memory-hygiene

Pass

Audited by Gen Agent Trust Hub on Mar 5, 2026

Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: Instructions specify destructive file system operations (rm -rf ~/.clawdbot/memory/lancedb/) to clear local database directories as part of the cleaning process.\n- [COMMAND_EXECUTION]: Suggests establishing a persistence mechanism by adding a recurring monthly cron job for automated memory maintenance tasks.\n- [DATA_EXFILTRATION]: Documentation encourages storing sensitive metadata, such as the locations of credentials and contact details, within the agent's long-term vector memory, which increases the risk of accidental data exposure.\n- [PROMPT_INJECTION]: Identifies an indirect prompt injection surface where unvalidated user data is ingested via the memory_store capability and stored in the agent's context.\n
  • Ingestion points: SKILL.md (via the memory_store command parameters).\n
  • Boundary markers: Absent for the stored text content.\n
  • Capability inventory: Shell command execution (rm -rf), system service management (clawdbot gateway restart), and configuration modification (gateway action=config.patch).\n
  • Sanitization: No input validation or filtering is specified for data before it is committed to memory.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 5, 2026, 07:56 AM