memory-hygiene

Run a memory hygiene audit that reads MEMORY.md, checks line count against the 200-line truncation limit, verifies index links resolve, and flags inline content that belongs in topic files. Fixes are proposed in a report and applied only after your approval.

It organizes memory into tiers: T0 axioms (10-12 items, always loaded), T1 MEMORY.md indexes, T1.5 path-scoped phase rules, T2 topic files loaded on demand, and T3 archives searched via grep only. This avoids the Lost in the Middle attention degradation from bulk-loading large files.

Yes. Staleness detection greps the codebase for referenced files, functions, and libraries that no longer exist, flags relative dates like 'last week', and surfaces conflicts between project and global lessons. It also groups incident memories by PR or commit to find conflicting cause claims.

No. The skill follows an audit-then-fix workflow: it produces a structured report grouped by severity, then executes fixes only after user approval. Destructive actions like deleting feedback files or merging superseded rules always require explicit confirmation.

The label_audit.py script belongs to the separate session-handoff plugin, and plugin installs place it under the plugin cache rather than ~/.claude/skills/. The audit resolves it across all three install roots and reports a skipped check instead of silently passing when the script is absent.