Part of #534 (Phase 3 EPIC). Pairs with #538 (loop). Without this, auto-extraction is a liability — with it, the system gets less annoying over time instead of more.

Thesis

Hooks alone don't solve memory hygiene; they're the trigger, not the mechanism. The actual mechanism is a small set of patterns working together: decay scoring drives ranking down for unused memories, a periodic review subagent surfaces removal candidates, contradiction detection (#537) prevents silent conflicts, and a trivially-fast "drop this" gesture lets users prune in one keystroke when they see noise.

Current behavior

• Memories live forever once written.
• No access tracking.
• No review pass.
• No quick-drop UI.

Proposed change

Decay scoring — every memory carries last_accessed, access_count, and a per-type decay rate. Recall (#536) already uses these as signals; the same scores power a "below-threshold for N days" candidate set for proposed removal. Decay is type-aware: episodic decays fast (days), procedural slowly (months), constitutional not at all by default.

Periodic review subagent — fires from PostCompact hook (#538) and on a daily-on-startup cadence. Reads memory + access stats, proposes a review queue:

• N memories below decay threshold → drop?
• N pairs with high text overlap → merge?
• N contradictions outstanding (MEMORY: topology — graph edges + contradiction detection #537) → resolve?

Always proposes, never auto-applies. Reference-cc's /remember skill is exactly this pattern; it's what makes auto-extraction tolerable instead of annoying.

One-key drop in /memory UI — the cheapest hygiene mechanism. When a user sees a bad memory, single keystroke removes it. Builds the user-driven feedback loop the decay scoring depends on.

Promotion lifecycle — when an episodic memory gets accessed N times across M sessions, propose promotion to procedural. When procedural proves durable, propose promotion to constitutional. Promotions never automatic — proposed via review queue.

Open questions

• Default decay rates per type.
• Review cadence: daily? Per-session? Both?
• How to surface the review queue without being noisy — a numbered count in the status bar that opens on demand?
• Should "auto-apply trivial cases" ever be allowed (e.g., dedup of literal duplicates)?

Acceptance signals

• Memory store size stabilizes over time instead of growing monotonically.
• User-initiated drops trend down across releases (proxy for "memory feels right").
• Review queue is opt-in to view, never blocks user flow.
• Promotion proposals visible and approvable from /memory.
• A long-running test: 10K simulated turns + extraction events leaves the store under a configurable size cap.