Self-Improving Agent
PassAudited by ClawScan on May 1, 2026.
Overview
The skill is coherent and does not show exfiltration or destructive behavior, but it intentionally creates persistent learning context and optional reminders that users should keep reviewed and sanitized.
Install if you want a persistent self-improvement log for your agent. Keep secrets and raw transcripts out of learning entries, review anything promoted into agent instruction or memory files, and enable hooks or cross-session sharing only when you intentionally want that behavior.
Findings (4)
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.
Incorrect, overly broad, or sensitive entries could persist and influence future agent behavior.
The skill deliberately stores learnings and can promote them into files that future agent sessions may read as guidance.
important learnings get promoted to project memory ... Promote to `CLAUDE.md`, `AGENTS.md`, and/or `.github/copilot-instructions.md`
Review .learnings and promoted memory files periodically, redact sensitive details, and require user approval before promoting entries into long-term agent instructions.
Once enabled, the agent will receive recurring reminders to capture learnings, which can affect future session behavior and context size.
The optional OpenClaw hook automatically adds a self-improvement reminder during agent bootstrap when enabled.
Only handle agent:bootstrap events ... Inject the reminder as a virtual bootstrap file
Enable the hook only if you want automatic reminders, and disable it if the persistent prompt behavior becomes intrusive.
If used carelessly, cross-session tools could expose transcript contents or start background work beyond the current task.
The skill documents cross-session transcript access, messaging, and sub-agent spawning, while also instructing that these be used only in trusted environments with explicit user intent.
**sessions_history** — Read another session's transcript ... **sessions_send** — Send a learning to another session ... **sessions_spawn** — Spawn a sub-agent
Use cross-session features only with explicit user approval, send sanitized summaries instead of raw transcripts, and avoid sharing secrets or full command output.
Manual installs from an unpinned repository may differ from the reviewed artifact version.
The manual installation example pulls the latest repository contents without pinning a commit or release.
git clone https://github.com/peterskoett/self-improving-agent.git ~/.openclaw/skills/self-improving-agent
Prefer the registry install for the reviewed version, or pin and inspect a specific commit when installing manually.