lifecycle: no exit ritual — 97% of memory hygiene happens only on an explicit user tidy-up command

[1Croydan1]

Problem

The product has an entry ritual (initiatives → awake → overview) and no exit ritual. A Claude Code session ends silently, the "end of work" moment never arrives — and all memory hygiene happens only when the user EXPLICITLY commands a cleanup.

Numbers (audit of 13,105 sessions, adversarially verified)

• ~5 cleanup threads (each triggered by an explicit "tidy up the memory" style user command) account for ~97% of organic layer traffic and 100% of forget/import/reflect. Not a single spontaneous "work is done, let me tidy up" was found.
• 157 of 180 organic layer calls come from one user-commanded cleanup thread.
• reflect promises itself a cron ("Good for a periodic tidy pass (e.g. a cron)") — no cron exists; reviews rot for weeks (one recorded closure: a review "had been open for 5 weeks", closed as stale).
• The instruction "Move things to archival when they stop changing" is tied to an unobservable moment: "stop changing" is only visible between sessions, and neither the verb nor the moment is named.
• A lost-arc example: a five-episode feature arc (research → impl → merged → deployed-to-prod-verified), the final episode literally reporting production verification — zero consolidation, all 5 nodes operational forever.

Proposal

1. Use the natural end-of-arc moments already present in traffic: done (14), close_review (8), confirm/refute — their responses should hint "arc closed — N linked operational episodes converge here; synthesise → settle?" (facilitator style, non-blocking).
2. An exit ritual in SKILL.md symmetric to re-entry, plus a ready Stop/PreCompact hook recipe for Claude Code.
3. Build the promised reflect-on-cron into a setup recipe (scheduled task), since the description already advertises it.