Learning Capture

Capture user corrections, failed tool runs, recurring gotchas, and useful lessons as short structured notes. Use when a mistake or workflow learning should be remembered for future work. Keep entries redacted, human-reviewable, and project-scoped. Do not change agent rules or configuration without explicit user approval.

Audits

Pass

Install

openclaw skills install self-improving-agent1

Learning Capture

Use this skill to turn mistakes and corrections into small reusable notes.

The goal is simple: if the same issue appears again, the agent can check the note before repeating the mistake.

Core Rules

  • Write short notes, not raw transcripts.
  • Redact credentials, auth tokens, private messages, customer data, and full command output.
  • Prefer project-scoped .learnings/ files.
  • Do not edit AGENTS.md, SOUL.md, CLAUDE.md, hooks, or config unless the user explicitly approves.
  • Treat every promotion to long-term instructions as a proposal first.

When To Use

SituationAction
User corrects the agentAdd a correction note
Tool/command fails unexpectedlyAdd an error note with redacted summary
A project gotcha is discoveredAdd a learning note
Same issue repeatsLink the related notes and propose a durable rule
User requests a missing capabilityAdd a feature-request note

Setup

Create these files in the project root if they do not exist:

.learnings/LEARNINGS.md
.learnings/ERRORS.md
.learnings/FEATURE_REQUESTS.md

Use normal file-writing tools. Never overwrite existing notes.

Learning Note Format

## LRN-YYYYMMDD-001 — short-title

**Logged:** ISO timestamp
**Priority:** low | medium | high
**Status:** pending | resolved | proposed
**Area:** frontend | backend | infra | tests | docs | workflow

### Summary
One sentence.

### What happened
Short redacted context.

### Better next time
Concrete behavior to apply in future.

### Tags
- tag-one
- tag-two

Error Note Format

## ERR-YYYYMMDD-001 — short-title

**Logged:** ISO timestamp
**Priority:** low | medium | high
**Status:** pending | resolved

### Summary
What failed, in one sentence.

### Redacted Evidence
Minimal safe excerpt. No credentials, tokens, private data, or full logs.

### Likely Cause
If known.

### Fix / Next Step
Concrete action.

Feature Request Format

## FEAT-YYYYMMDD-001 — short-title

**Logged:** ISO timestamp
**Status:** proposed

### Requested Capability
What the user wanted.

### Why It Matters
What problem it solves.

### Possible Implementation
Short option list, if useful.

Promotion Rule

If a note is broadly useful:

  1. Propose the distilled rule to the user.
  2. Wait for approval.
  3. Only then update durable project or agent instructions.
  4. Mark the original note as proposed or resolved with a link to the approved location.

Output

When done, respond briefly:

Logged: .learnings/LEARNINGS.md#LRN-YYYYMMDD-001

If no durable learning exists, do nothing.