[Feature] Memory hygiene — stale detection, contradiction flagging, obsolete archival

[rahilp]

Why
As the second brain grows, it accumulates noise. Old decisions get superseded, facts become stale, and duplicate-adjacent entries pile up. Right now there's no way to surface any of this automatically — the user has to manually notice and forget things. That doesn't scale.

What this covers

• Stale memory detection: flag entries that haven't been recalled in N days and contain time-sensitive language ("we're planning to", "next week", "current version is")
• Contradiction detection: when a new memory is stored, check if it contradicts an existing one (e.g. "auth token is X" stored twice with different values) and surface the conflict
• Obsolete memory archival: soft-delete or archive entries that have been explicitly superseded, rather than hard-deleting them
• Duplicate clustering: go beyond the current 0.95 block threshold and surface clusters of highly similar entries for the user to review and consolidate
  Implementation notes
• Stale detection can run as a scheduled Cloudflare Worker (cron trigger)
• Contradiction detection could be a post-store step using the AI binding to compare new content against top recall results
• Archival needs a new status column in D1 (active, archived, superseded) and a filter in the recall query
• Web UI needs an "Inbox" or "Review" view to surface flagged memories
  Related
• Builds on existing duplicate detection (0.85/0.95 thresholds)
• Pairs well with semantic compression ([Feature] Semantic compression — rolling summaries and project digests #9 )

[rahilp]

Contradiction detection is now shipped (merged via #61 and #64).

Remaining scope for this issue:

• Stale memory detection (cron-based, flag entries with time-sensitive language not recalled in N days)
• Obsolete memory archival (soft-delete / status column: active, archived, superseded)
• Duplicate clustering (surface clusters of similar entries in a review UI)

Updating scope to reflect what's left.