Memory Self-Heal
PassAudited by ClawScan on May 1, 2026.
Overview
This instruction-only skill is coherent and safety-bounded, but it relies on reading and writing workspace memory/logs and may retry actions automatically.
Before installing, make sure your workspace memory and logs do not contain secrets, and be comfortable with the agent performing up to three bounded recovery attempts before escalating. Review the persistent self-heal entries occasionally so incorrect or sensitive fixes are not reused later.
Findings (2)
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
The agent may take limited corrective actions without asking each time, which can save effort but should still be watched for high-impact tasks.
The skill allows the agent to retry actions and change fallback paths during a task, but it also sets a retry limit and requires validation.
Apply best-known fix from memory for same class/signature - Re-run the smallest validating action ... - Max 3 retries per blocker signature per task
Use this skill where bounded retries are acceptable, and require explicit confirmation before any destructive, account-changing, or production-impacting action.
Past logs or memory entries may affect the agent's future troubleshooting decisions, and sensitive data in those files could be brought into the agent's context.
The skill reads prior memory/tasks/logs and writes persistent self-heal entries, so stale, sensitive, or poisoned context could influence future runs.
Scan these in order; skip missing paths silently: 1. `memory/` ... 2. `tasks/` or queue files 3. runtime logs / channel logs ... Append one concise entry after each self-heal cycle
Keep memory and logs free of secrets, review self-heal writebacks periodically, and avoid treating untrusted memory/log content as authoritative instructions.