Append one fact to this project's ledger, at the moment it is produced. Call it right after: the user states a rule or a boundary (constraint); the user corrects something you did (correction); you choose between alternatives (decision — say why in text, and ref where it is recorded); a tool result yields a figure you will rely on (measurement — ref the command, refresh says whether it reproduces); work is parked, blocked, or handed to someone (thread — ref the branch, PR, or issue). Decisions, measurements and threads must carry a ref. Write text so it says WHAT the thing is before WHY: recall scores entries against a future task by their text, and an entry that only explains a reason will not be found by the task it belongs to. To retire an earlier entry pass its id in supersedes; the old entry stays on disk, marked replaced. Nothing here is ever deleted.

Bring forward what earlier sessions recorded, scored against the task you are about to do. Call it once, at the start of a task, with the task in one sentence. Constraints and corrections come back in full every time. Everything else is scored for whether not knowing it would cause wrong or repeated work: high scores come back in full, middling ones as one line, the rest are omitted. Without a task it lists everything one line each. If scoring is unavailable it returns everything and says so — it never returns less than the ledger holds without telling you.