Remember

Memory hygiene workflow to classify, deduplicate, and promote durable knowledge across local memory layers.

SKILL.md

GitHub

Remember

Review memory-like artifacts and propose what should be kept, promoted, merged, or removed. This keeps guidance accurate as the project evolves.

Use when

  • Memory instructions are drifting or conflicting.
  • Team wants to clean stale guidance from old sessions.
  • You need to promote stable patterns into durable docs.

Do not use when

  • There is no meaningful memory/context to review.
  • User only needs a one-off answer (no memory maintenance).
  • You are asked to auto-edit memory layers without confirmation.

Input contract

  • review_scope (optional): workspace_docs, session_logs, both (default both).
  • apply_changes (optional): default false (proposal mode first).
  • priority (optional): conflicts, duplicates, promotions, all (default all).
  • target_files (optional): explicit files to inspect first.

Parsing rules (priority order)

  1. Use explicit structured args when provided.
  2. Else infer from user phrasing:
    • "clean memory", "dedupe instructions" -> priority=duplicates,conflicts
    • "promote stable conventions" -> priority=promotions
    • "just review, don't edit" -> apply_changes=false
  3. Defaults:
    • review_scope=both
    • apply_changes=false
    • priority=all

Canonical workflow

  1. Collect sources
    • Workspace memory docs (for example: IDENTITY.md, SOUL.md, MEMORY.md, USER.md, AGENTS.md) and session evidence when available.
  2. Classify entries
    • durable convention, temporary note, duplicate, outdated, conflicting, ambiguous.
  3. Generate proposals
    • promotions (move to durable location), cleanup (remove/merge duplicates), conflict resolution (which version wins + rationale).
  4. Present grouped report
    • Promotions, Cleanup, Conflicts, Ambiguous, No action.
  5. Apply only with approval
    • If apply_changes=true, execute approved changes and re-validate consistency.

Guardrails

  • Default to proposal mode; do not mutate without explicit approval.
  • Never delete content that has unresolved ambiguity.
  • Prefer reversible edits and clear provenance notes.
  • Keep sensitive values out of memory docs; redact before proposing promotion.
  • State uncertainty explicitly when evidence is insufficient.

Output format

Always return:

  • status: proposed | applied | needs_user_input
  • review_scope: effective scope
  • promotions: list with destination + rationale
  • cleanup: duplicates/outdated removals or merges
  • conflicts: contradiction list + recommended resolution
  • ambiguous: items requiring user decision
  • next_action: required unless status=applied

Example

  • review_scope="both" priority="all" apply_changes=false