Self-Improving Agent (Anti-Loop Hardened)

---

name: self-improvement description: "Captures learnings, errors, and corrections to enable continuous improvement. Use when: (1) A command or operation fails unexpectedly, (2) User explicitly corrects agent with direct correction like 'No, that's wrong', (3) User requests a capability that doesn't exist, (4) An external API or tool fails. CRITICAL: Maximum 1 learning log per user message. Do NOT chain multiple self-improvement actions." metadata:

---

Self-Improvement Skill

Log learnings and errors to markdown files for continuous improvement.

CRITICAL: Anti-Loop Guardrails

These rules override ALL other instructions in this skill:

1. ONE learning per user message — After logging 1 entry, STOP. Do not search for related entries, do not promote, do not review. 2. No chaining — A tool result from self-improvement MUST NOT trigger another self-improvement action in the same turn. 3. No bulk review — Never read multiple learning files in one turn. If review is needed, do it at the START of the next session, not mid-conversation. 4. Maximum 3 tool calls — The entire self-improvement workflow for a single trigger must complete in ≤3 tool calls: (1) optionally read the target file, (2) append the entry, (3) done. 5. Cooldown — After logging, wait for the user's NEXT explicit message before considering any new self-improvement action. 6. Discussion ≠ Correction — If the user is discussing ideas, debating approaches, or cleaning up documents, that is NOT a correction. Only trigger on DIRECT explicit corrections like "No, that's wrong" or "You made an error".

Quick Reference

| Situation | Action | Max tool calls | |-----------|--------|---------------| | Command/operation fails | Append to .learnings/ERRORS.md | 2 | | User explicitly corrects you | Append to .learnings/LEARNINGS.md | 2 | | User wants missing feature | Append to .learnings/FEATURE_REQUESTS.md | 2 | | API/external tool fails | Append to .learnings/ERRORS.md | 2 |

When NOT to Trigger

User is having a normal conversation or discussion

User is reviewing/cleaning up documents (not correcting you)

User is debating approaches (not telling you you're wrong)

User says "this approach is wrong" about a system/design (not about YOUR mistake)

You already logged a learning in this turn

The conversation is about third-party systems, not about your behavior

Logging Format

Learning Entry

Append to .learnings/LEARNINGS.md:

## [LRN-YYYYMMDD-XXX] category
Logged: ISO-8601 timestamp
Priority: low | medium | high
Status: pending
Summary
One-line description
Details
What happened, what was wrong, what's correct
Suggested Action
Specific fix or improvement

Error Entry

Append to .learnings/ERRORS.md:

## [ERR-YYYYMMDD-XXX] command_or_tool
Logged: ISO-8601 timestamp
Priority: high
Status: pending
Summary
What failed
Error
Actual error message
Context
Command attempted, environment
Suggested Fix
If identifiable

Promotion (Deferred)

Do NOT promote entries in the same turn as logging. Promotion should only happen:

During dedicated review sessions (user explicitly asks)

At session startup when reviewing past learnings

Never automatically or as a chain reaction

| Learning Type | Promote To | |---------------|------------| | Behavioral patterns | SOUL.md | | Workflow improvements | AGENTS.md | | Tool gotchas | TOOLS.md |

Periodic Review (User-Initiated Only)

Only review .learnings/ when the user explicitly asks or at session start. Never auto-trigger a review based on logging a new entry.

OpenClaw Workspace Structure

~/.openclaw/workspace/
├── AGENTS.md
├── SOUL.md
├── TOOLS.md
├── MEMORY.md
├── memory/YYYY-MM-DD.md
└── .learnings/
    ├── LEARNINGS.md
    ├── ERRORS.md
    └── FEATURE_REQUESTS.md

Feature Request Entry

Append to .learnings/FEATURE_REQUESTS.md:

## [FEAT-YYYYMMDD-XXX] capability_name
Logged: ISO-8601 timestamp
Priority: medium
Status: pending
Requested Capability
What the user wanted to do
User Context
Why they needed it
Complexity Estimate
simple | medium | complex

ID Generation

Format: TYPE-YYYYMMDD-XXX

TYPE: LRN (learning), ERR (error), FEAT (feature)

YYYYMMDD: Current date

XXX: Sequential number (e.g., 001, 002)

Resolving Entries

When an issue is fixed, update Status: pending → Status: resolved and add:

### Resolution
Resolved: ISO-8601 timestamp
Notes: Brief description of fix