什么是 学习引擎？

学习引擎是一个复杂的元层，旨在 AI 环境中实现持续改进原则。通过处理错误日志、自我评估数据和性能指标，它能识别导致失败或成功的模式。作为 Openclaw Skills 生态系统的重要组成部分，它确保一旦找到解决方案或识别出错误，相关知识就会被编纂成永久规则。

该技能将原始操作数据转化为可操作的智能，允许智能体随着时间的推移演化自身的逻辑和配置文件。它有效地弥合了静态指令与动态操作经验之间的差距，确保您的 Openclaw Skills 在每次执行中都变得更加健壮。通过自动化“不重复同样错误”的原则，它显著减少了技术债务和手动故障排除。

下载入口:https://github.com/openclaw/skills/tree/main/skills/mupengi-bot/learning-engine

安装与下载

1. ClawHub CLI

从源直接安装技能的最快方式。

2. 手动安装

将技能文件夹复制到以下位置之一

优先级：工作区 > 本地 > 内置

3. 提示词安装

将此提示词复制到 OpenClaw 即可自动安装。

学习引擎 应用场景

• 通过基于过去的失败生成验证规则，自动防止重复的 API 或上传错误。
• 通过学习历史性能数据和发布计划，优化社交媒体参与度。
• 通过识别并用直接 API 调用替换低效的浏览器自动化模式，减少 Token 消耗。
• 生成全面的周报，以审计 Openclaw Skills 框架内智能体的进展和学习里程碑。

1. 坚控多种数据源，包括 memory/errors/ 中的错误日志、性能 JSON 数据以及每周自我评估结果。
2. 执行模式分析，以区分一次性故障与系统性失败模式或成功触发因素。
3. 生成存储在 memory/learned-rules/ 目录中的结构化行为规则。
4. 自动将这些学到的教训注入相关的 SKILL.md 文件中，以永久更新智能体行为。
5. 发布结构化事件和周报，向用户和其他系统通知新获取的知识和更新后的 Openclaw Skills。

学习引擎 配置指南

要将学习引擎集成到您的环境中，请确保您的工作区遵循标准目录结构。使用以下命令为您的 Openclaw Skills 学习流水线初始化必要的目录：

# 创建所需的 memory 结构
mkdir -p memory/errors memory/learned-rules memory/learning memory/self-eval

# 确保您的智能体对 skills 目录具有写权限
# 配置 hook-engine 以在错误事件上触发 learning-engine

学习引擎 数据架构与分类体系

系统在多种格式中维护结构化的知识分类，以确保与其他 Openclaw Skills 的互操作性：

name: learning-engine
description: Auto-analyze mistake and success patterns and reflect in skills
author: ??? ??

learning-engine

System records mistakes and successes, automatically learns patterns to improve skills. Automates "don't repeat same mistake" principle.

Learning Sources

1. memory/errors/

Extract failure patterns from error logs

# memory/errors/2026-02-14.md

## 10:30 - insta-post failure
- Cause: PNG file upload → "Problem occurred" error
- Fix: Retry after JPG conversion → Success
- Lesson: Always convert to JPG before In@stagram upload

2. self-eval Results

Extract improvement points from weekly self-evaluation

# memory/self-eval/2026-W07.md

## This Week's Mistakes
- Too many browser snapshots (token waste)
- → Improvement: Call API directly via exec

## This Week's Successes
- 95% token savings with insta-cli v2 DM check

3. performance Data

Learn successful/unsuccessful patterns from performance tracking

{
  "insight": "Posts at 7-9 PM get +30% likes",
  "rule": "In@stagram posts recommended 19:00-21:00"
}

Auto Rule Generation

Convert learned patterns to rules:

Location: memory/learned-rules/

memory/
  learned-rules/
    instagram-posting.md
    browser-automation.md
    api-usage.md
    error-recovery.md

Rule Format

# In@stagram Posting Rules

## Rule #1: Always Convert to JPG
- **Situation**: Upload image to In@stagram
- **Failure Pattern**: PNG file → "Problem occurred"
- **Solution**: `convert input.png -quality 92 output.jpg`
- **Evidence**: 2026-02-10, 2026-02-14 error logs
- **Applied Skills**: insta-post, cardnews, social-publisher

## Rule #2: 1:1 Ratio Required
- **Situation**: In@stagram card news
- **Failure Pattern**: 16:9 horizontal → Cropped in feed
- **Solution**: Generate as 1024x1024 square
- **Evidence**: 2026-02-13 feedback
- **Applied Skills**: cardnews, nano-banana-pro

Inject Rules into Skills

Auto-add learned rules to relevant skill SKILL.md:

Location: skills/{skill-name}/SKILL.md

# insta-post

...

## Learned Lessons

### Image Processing
- ? Always convert to JPG (PNG causes errors)
- ? 1:1 ratio required (1024x1024 recommended)
- ? File size < 8MB

### Timing
- ? Posts at 19:00-21:00 get +30% engagement
- ? Avoid early morning posts

### Automation
- ? Call API via exec (0 snapshots)
- ? Minimize browser automation

Weekly Learning Report

Auto-generated every Monday:

Location: memory/learning/weekly-YYYY-Www.md

# 2026-W07 Learning Report

## New Learnings (5)

1. **In@stagram PNG Ban**
   - 3 mistakes → Rule created
   - Applied: insta-post, cardnews

2. **Token Saving: exec > Browser**
   - v1: 5 snapshots → v2: 1 exec
   - 95% savings

3. **Optimal Posting Time**
   - 19:00-21:00 +30% likes

4. **Brand Tone Effect**
   - ??? tone +40% engagement

5. **Auto Error Recovery**
   - browser-dependent failure → Browser restart

## Applied Skills
- insta-post (2 rules)
- cardnews (1 rule)
- performance-tracker (1 insight)

## Next Week Goals
- [ ] Build A/B testing system
- [ ] Add 3 auto-recovery patterns

Event Publishing

Publish event when learning complete:

Location: events/lesson-learned-YYYY-MM-DD.json

{
  "timestamp": "2026-02-14T23:00:00Z",
  "source": "learning-engine",
  "new_rules": 2,
  "updated_skills": ["insta-post", "cardnews"],
  "summary": "Learned 2 In@stagram image rules"
}

hook-engine Integration

• on-error hook: Error occurs → Record to memory/errors/ → learning-engine analysis
• post-hook (self-eval): After weekly evaluation → Update learning rules
• post-hook (performance): After collecting performance data → Learn patterns
• scheduled hook: Every Monday → Generate weekly learning report

Learning Pipeline

Error occurs
  ↓
Record to memory/errors/
  ↓
learning-engine analysis
  ↓
Extract patterns + Create rules
  ↓
Save to memory/learned-rules/
  ↓
Auto-update relevant skill SKILL.md
  ↓
Publish event (lesson-learned)
  ↓
Reflect in weekly report

Trigger Keywords

• "what did I learn"
• "learning"
• "lessons"
• "mistake patterns"
• "improvements"
• "learning report"
• "add rule"

Usage Examples

"What did I learn this week?"
→ Generate weekly learning report

"Organize In@stagram posting mistake patterns"
→ Analyze memory/errors/ + Create rules

"Learn from performance data"
→ Extract successful patterns + Update rules

Auto-improvement Examples

Before (Pre-learning)

In@stagram post fails → Manually convert to JPG → Retry
(Repeat every time)

After (Post-learning)

Execute insta-post → Auto-check/convert JPG → Success
(Rule injected into SKILL.md)

Meta Learning

learning-engine itself also learns:

• "Which rules are used most?"
• "Which skills improve most?"
• "Which areas have slow learning?"

Meta Learning Report: memory/learning/meta-YYYY-MM.md

Future Improvements

• Rule conflict detection (Rule A vs Rule B)
• Rule confidence score (based on usage frequency)
• Auto A/B testing (rule validation)
• Share learning with other agents

?? Built by ??? — Mupengism ecosystem skill