Memory sanitize
Sanitize a memory directory so it is safe to share externally.
Contract
| Field |
Bound contract |
| Trigger |
The user asks to sanitize memory for sharing, redact PII, or scan memory for credentials. |
| Authority |
Reversible local: writes only a fresh /tmp/memory-sanitized-<timestamp> directory; rollback is undo. No remote mutation. Never modifies originals. |
| Side effect |
Create redacted copies of top-level Markdown memory files under a new temporary directory. Rollback is deletion of that directory. |
| Done |
A temporary directory containing redacted copies, a JSON report of redactions, and a user-facing diff review. If any Tier-1 credential is found in a source file, stop before generating copies and report the affected files. |
Inputs
- Memory directory (required): the directory to sanitize, supplied explicitly or resolved from the current project by
../memory-update/scripts/resolve-paths.sh memory_dir. If MEMORY_UPDATE_SKILL_SCRIPTS is set, use $MEMORY_UPDATE_SKILL_SCRIPTS/resolve-paths.sh instead.
- Requires the sibling skill memory-update in the same plugin.
- Optional:
MEMORY_DIR may override resolution after path validation.
- Process only top-level
*.md files. Do not read session histories or nested Markdown files.
- Use a fresh destination named
/tmp/memory-sanitized-<timestamp>; it must not already exist.
Procedure
- Resolve source and validate destination freshness. Run
../memory-update/scripts/resolve-paths.sh memory_dir to resolve the source directory; the resolver rejects a missing directory, control bytes, or shell metacharacters, and refuses a memory directory git would track. Confirm the destination is a fresh path under /tmp before any write. Done when: the source resolves and the destination is confirmed fresh.
- Scan sources for Tier-1 credentials. Run
scripts/sanitize-memory.py --scan-only <memory-dir> to check every top-level Markdown file for Tier-1 credential patterns (OpenAI, GitHub, AWS, Slack, bearer-token, ECR). If any Tier-1 credential is found, stop before generating any copies and report the affected files. The user must remediate the originals before a new run. Done when: the scan completes with zero Tier-1 hits, or the run stops with affected files reported.
- Generate redacted copies. Run
scripts/sanitize-memory.py <memory-dir> <destination> to produce redacted copies of all top-level Markdown files. Tier-2 redactions (home paths, email addresses, session IDs, dates older than 30 days) are applied in the copies. Done when: the JSON report and redacted copies are generated.
- Compute and display per-file diffs. Compare each reported source with its reported destination and show the complete diff. Verify that the report accounts for every processed top-level Markdown file and states redaction counts. Done when: every processed file has a complete diff and the report accounts for all files.
- Classify output as safe for human review. Report the generated directory and ask the user to review the displayed diff before sharing. Leave originals unchanged; the generated directory is a disposable review artifact, not a replacement memory store. Done when: the generated directory is reported and the user is asked to review the diff.
Failure and recovery
- Invalid source or destination: stop before sanitization. Report the rejected path and reason; choose a new timestamp only for a destination collision.
- Memory directory git would track: stop before snapshot or mutation and return
blocked: MEMORY_DIR is inside version control.
- Tier-1 credential detected in source: stop before generating copies. Report the affected source files and credential classes. Require manual remediation of the originals before a new run. Never approve or publish copies when a credential is present in the source.
- Sanitizer execution failure: return
blocked: sanitization or proof incomplete, including the command failure and any generated paths. Do not infer missing results or claim done.
- Nested Markdown: report that nested files were skipped and classify the result as partial rather than claiming the directory was fully sanitized.
- Originals require no rollback because they are never written. To roll back local output, delete only the named generated
/tmp/memory-sanitized-<timestamp> directory after preserving any report the user needs.
Output
Return the destination path, the sanitizer JSON report, the original-to-copy diff for every processed file, warnings for skipped nested files, and one terminal classification: sanitized for human review, blocked: credential detected, or blocked: sanitization or proof incomplete.
1---2name: memory-sanitize3description: Use when the user asks to sanitize memory for sharing, redact PII, or scan memory for credentials. Not for auditing memory: use memory-clean.4---56# Memory sanitize78Sanitize a memory directory so it is safe to share externally.910## Contract1112| Field | Bound contract |13|---|---|14| Trigger | The user asks to sanitize memory for sharing, redact PII, or scan memory for credentials. |15| Authority | Reversible local: writes only a fresh `/tmp/memory-sanitized-<timestamp>` directory; rollback is undo. No remote mutation. Never modifies originals. |16| Side effect | Create redacted copies of top-level Markdown memory files under a new temporary directory. Rollback is deletion of that directory. |17| Done | A temporary directory containing redacted copies, a JSON report of redactions, and a user-facing diff review. If any Tier-1 credential is found in a source file, stop before generating copies and report the affected files. |1819## Inputs2021- **Memory directory** (required): the directory to sanitize, supplied explicitly or resolved from the current project by `../memory-update/scripts/resolve-paths.sh memory_dir`. If `MEMORY_UPDATE_SKILL_SCRIPTS` is set, use `$MEMORY_UPDATE_SKILL_SCRIPTS/resolve-paths.sh` instead.22- Requires the sibling skill memory-update in the same plugin.23- Optional: `MEMORY_DIR` may override resolution after path validation.24- Process only top-level `*.md` files. Do not read session histories or nested Markdown files.25- Use a fresh destination named `/tmp/memory-sanitized-<timestamp>`; it must not already exist.2627## Procedure28291. **Resolve source and validate destination freshness.** Run `../memory-update/scripts/resolve-paths.sh memory_dir` to resolve the source directory; the resolver rejects a missing directory, control bytes, or shell metacharacters, and refuses a memory directory git would track. Confirm the destination is a fresh path under `/tmp` before any write. Done when: the source resolves and the destination is confirmed fresh.302. **Scan sources for Tier-1 credentials.** Run `scripts/sanitize-memory.py --scan-only <memory-dir>` to check every top-level Markdown file for Tier-1 credential patterns (OpenAI, GitHub, AWS, Slack, bearer-token, ECR). If any Tier-1 credential is found, stop before generating any copies and report the affected files. The user must remediate the originals before a new run. Done when: the scan completes with zero Tier-1 hits, or the run stops with affected files reported.313. **Generate redacted copies.** Run `scripts/sanitize-memory.py <memory-dir> <destination>` to produce redacted copies of all top-level Markdown files. Tier-2 redactions (home paths, email addresses, session IDs, dates older than 30 days) are applied in the copies. Done when: the JSON report and redacted copies are generated.324. **Compute and display per-file diffs.** Compare each reported source with its reported destination and show the complete diff. Verify that the report accounts for every processed top-level Markdown file and states redaction counts. Done when: every processed file has a complete diff and the report accounts for all files.335. **Classify output as safe for human review.** Report the generated directory and ask the user to review the displayed diff before sharing. Leave originals unchanged; the generated directory is a disposable review artifact, not a replacement memory store. Done when: the generated directory is reported and the user is asked to review the diff.3435## Failure and recovery3637- Invalid source or destination: stop before sanitization. Report the rejected path and reason; choose a new timestamp only for a destination collision.38- Memory directory git would track: stop before snapshot or mutation and return `blocked: MEMORY_DIR is inside version control`.39- Tier-1 credential detected in source: stop before generating copies. Report the affected source files and credential classes. Require manual remediation of the originals before a new run. Never approve or publish copies when a credential is present in the source.40- Sanitizer execution failure: return `blocked: sanitization or proof incomplete`, including the command failure and any generated paths. Do not infer missing results or claim done.41- Nested Markdown: report that nested files were skipped and classify the result as partial rather than claiming the directory was fully sanitized.42- Originals require no rollback because they are never written. To roll back local output, delete only the named generated `/tmp/memory-sanitized-<timestamp>` directory after preserving any report the user needs.4344## Output4546Return the destination path, the sanitizer JSON report, the original-to-copy diff for every processed file, warnings for skipped nested files, and one terminal classification: `sanitized for human review`, `blocked: credential detected`, or `blocked: sanitization or proof incomplete`.
OutlineDriven