Claude Memory Curator

Claude Memory Curator

Goal

Audit Claude Code durable context as user-owned agent state: expose stale, unsafe, duplicated, over-broad, conflicting, misplaced, or unenforceable entries; propose better destinations; and route review, planning, persistence, and cleanup through one explicit contract.

Keep the subject scoped to Claude Code surfaces: CLAUDE.md, CLAUDE.local.md, .claude/rules/, user-level Claude rules, settings, hooks, managed policy evidence, and auto memory files. Do not treat this as a Codex, Cursor, Claude app, or Anthropic API Memory Stores curator.

When to use

• Use for review, placement, planning, or cleanup of the Claude Code durable surfaces named in the description.
• Use when those surfaces are stale, conflicting, sensitive, over-broad, ignored, or causing Claude to reuse old guidance.

When not to use

• Do not use for Codex, Cursor, Claude web/app, Anthropic API Memory Stores, generic prompt engineering, or repo-doc cleanup without Claude durable context.
• Keep review requests read-only. Treat managed policy, including /etc/claude-code/CLAUDE.md and managed-settings.json, as read-only unless exact edit authority is explicit.

Workflow selection

Always expose these workflows in this order. plan-run-cleanup-file is always first and Recommended:

Route from intent instead of adding an auto workflow:

• Direct review defaults to review-chat; explicit persistence selects review-file.
• Explicit cleanup without a delivery preference selects plan-run-cleanup-file.
• A clear direct request may select another matching route when delivery and execution intent are explicit.
• Agent-initiated activation may select only a relevant read-only route. Use review-file only when the existing task already requests persistence; never infer cleanup.
• A bare invocation, conflicting cues, or ambiguity about review versus cleanup, chat versus file, execution, target paths, or mutation authority exposes the table and asks the user to choose.
• A mutating route may be selected only when the user already requested cleanup of the identified Claude context scope.

Before substantive inspection, show the complete table plus Selected, Reason, target paths, write scope, expected artifacts, protected state, Plan-mode capability, and remaining separate approvals. If selection is unambiguous, announce it and proceed. If it is ambiguous, stop before inventory and ask.

Workflow selection does not authorize whole-file deletion, destructive recovery, managed-policy edits, paid or external actions, deployment, publication, or scope expansion.

Inputs to inspect

• Resolve the repo, Claude home, configured auto-memory directory, and only the applicable project/user instruction, rule, setting, hook, managed-policy, and memory surfaces.
• Inspect current repository evidence only as needed to verify a disputed claim.
• Load the placement, classification, conflict, enforcement, and safe-editing references below only when their decision is active. Load the report or plan asset whenever producing that artifact.

Workflow

Every route performs the same full-depth review before planning or cleanup:

1. Resolve the selected route, target repo, Claude home, auto-memory path, persistence path when applicable, and protected state.

2. Inventory Claude context without dumping contents:

   node scripts/inventory-claude-memory.mjs --repo .
   

   Add --claude-home PATH or --memory-dir PATH when the user provides non-default locations. Use --json when structured evidence is useful.

3. Run the redacted risk scanner when looking for sensitive, stale, broad, local, repo-specific, ignored, or conflict-prone entries:

   node scripts/scan-claude-memory-risks.mjs --repo . --json
   

   Exit code 1 means findings were found, not that the scan failed. Summarize counts and highest-signal redacted findings instead of pasting the full payload.

4. Parse .claude/rules/**/*.md frontmatter for paths so path-scoped guidance is not treated like always-loaded guidance.

5. Inspect visible settings for autoMemoryEnabled, autoMemoryDirectory, claudeMdExcludes, permission rules, and hooks when they affect placement.

6. Treat managed policy files and managed settings as higher-precedence read-only evidence by default.

7. Treat MEMORY.md as the auto-memory entrypoint. Treat topic files as on-demand detail unless the user asks to audit the whole memory directory.

8. Read context files in bounded chunks and redact sensitive values.

9. Extract one atomic claim per row. Split compound entries before classification.

10. Verify disputed claims against only the repo files needed for the dispute. Load references/conflict-resolution.md when precedence is unclear.

11. Assign exactly one primary classification per atomic claim: KEEP, KEEP BUT REWRITE, MOVE TO CLAUDE.md, MOVE TO CLAUDE.local.md, MOVE TO CLAUDE RULE, MOVE TO AUTO MEMORY TOPIC, MOVE TO AGENTS.md, MOVE TO REPO DOCS, MOVE TO SKILL, MOVE TO SETTINGS, MOVE TO HOOK, MOVE TO MANAGED POLICY, DELETE, or ASK USER.

12. Tag high-risk entries as useful context only: stale, duplicated, too-broad, too-specific, repo-specific, workflow, config, sensitive, conflicting, unenforced, managed-policy, or useful.

13. Add confidence (high, medium, or low) and a proposed action to every entry.

14. Produce the complete review before planning or editing. Route delivery must not reduce review depth.

Route execution

• review-chat: return the review in chat and create no durable curation report.
• review-file: persist the single curation record and make no context change.
• cleanup-chat: derive only high-confidence atomic actions from the completed review, back up every exact file to be changed, apply them, re-read changed sections, and report verification in chat. Create no durable curation report.
• cleanup-file: create the curation record before mutation; if persistence fails, stop. Then back up exact files, apply only high-confidence atomic actions, and complete the same record with execution and verification.
• plan-cleanup-chat and plan-cleanup-file: enter the Plan lifecycle, resolve the cleanup plan with the user, and stop after approval without changing Claude context.
• plan-run-cleanup-chat and plan-run-cleanup-file: enter the Plan lifecycle, resolve and approve the complete cleanup plan, recheck state, exit Plan mode, back up exact files, execute only the unchanged plan, and verify. Do not ask a generic second cleanup question after plan approval.

Direct cleanup (cleanup-chat or cleanup-file) is limited to high-confidence atomic edits, moves, or entry deletion in existing, editable, runtime-owned Claude context. Defer whole-file deletion, new context files, settings, hooks, managed policy, AGENTS.md, repository docs, skills, UI-only User/Team settings, uncertain schemas, medium/low-confidence changes, and any scope expansion. A plan-run route may execute broader curation changes only when the approved plan names each destination, write path, backup, rollback, and separate approval boundary.

Plan lifecycle

The four plan-* routes require native Plan mode when the host supports it:

1. Detect support before substantive planning.
2. If supported and active, plan there. If supported but inactive, or support is indeterminate, stop and ask the user to enter or confirm Plan mode.
3. Use an in-chat portable fallback only when native Plan mode is definitely unavailable.
4. Before execution, record plan approval, recheck target files and protected state, stop on material drift, and exit Plan mode before mutation.

Do not invoke claude-spec-interviewer inside this curation workflow. If findings require a broader durable rule, repository spec, or unresolved product decision, finish the selected curation route and offer the interviewer as a separate follow-up.

File delivery contract

File routes persist exactly one redacted curation record. Prefer an existing repository-native report location; otherwise use <repo>/.agent-reports/claude-memory-curation/<UTC timestamp>-<selection-id>.md. Create a new path without overwriting and keep all route output in that record.

The record contains Review, Plan, Execution Receipt, Deferred Work, Backup, and Verification. Use not applicable with a reason for phases the route does not perform. Create the record before mutation for cleanup-file and plan-run-cleanup-file; persistence failure blocks cleanup. Chat routes create no report file. Backup directories remain mandatory safety artifacts and do not count as curation reports.

Explicit --backup-root requires a stable non-sensitive --backup-root-alias; file routes persist the script-reported portable storage locator and <storage-locator>/backup-manifest.json. Report exact absolute backup and manifest paths only in non-persisted chat, never in repository artifacts.

Safety rules

• Do not inspect when the route is unresolved, and do not mutate unless the selected route and user request authorize cleanup of the exact target scope.
• Never silently delete, rewrite, truncate, or move Claude context files.
• Back up every exact file before an approved edit and report the backup path.
• Do not print full secrets, tokens, credentials, customer data, private identifiers, internal hostnames, or sensitive personal data.
• If secret-like data is found, redact values in output, identify file and line when possible, recommend removal, and recommend rotation for real credentials.
• If an auto-memory topic schema is unclear, do not edit it directly. Defer it in the current record or chat result.
• Treat CLAUDE.md and auto memory as context, not enforcement. Recommend settings or hooks for deterministic restrictions.
• Do not apply repo-specific assumptions globally. Prefer project CLAUDE.md, .claude/rules, AGENTS.md, or repo docs for repo rules.
• Do not run broad destructive commands.

References

Read only when needed:

• Placement and classification: references/context-surface-anatomy.md, references/classification-rubric.md.
• Conflicts and enforcement: references/conflict-resolution.md, references/settings-and-hooks.md.
• Safe mutation and examples: references/safe-editing-procedure.md, references/example-review-report.md.
• Output artifacts: assets/review-report-template.md, assets/cleanup-plan-template.md.

Scripts

Use only when needed. All scripts are non-interactive, use Node.js stdlib only, and accept --help.

node scripts/inventory-claude-memory.mjs [--repo PATH] [--claude-home PATH] [--memory-dir PATH] [--json]
node scripts/scan-claude-memory-risks.mjs [--repo PATH] [--claude-home PATH] [--memory-dir PATH] [--json] [--max-findings N]
node scripts/backup-claude-memory.mjs [--repo PATH] [--claude-home PATH] [--memory-dir PATH] [--backup-root PATH --backup-root-alias NAME] [--include PATH ...]

• Inventory is read-only. The scanner is read-only, redacts by default, bounds findings, and uses exit 1 for findings rather than execution failure.
• backup-claude-memory.mjs creates a no-clobber backup plus backup-manifest.json. Unredacted backup payloads and manifests stay outside Git worktrees; the backup scripts default to deterministic user state and reject an unsafe --backup-root before copying. One or more repeatable --include PATH values select exact-only mode; zero includes retain legacy context discovery. Selected paths and explicit discovery roots must exist and be readable; every symlink path component and legacy traversal error fails before root creation. A malformed discovered settings file or present invalid autoMemoryDirectory also fails before root creation. Only an absent key permits the documented derived auto-memory path. It does not edit or delete context files.

Output format

Start with the selected workflow, rationale, target paths, write scope, expected artifacts, protected state, Plan-mode state, persistence path or chat only, and remaining approvals.

Before producing a report, load and follow assets/review-report-template.md as the canonical heading and field contract. File routes copy that complete template into the one curation record; chat routes render only applicable sections in chat and create no report file. Populate every applicable field, use not applicable with a reason for skipped phases, and redact sensitive values.

Before edits, complete the review and decision tables. After edits, complete the same record's receipt, including Manifest reconciliation and unmatched paths; New paths (created-no-preimage) and rollback; Backup mode and manifest path; Backup integrity result; and the row schema | Changed path | Backup destination | Bytes | SHA-256 | Verification |.

Completion criteria

• One of the eight canonical workflows was selected from clear authority or resolved ambiguity; agent activation never inferred cleanup.
• Applicable context, .claude/rules metadata, settings, and auto-memory boundaries were inventoried or reported missing.
• Every atomic claim has one classification, risk tags, confidence, action, and higher-precedence conflict evidence when applicable.
• Deterministic restrictions route to settings/hooks; Plan and direct-cleanup boundaries remain satisfied.
• Chat/file delivery matches the canonical template contract, and every authorized edit has an exact-file backup, manifest reconciliation, re-read, and integrity result.

Failure modes

• Missing surfaces or auto-memory path: report what is unavailable and request /memory, claude --version, or an explicit --memory-dir only when needed.
• Managed policy or unknown schema: keep read-only and recommend/defer the manual action.
• Sensitive or conflicting content: redact; cite location and higher source; recommend removal/rotation or classify for rewrite, move, deletion, or user decision.
• Persistence, backup, or approved-state recheck failure: stop before mutation (or before further mutation), report the failure, and return to planning after drift.