What problem does it solve?
This Skill addresses the "stale-but-confident" failure mode in AI memory systems, where verified information silently drifts over time but remains highly trusted during retrieval, leading to incorrect or outdated outputs from AI agents. This issue is especially prevalent in online-first harnesses that rely on frequently updated live external resources like GitHub repositories, arXiv records, and DOIs.
Core Features & Use Cases
- Automated Staleness Scanning: Classifies memory entries by type (project, reference, feedback, user) and applies type-specific staleness thresholds to flag outdated content without manual review.
- Live Re-verification: Automatically checks stale external references via GitHub CLI or web fetch to confirm if content has drifted, updating verification timestamps for valid entries.
- Safe Archival Proposals: Flags unused, superseded memory entries for archival with mandatory pre-archive snapshots to prevent accidental data loss, requiring human confirmation for all destructive actions.
- Use Case: For a team using Claude Code across multiple software projects, this skill runs automatically during harvest loops to catch outdated GitHub repository statuses or obsolete operating rules before they cause incorrect work or wasted effort.
Quick Start
Invoke the memory-hygiene skill to scan all your project memory entries for staleness and receive a list of proposed updates, re-verifications, or archival candidates.
