What problem does it solve? Valuable knowledge from a working session — architectural decisions, non-obvious learnings, user feedback, diagnosed failures — is easily lost when the session ends. This Skill reviews the recent session, identifies what is worth keeping, and writes each approved item to the correct durable home so nothing has to be rediscovered later. ## Core Features & Use Cases - Evidence-anchored capture: Reviews recent turns (and an optional --since <git-ref> diff) and proposes candidates with file:line, ID, or quote evidence — never fabricated. - One fact, one home: Routes each item to the right destination — decisions to docs/decisions.md via decision_register.py, facts and feedback to auto-memory files with a MEMORY.md pointer, deferred work to docs/backlog.yaml, repeatable failures to harness/LESSONS.md, and coined terms to docs/glossary.yaml. - Human-in-the-loop writes: Presents a numbered proposal list and writes only what the user approves; it never auto-writes a decision, gate config, or enforcement rule. - Governed decision flips: Replacing an existing DEC triggers a blast-radius scan (--scan-flip), and cross-scope flips require explicit user confirmation via a confirm token before the register is updated. - Use Case: After settling on a retry threshold and fixing a recurring wiring bug during a session, invoke this Skill to record the threshold as a DEC, log the bug pattern in LESSONS.md, and capture the user's correction as a feedback memory — each approved individually. ## Quick Start Ask the assistant to run the remember skill to review this session and propose durable entries for the decisions and lessons we just made.
