Project Memory Keeper

Every AI session starts from nothing. You re-explain the project, decisions get re-argued, and notes pile up until nobody reads them, including the AI.

This skill keeps a small, honest memory in your repo. PROJECT_MEMORY.md holds the current state, goals, constraints, a one-line index of key decisions, conventions, open questions with owners, and a glossary. It is read at every session start and kept under a size budget. A decisions folder holds ADR-style records with context, alternatives, consequences and status; changed decisions are superseded, never rewritten. Each session ends with a short handoff note saying exactly where to start next time.

The procedure covers what is worth remembering versus noise, how to update without bloat, archiving, and flags for stale or contradicted entries. A Node.js lint script with no dependencies checks the folder: broken links between ADRs, superseded ADRs not marked as such, duplicate decision titles, files over budget, undated entries, bad statuses and duplicate numbers, plus warnings for stale entries and open flags. It reads most existing ADR formats.

Examples include a clean memory folder, a deliberately broken one (14 errors found), and a walk-through of a full session.

The lint checks structure and dates, not truth. Memory lives in your repo, so it never stores access values or personal data.