Your agent has amnesia, and you pay the tax every morning. Agent OS makes it write down what it learned before the session dies — and read it back the moment the next one starts.
That is why it stops working at project three. A README describes. What you need is a constitution — and a constitution without enforcement is just a wish written down.
Agent OS splits the one file almost everyone overloads into three, divided by how fast each one changes:
The separation is the whole trick. State churns daily, so it has to be compressed or you bleed tokens on every session start. Wisdom is permanent, so compressing it destroys the exact thing you kept it for. Share one file and no setting satisfies both — you either lose the lessons or pay for the bloat.
Master CLAUDE.md with a project registry, plus the three-file
continuity kit. The registry is what lets the agent answer questions about
project seven while you're sitting in project two.
One denies every write to .env and friends. One runs your tests
the instant code changes and feeds failures straight back to the agent.
Enforcement, not polite requests.
new_project.py lays down the full doc kit and hook wiring in one
command. It never overwrites an existing file, so it's safe on a project
that's already half set up.
registry_scan.py walks every project you own and reports which
ones are missing docs or have gone stale. Reads only filenames and
timestamps — safe to run against folders holding secrets.
Each rule written next to the specific failure that produced it. A rule without its scar tissue gets deleted by the next person who finds it inconvenient.
What breaks first at 1, 5, and 20 projects — session amnesia, then cross-project blindness, then repeated mistakes, then token bleed, then silent write gaps. In that order.
You should not have to trust a stranger's hook with your repository. Run these the minute you unzip it — no install, no dependencies, standard library only:
Fair question — I asked it myself two days after launch. Auto Memory and claude-mem (71k stars, free) are both real and both free. Here's what they don't cover:
They solve session memory — what happened last time. They're not
opinionated about rules you want enforced, not just remembered. A
PreToolUse hook that hard-blocks writes to .env is a
wall. A memory note that says "don't touch .env" is a suggestion the model
can still talk itself past.
They're single-project. Run more than one thing at once and the failure that actually costs time isn't forgetting — it's one project's agent silently not writing to the shared log the others read from. Neither has a concept of "other projects."
If you run one project and just want to stop re-explaining yesterday, native Auto Memory genuinely covers that — save your $29. This is for the point past that, where "remembered" isn't the problem anymore and "enforced, and consistent across projects" is.
Fifteen minutes to set up. The quickstart ends with three concrete tests, so you'll know it's live rather than hoping.