide — what breaks at 1, 5, and 20 projects, in the order it breaksTHE CORE IDEAMost people treat CLAUDE.md as a README. That's why it stops working at project three.Agent OS treats it as a constitution with enforcement, split across three layers:LAW → CLAUDE.md → changes rarely → never compressedSTATE → PROGRESS.md → changes every turn → compressed after 24hWISDOM → LESSONS.md → changes on correction → never compressedThe separation is the whole trick. State churns and must be compressed. Wisdom is permanent and must not be. Collapse them into one file — which is what almost everyone does — and compression either destroys your lessons or leaves your logs bloated. There's no setting that works for both.WHO THIS IS FOR• You run Claude Code on more than one project• You've re-explained the same context more than twice• You've watched the agent repeat a mistake it already madeWHO THIS ISN'T FOR• One weekend project, one session — you don't have this problem yet• Looking for "act as a senior engineer" prompts. There are none here.• Want a plugin to install. This is plain markdown and four Python scripts you can read in fifteen minutes.VERIFIED, NOT ASSERTEDEvery script ships with a --demo self-check. Run them before you trust them:python hooks/guard_secrets.py --demopython hooks/verify_on_save.py --demopython scripts/new_project.py --demopython scripts/registry_scan.py --demoRequirements: Claude Code, Python 3.9+. Works on macOS, Linux, Windows.Setup takes 15 minutes. The included QUICKSTART ends with three concrete tests, so you'll know it's live rather than hoping.30-day refund, no questions asked.