| name | memory-sweep |
| description | Periodic review of the memory system at ~/.claude/memory/ for duplicates, redundancies, conflicts, and stale entries. Use when the user asks to audit, clean up, or review memory. Presents findings with enforcement trade-off context — never retires anything without user confirmation. Scoped to the agent memory files under ~/.claude/memory/; for the same kind of duplicate/stale/conflict audit over the knowledge base at /path/to/works/for/you/knowledge_base/ use kb-sweep instead. |
Memory Sweep
Reviews all memory files and presents a structured findings report. The lead agent runs this inline — no subagent needed while memory is small (<20 files). Revisit this design when memory grows past that threshold.
Step 1 — Inventory
Read all files in ~/.claude/memory/ and ~/.claude/CLAUDE.md.
Note for each memory file: name, type (user/feedback/project/reference), and a one-line summary of what it enforces or describes.
Step 2 — Check for duplicates
Flag any two memory entries that enforce substantially the same principle or describe the same fact.
Step 3 — Check for redundancy with CLAUDE.md
Flag any memory entry whose principle is already stated verbatim (or near-verbatim) in ~/.claude/CLAUDE.md.
Why this matters: Both memory and CLAUDE.md are loaded every session — true duplicates between them are the only case where a memory is genuinely redundant from an enforcement standpoint. A memory that overlaps with a skill is NOT redundant — skills are loaded on invocation only, not every session.
Step 4 — Check for conflicts
Flag any two memory entries that give contradictory guidance.
Step 5 — Check for staleness
Flag any entry that appears stale — regardless of type. Signs of staleness:
- References a project, deadline, or state that has clearly passed
- References a path, repo, or tool that no longer exists
- Describes a behavior or workflow that has since been superseded
Step 6 — Present findings
Present a structured report:
Memory Sweep — <date>
Files reviewed: N
Total lines (MEMORY.md): N / 200 cap
DUPLICATES
- [file A] and [file B]: <what overlaps>
REDUNDANT WITH CLAUDE.md
- [file]: <which CLAUDE.md section covers it>
⚠ Retiring this would have no enforcement impact (same load scope)
REDUNDANT WITH SKILL
- [file]: <which skill mentions it>
⚠ Retiring this would WEAKEN enforcement — skill only loads on invocation
→ Recommend: keep in memory unless user explicitly accepts weaker enforcement
CONFLICTS
- [file A] vs [file B]: <what conflicts>
STALE
- [file]: <why stale>
NO ISSUES
- [file]: ok
Step 7 — Confirm before acting
For each flagged entry, state the enforcement trade-off explicitly and ask the user whether to retire, merge, or keep it.
Never delete or modify a memory file without user confirmation.
When to upgrade to an agent
If ~/.claude/memory/ grows past ~20 files, or reading all files starts consuming meaningful lead context, promote this to an agent pattern (like kb-sweep) where a specialized agent does the review and returns a compact findings report.
