Self Improving Agent
AdvisoryAudited by Static analysis on Apr 30, 2026.
Overview
No suspicious patterns detected.
Findings (0)
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.
A mistaken, sensitive, or malicious learning could become future guidance for the agent and affect later work.
The skill tells the agent to turn learnings into persistent prompt/memory files that are reused across sessions. The visible instructions do not require user review or validation before promotion.
"important learnings get promoted to project memory"; "Broadly applicable learning | Promote to `CLAUDE.md`, `AGENTS.md`, and/or `.github/copilot-instructions.md`"; "OpenClaw injects these files into every session"
Keep `.learnings/` as non-authoritative notes by default. Require explicit user approval before editing `CLAUDE.md`, `AGENTS.md`, `SOUL.md`, `TOOLS.md`, `MEMORY.md`, or Copilot instructions, and periodically audit those files.
Private context or untrusted instructions from one session could be exposed to or reused by another session.
The skill promotes cross-session transcript access, message passing, and sub-agent spawning, but does not define permission checks, sanitization rules, or data boundaries for what may be shared.
"sessions_history — Read another session's transcript"; "sessions_send — Send a learning to another session"; "sessions_spawn — Spawn a sub-agent for background work"
Use cross-session tools only with explicit user approval. Share minimal sanitized summaries rather than full transcripts, and treat imported learnings as untrusted until verified.
If enabled, future sessions will automatically receive self-improvement reminders.
The optional hook persistently modifies bootstrap context after being enabled. The supplied handler only injects a reminder, which is disclosed and purpose-aligned.
"Fires on `agent:bootstrap`" and "Adds a reminder block to check `.learnings/` for relevant entries"
Enable the hook only if you want recurring reminders, and disable it through OpenClaw hook settings if it becomes intrusive.
Users may have less certainty that the reviewed package identity matches the registry entry they intended to install.
The embedded metadata differs from the supplied registry metadata, which lists slug `self-improving-agent-zh` and version `1.0.0`. This is a provenance/identity ambiguity, not evidence of malicious code.
"slug": "self-improving-agent", "version": "3.0.4"
Verify the package source and publisher before enabling hooks or allowing persistent memory-file edits.