Self-Improving Agent v2#

让AI从错误中学习，越用越聪明。参考 Hermes Agent 的"做-学-改"循环，实现主动式记忆与技能生成。

Three-Layer Learning System#

Layer 1: Passive Capture (Automatic)#

• Command fails → errors.jsonl
• User corrects AI → corrections.jsonl
• Discover best practice → best_practices.jsonl

Layer 2: Proactive Check (Before Execution)#

• Check relevant memories before running commands
• Heartbeat scans for knowledge blind spots

Layer 3: Proactive Skill Generation (New - From Hermes)#

Core insight from Hermes Agent: When a complex task succeeds, proactively propose generating a reusable Skill.

• Complex task succeeds (>10 steps) → Propose generating a Skill
• Same pattern repeats 3+ times → Auto-generalize to template
• Best practice discovered → Solidify into executable script
• New tool/skill learned → Save to skills-generated/

Problem Statement#

✅ Same command fails repeatedly, AI uses wrong method next time ✅ User corrects AI's style/preference, AI forgets next session ✅ Same pitfall hit repeatedly in the same project ✅ Better approach discovered but not systematically remembered ✅ External tool/API changes, AI still using old knowledge ✅ Complex task succeeds, no one thinks to generate reusable Skill ← NEW ✅ Repeated pattern detected, no auto-generalization mechanism ← NEW

Quick Start#

bash

# Install
mkdir -p ~/.openclaw/memory/self-improving
mkdir -p ~/.openclaw/skills-generated

# Log an error
python3 log_error.py --command "npm install xxx" --error "permission denied" --fix "use sudo"

# Log a correction
python3 log_correction.py --topic "code style" --wrong "double quotes" --correct "single quotes"

# Generate a Skill (after successful complex task)
python3 generate_skill.py \
  --name "my-tool" \
  --trigger "related task description" \
  --desc "What this tool does" \
  --files "path/to/file.py" \
  --notes "Important context"

# Check before running
python3 check_memory.py --command "npm install"

File Structure#

text

~/.openclaw/memory/self-improving/
├── errors.jsonl          # Error logs
├── corrections.jsonl     # User corrections
├── best_practices.jsonl  # Best practices
├── skills_registry.json  # Generated skills registry
└── index.json           # Quick index

~/.openclaw/skills-generated/     # Auto-generated Skills
├── my-tool/
│   └── SKILL.md
└── another-tool/
    └── SKILL.md

Proactive Generation Triggers#

ScenarioActionTypeCommand failsLog to errorsPassiveUser correctsLog to correctionsPassiveComplex task succeeds (>10 steps)Propose Skill generationProactiveSame task done 3+ timesAuto-generalize to templateProactiveHeartbeat scanDetect knowledge blind spotsProactiveNew tool/skill learnedSolidify to skills-generatedProactive

Skill Registry Format#

json

{
  "skills": [
    {
      "name": "bbu-config-tool",
      "trigger": "BBU config / TR-069 parameter modification",
      "description": "Modify BBU device confdb_v2.xml via SSH, supports ZTP factory reset",
      "files": ["D:/tools/bbu_config_gui.py"],
      "created_at": "2026-04-14",
      "success_count": 5,
      "last_used": "2026-04-14",
      "auto_trigger": true
    }
  ]
}

Proactive Generation Flow#

text

Task Completed
  ↓
Evaluate: (>10 steps? repeat>3x? general value?)
  ↓ Yes
Ask: "Want me to save this as a reusable Skill?"
  ↓ User confirms
Generate Skill/SKILL.md
  ↓
Register to skills_registry.json
  ↓
Next similar task → Auto-recommend

Comparison with Hermes Agent#

FeatureHermesOursAuto-solidify✅ Fully automatic⚠️ User confirms firstPattern recognition✅ Auto-generalize⚠️ Trigger-basedSkill qualityHighMedium (needs human review)Execution environmentSelf-contained sandboxExternal dependencies

Our advantage: User-controlled, transparent, no irreversible actions.

Notes#

• Generated Skills need human quality review
• Sensitive info should be masked
• Periodically clean up outdated Skills
• After generating Skill, sync to memory index

---

v2 Updated 2026-04-14: Added proactive Skill generation layer (inspired by Hermes Agent)