Self Improving Agent
ReviewAudited by ClawScan on May 10, 2026.
Overview
This is mostly a disclosed self-improvement logger, but it can persist and promote learnings into future-session instruction files without clear approval or retention boundaries.
Install this only if you want the agent to maintain persistent learning files. Before enabling hooks or promotions, decide where learnings may be written, require approval for changes to agent instruction files, and periodically review or delete stored entries.
Findings (5)
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.
Future agent behavior may be changed by stored learnings, including incorrect or sensitive information captured from prior work.
The skill tells the agent to move learnings into persistent instruction/context files. If a mistaken, sensitive, or malicious correction is promoted, it can influence future sessions.
Broadly applicable learning | Promote to `CLAUDE.md`, `AGENTS.md`, and/or `.github/copilot-instructions.md`
Require explicit user approval before promoting learnings into CLAUDE.md, AGENTS.md, SOUL.md, TOOLS.md, or Copilot instructions; keep entries scoped, reviewable, and easy to remove.
A bad or over-broad learning could redirect how the agent behaves, delegates work, or uses tools in future sessions.
The promoted workspace files are injected as prompt context, so logged behavioral or workflow guidance can become authoritative instructions in later sessions.
OpenClaw uses workspace-based prompt injection ... Context is injected from workspace files at session start
Treat promoted learnings as prompt changes: review diffs, avoid copying untrusted text verbatim, and separate factual notes from behavioral instructions.
If enabled, hook scripts run automatically during agent workflows.
The optional hook setup runs local shell scripts on prompt or tool-use events. The included scripts only print reminders, but they still execute with the agent environment's permissions.
"command": "./skills/self-improvement/scripts/activator.sh"
Enable hooks only if you want automatic reminders, inspect the scripts first, and keep hook paths limited to this skill's reviewed files.
Project or conversation details could be shared into other sessions if these tools are used carelessly.
The documentation describes cross-session transcript reading, messaging, and sub-agent spawning as part of sharing learnings. This is disclosed and purpose-aligned, but it can move context across session boundaries.
sessions_history ... Read transcript from another session; sessions_send ... Send message to another session; sessions_spawn ... Spawn a background sub-agent
Share only the minimum learning needed, avoid sensitive transcript content, and confirm the target session before sending.
Users may not be sure which upstream repository or maintainer they are trusting.
The documented manual install repository differs from the supplied homepage and registry identity, creating a provenance ambiguity even though the included code is simple.
git clone https://github.com/peterskoett/self-improving-agent.git ~/.openclaw/skills/self-improving-agent
Install from a verified source, compare the reviewed files with the repository you install from, and avoid enabling hooks from unverified copies.