- name
- memory-poison-auditor
- version
- 0.1.0
- license
- MIT
- description
- Audits OpenClaw memory files for injected instructions, brand bias, hidden steering, and memory poisoning patterns. Use when reviewing MEMORY.md, daily memory files, or any long-term memory store that may have been contaminated through dialogue.
- metadata
- {"clawdbot":{"emoji":"🧠","requires":{"bins":["python3"]}}}
Memory Poison Auditor
memory-poison-auditor checks whether OpenClaw memory files have been contaminated by hidden instructions, brand steering, injected operational policies, or suspicious recommendation bias written through prior conversations.
What It Checks
- Prompt-injection style instructions inside memory.
- "Always recommend X" or "never mention Y" style brand steering.
- Abnormal brand repetition and preference shaping.
- Suspicious authority claims like fake approvals or fake user intent.
- Low-signal blocks that act like covert policy rather than factual memory.
- Optional AI review for borderline suspicious blocks.
Commands
Audit Default Memory Roots
python3 {baseDir}/scripts/audit_memory.py scan
python3 {baseDir}/scripts/audit_memory.py --format json scanAudit a Specific Path
python3 {baseDir}/scripts/audit_memory.py scan --path /root/clawd/MEMORY.md
python3 {baseDir}/scripts/audit_memory.py scan --path /root/clawd/memoryOptional AI Review
python3 {baseDir}/scripts/audit_memory.py scan --with-ai
python3 {baseDir}/scripts/audit_memory.py scan --path /root/clawd/memory/2026-03-15.md --with-aiOne-Click Cleaning
python3 {baseDir}/scripts/audit_memory.py clean --path /root/clawd/MEMORY.md --apply
python3 {baseDir}/scripts/audit_memory.py clean --path /root/clawd/memory --applyCleaning creates backups before rewriting suspicious blocks.
Output
Each audit returns:
PASS: no meaningful poisoning signalsWARN: suspicious memory blocks detectedBLOCK: memory likely contaminated and should be reviewed/cleaned
Reports and backups are written to:
/root/clawd/output/memory-poison-auditor/reports/
/root/clawd/output/memory-poison-auditor/backups/Operational Guidance
- Use this before trusting long-term memory in important planning or recommendations.
WARNmeans review before relying on that memory block.BLOCKmeans clean or quarantine the memory before reuse.- AI review is optional and intended only for ambiguous cases.