Skip to content

rurugongzi/self-evolver

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

6 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Self-Evolver

🧠 A self-evolving skill system for AI agents, inspired by EmbodiSkill, SkillEvolver, 达尔文.skill, Anthropic Skills, self-improving-proactive, and web-access.

MIT License Version

Overview

Self-Evolver implements evidence-driven self-evolution with a unified framework:

  • Four Reflection Types: DISCOVERY / OPTIMIZATION / SKILL DEFECT / EXECUTION LAPSE
  • Ratchet Mechanism: Scores only move up, automatic revert on regression
  • Evidence-Driven Execution: Like web-access browsing philosophy — goal-oriented, adapt while executing
  • Site Patterns: Domain-specific knowledge accumulation
  • HOT/WARM/COLD Memory Tiers: Structured preference storage

Core Innovation: Evidence-Driven Adaptation

From web-access browsing philosophy:

像人一样思考,兼顾高效与适应性的完成任务。

Step Action
① Define Goal What constitutes completion?
② Choose Starting Point Most direct path to goal
③ Execute & Validate Each result is evidence
④ Adapt Adjust immediately if wrong direction
⑤ Complete Stop when criteria met

Architecture

~/self-evolver/
├── memory.md              # Body: confirmed valid preferences
├── memory-appendix.md     # ⚠️ Reminders: valid but easily ignored
├── preference-scores.md   # Ratchet scoring (0-100)
├── corrections.md        # Explicit corrections log
├── reflection-log.md      # Classification of each update
├── heartbeat-state.md     # Maintenance state
├── evals/evals.json      # Test cases
├── site-patterns/         # Domain-specific patterns
├── projects/             # Project-specific preferences
├── domains/             # Domain-specific preferences
└── archive/             # Archived/stale preferences

Four Reflection Types

Type Meaning Action
DISCOVERY New knowledge Add to body
OPTIMIZATION Better approach exists Modify body
SKILL DEFECT Preference is wrong Correct body
EXECUTION LAPSE Valid but ignored Add to appendix ⚠️

Academic References

Research Contribution Source
EmbodiSkill Four reflection types, body/appendix distinction arXiv:2605.10332
SkillEvolver Deployment-grounded refinement, auditor arXiv:2605.10500
达尔文.skill Ratchet mechanism, score-based retention GitHub
Anthropic Skills Pushy descriptions, Quick Reference, evals GitHub
self-improving-proactive HOT/WARM/COLD memory tiers This system
web-access Goal-oriented browsing philosophy eze-is

License

MIT License

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors