No jargon here. The technical spec is PRIORS.md; this page is the version you'd sketch on a whiteboard.
You ask your agent to review your code. It finds 35 things. You fix 22, skip a few on purpose — and the next run hands you 25 "new" findings. Nothing you fixed is acknowledged, and the ones you skipped are re-argued as if you'd never decided.
priors puts a small ledger — plain text, inside your repo — between the
agent and your work. The agent must read it before speaking, and a tiny
script makes sure it did: a run that ignores the notebook simply doesn't
count.
Instead of a fresh pile, you get accounting: what got fixed (by name), what carries over (same items, never dressed up as new), what needs your decision, and only then — what's genuinely new.
Say "skip it" once and it's locked. No future run may bring it back — the only key is a real change in the thing you decided about. If the agent ever thinks a decision deserves revisiting, that arrives as a question to you, never as finding #26.
The same notebook holds what the agent learns by doing: the CLI flag that works, the login flow of that one website, the vendor whose PDFs need decrypting first. Next session starts already knowing.
Things do leave the stage — but through three visible doors, never by vanishing. Settled things stop acting but stay in the book. Findings whose code was deleted go to rest, and wake with their memory if the code comes back. And if a secret ever slips into the notebook, a human can remove the content — while the removal itself stays on record.
The whole system in one sentence: the agent proposes, the notebook disposes — so re-runs get shorter, decisions stay decided, and one day the agent tells you the best sentence a reviewer can say: "everything settled still holds — nothing new."