Self-Learn

Continuous self-improvement through learning from corrections and task self-evaluation. Use when: (1) User corrects the agent (No that is wrong, Actually, I...

Security Scan

Instruction-only skill that logs agent corrections and self-evaluations to a local file and the platform memory; behavior is coherent with its description but it can accidentally persist sensitive user data if not carefully used.

Purpose & Capability

The skill's name/description (self-learning from corrections and self-evaluation) matches its instructions: append human-readable entries to memory/corrections.md and call the platform memory_store API. It does not request unrelated credentials, binaries, or system paths.

Instruction Scope

Instructions are focused on creating/appending memory/corrections.md, calling memory_store and memory_recall, and returning recent entries. This stays within the stated purpose. Caveat: the SKILL.md relies on the agent to follow a rule ('No secrets — never log credentials...') which is a behavioral constraint, not an enforced technical control; the skill will persist whatever the agent is told to store unless the platform enforces filtering.

Install Mechanism

No install spec and no code files — instruction-only. Nothing is downloaded or written by an installer. Lowest-risk mechanism.

Credentials

The skill declares no required env vars or credentials, which is consistent. It calls memory_store / memory_recall (platform functions) and writes a workspace file; ensure the platform memory backend (LanceDB) and workspace are trusted. The absence of declared credentials is normal, but memory_store may use implicit platform credentials — confirm what the memory backend stores/retains.

Persistence & Privilege

always:false and user-invocable:true. The skill requests local persistence of logs in the agent workspace and platform memory but does not request elevated or always-on privileges. It does not modify other skills or global agent settings.

Assessment

This skill is coherent and low-risk in structure: it writes correction/lesson entries to a workspace file (memory/corrections.md) and uses the platform memory APIs. Before installing: (1) confirm where the platform memory_store (LanceDB) persists data and its retention/ACLs, (2) decide or enforce a policy to prevent accidental logging of secrets or PII (the skill's 'No secrets' rule is not a technical guard), (3) review and periodically purge or restrict access to the corrections.md file, and (4) test the skill in an isolated workspace to verify memory_store behavior. If you need stronger guarantees about not storing sensitive data, add content-filtering or explicit redaction steps before writing to memory.

Self-Learn Skill

Two-part continuous improvement loop: learn from corrections, learn from self-evaluation.

On Every Task Completion

After finishing any task, briefly self-evaluate (1–3 sentences internally):

• Did the output meet the intent?
• What could be better next time?
• Is this a repeatable pattern worth storing?

If yes → log it. If it was routine with no issues → skip logging (don't over-log).

Log format:

CONTEXT: [task type]
REFLECTION: [what I noticed]
LESSON: [what to do differently / keep doing]

Append to memory/corrections.md under today's date. Also call memory_store with:

• category: decision
• importance: 0.75
• text: [LESSON] <lesson text> + relevant keywords

On User Corrections

Trigger phrases (detect these):

• "No, that's wrong / not right"
• "Actually..." / "I prefer..." / "Remember that I..."
• "Stop doing X" / "Why do you keep..."
• "I told you before..." / "Always do X"

When triggered:

1. Acknowledge the correction briefly
2. Append to memory/corrections.md under ## Corrections with today's date
3. Call memory_store with:
   • category: preference (style/tone) or decision (behaviour/approach)
   • importance: 0.85
   • text: [CORRECTION] <what was wrong> → <correct behaviour> + keywords
4. Recall to verify it stored correctly

Log format:

[YYYY-MM-DD] CORRECTION: <what was wrong> → <correct behaviour>

memory/corrections.md Structure

# Corrections & Learnings Log

## Corrections
[YYYY-MM-DD] CORRECTION: ...

## Self-Evaluations
[YYYY-MM-DD] CONTEXT: ... | LESSON: ...

Create the file if it doesn't exist.

On "What have you learned?" / "Show my patterns"

Read memory/corrections.md and show last 10 entries. Also memory_recall with query "CORRECTION OR LESSON" for LanceDB results.

Installation on a New OpenClaw

1. Copy skills/self-learn/ into your workspace skills/ folder — skill activates automatically
2. Create memory/corrections.md (copy from references/corrections-template.md)
3. Optionally update AGENTS.md skill table for easy reference

Rules

• Don't over-log — skip routine tasks with no notable outcome
• Atomic entries — one lesson per entry, under 100 words
• Keywords matter — include domain keywords in memory_store text for good recall
• No secrets — never log credentials, personal data, or sensitive info
• Corrections always log — user corrections are always worth storing (importance ≥ 0.85)