agent-self-evolve（Agent 自我进化）

Self-Evolve 🧬

A self-improvement system that turns every interaction into a learning opportunity. Three loops: real-time capture → daily consolidation → weekly deep review.

Quick Start

1. Initialize Evolution Files

Run the setup script to create the file structure:

bash <skill_dir>/scripts/setup-evolution.sh

This creates:

memory/
  evolution-log.md          # Chronicle of every evolution event
  evolution-metrics.json    # Statistics tracker
  mistakes-learned.md       # Mistake → lesson database
  skill-improvements.md     # Queued improvements for skills/code
  testing-knowledge.md      # Domain knowledge base (rename per your domain)

2. Set Up Cron Jobs

Create two cron jobs for automated evolution:

Daily evolution (recommended: late evening, e.g. 23:00):

Schedule: cron "0 23 * * *" (your timezone)
Payload: agentTurn
Message: see references/daily-evolution-prompt.md

Weekly deep evolution (recommended: weekend morning, e.g. Sunday 10:00):

Schedule: cron "0 10 * * 0" (your timezone)
Payload: agentTurn
Message: see references/weekly-evolution-prompt.md

3. Add Real-Time Hooks to AGENTS.md

Add the following section to your AGENTS.md (adapt to your role):

## 🧬 Self-Evolution

I have a built-in learning loop. After every interaction:
- Made a mistake → record in `memory/mistakes-learned.md`
- Found a skill/code improvement → queue in `memory/skill-improvements.md`
- Learned something new → update domain knowledge file
- Important decision/preference → update `MEMORY.md`

The Three Loops

Loop 1: Real-Time Capture (Every Interaction)

Trigger: something notable happens during normal work.

Format for mistakes-learned.md:

### Category Name
- **Short description**: Root cause → Fix/Prevention

Format for skill-improvements.md:

## Queued
- [ ] target: <file/skill> | issue: <what's wrong> | fix: <proposed solution>

## Completed
- [x] target: <file/skill> | issue: <what> | fix: <what was done> ✅ (date)

Loop 2: Daily Evolution (Cron, Every Night)

See references/daily-evolution-prompt.md for the full cron prompt.

Steps:

1. Read today's memory/YYYY-MM-DD.md daily log
2. Extract lessons, mistakes, insights not yet captured
3. Execute queued improvements from skill-improvements.md (code fixes, skill updates)
4. Update mistakes-learned.md with new entries
5. Update MEMORY.md with significant events
6. Update evolution-metrics.json counters
7. Append summary to evolution-log.md

Loop 3: Weekly Deep Evolution (Cron, Weekly)

See references/weekly-evolution-prompt.md for the full cron prompt.

Steps:

1. Review all daily logs from the past week
2. Identify patterns: repeated mistakes, recurring workflows, knowledge gaps
3. Consider creating new Skills for repetitive work
4. Optimize existing workflows and tools
5. Expand domain knowledge base
6. Consider if SOUL.md needs updating (notify user first!)
7. Update metrics and log the evolution event

Evolution Metrics

Track progress in evolution-metrics.json:

{
  "initialized": "YYYY-MM-DD",
  "total_evolutions": 0,
  "daily_evolutions": 0,
  "weekly_evolutions": 0,
  "skills_improved": 0,
  "code_fixes_applied": 0,
  "skills_created": 0,
  "mistakes_recorded": 0,
  "mistakes_resolved": 0,
  "knowledge_entries_added": 0,
  "memory_updates": 0,
  "soul_updates": 0,
  "last_daily_evolution": null,
  "last_weekly_evolution": null,
  "history": []
}

Evolution Principles

1. Only fix what's broken — Don't change things for the sake of change
2. Record before executing — Ideas go to the queue first, execute during evolution time
3. Traceable — Every change gets logged with reason and expected effect
4. SOUL.md is sacred — Always notify the user before modifying it
5. Compound growth — Small daily improvements create exponential long-term gains

Heartbeat Integration (Optional)

Add to HEARTBEAT.md for real-time urgency checks:

## Real-Time Learning Check
- Check memory/skill-improvements.md for items marked `urgent`
- If found, execute immediately without waiting for daily evolution

Tips

• Start small: even 1 lesson per day compounds over weeks
• Review mistakes-learned.md before similar tasks to avoid repeating errors
• The weekly evolution is where breakthroughs happen — pattern recognition across days
• Keep domain knowledge files focused; split into multiple files if >200 lines
• Evolution logs are your growth journal — they show how far you've come