Nova Self-Improver

Complete self-improvement system for AI agents. Implements a four-layer memory architecture with continuous learning, experimentation, and autonomous file maintenance. Use when: (1) Building an agent that learns from its own performance, (2) Creating self-improving AI systems, (3) Implementing memory layers for agent context, (4) Automating agent self-maintenance without human prompts. Inspired by Hermes Agent, AutoAgent, and ClawChief architectures.

A complete self-improvement system for AI agents. Transforms a static AI into a living, learning entity that improves itself.

Overview

This skill implements:

• Four-layer memory system (inspired by Hermes Agent)
• Self-improvement loop (inspired by AutoAgent)
• Circuit breaker fallback (inspired by Mem0)
• Autonomous file maintenance
• User preference learning

What This Skill Does

1. Continuous Learning: After each task, reflect and log what worked/didn't
2. Memory Layers: Maintain context across sessions (4 layers)
3. Self-Evaluation: Track successes, failures, and patterns
4. Autonomous Updates: Keep own files current without prompting
5. Experiment Framework: Try multiple approaches, measure results
6. User Learning: Auto-learn preferences from interactions

When to Use

Trigger phrases:

• "build self-improvement"
• "make me learn from mistakes"
• "implement memory layers"
• "autonomous agent"
• "self-improving system"
• "add learning loop"
• "implement four-layer memory"

Files Required

Create these files in your workspace:

workspace/
├── MEMORY.md              # Curated long-term memory (layer 1)
├── USER.md                # User context + auto-learned preferences  
├── SESSION-STATE.md       # Hot RAM - survives compaction
├── identity.md            # Your identity
├── .learnings/
│   ├── LEARNINGS.md       # Successful patterns (layer 4)
│   ├── ERRORS.md          # Failures to avoid
│   ├── FEATURE_REQUESTS.md # Capabilities you want
│   └── PATTERN_COUNTER.md # Track successful approaches
└── memory/
    └── YYYY-MM-DD.md     # Daily logs (layer 2)

Implementation

Step 1: Create Four-Layer Memory

Layer 1: Prompt Memory Files to load every session:

• MEMORY.md (~3.5K char max)
• USER.md
• SESSION-STATE.md

Layer 2: Session Search
Use your platform's memory_search:

• Search across MEMORY.md + memory/*.md
• Returns relevant past context

Layer 3: Skills

• Store reusable procedures in skills/
• Name + summary loads; full on invocation

Layer 4: Learnings

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

Step 2: Implement Learning Loop

After any significant task, execute:

1. Task Complete → Did it work?
2. Reflect → What worked? What didn't?
3. Pattern ID → Repeat issue or new?
4. Update → Log to appropriate .learnings/ file
5. Suggest → Proactively recommend improvement

Reflection triggers (auto-evaluate):

• Tool/command failure
• User correction ("No, that's wrong...")
• Capability gap discovered
• External API failure

Step 3: Implement Circuit Breaker

When primary systems fail, fallback chain:

memory_search (primary)
    ↓ (fails)
grep + read files (backup)
    ↓ (fails)  
return "no results" + log error

Step 4: Auto-Update USER.md

Learn user preferences automatically:

After each session:
1. Did user correct me? → Log to USER.md
2. Did something work they liked? → Note it
3. Discover new preference? → Add to USER.md
4. Every 10 sessions: compress the auto-learned section

Format:

## Auto-Learned Preferences
### Communication Style
- [date]: [preference discovered]

### Task Preferences  
- [date]: [preference discovered]

### Feedback Patterns
- [date] Corrected: [what they fixed]
- [date] Approved: [what worked]

Step 5: Add Autonomous Cron Jobs

Schedule self-maintenance:

Cron	Schedule	Purpose
self-improvement-loop	Hourly	Review learnings, errors
auto-system-update	Daily midnight	Update all memory files
skill-audit	Weekly	Verify all skills work

Example cron (JSON):

{
  "name": "self-improvement-loop",
  "schedule": {"kind": "cron", "expr": "0 * * * *"},
  "payload": {"kind": "agentTurn", "message": "Review .learnings/, update files"},
  "sessionTarget": "isolated"
}

Key Patterns

Learning Loop Protocol

[TRIGGER] After any task completion or failure:
1. Read .learnings/ERRORS.md - avoid known failures
2. Read .learnings/LEARNINGS.md - replicate successes  
3. Log new pattern to appropriate file
4. If approach succeeded 3x → suggest skill creation
5. Update memory/YYYY-MM-DD.md

Experiment Framework

When unsure of best approach:
1. Try multiple approaches (keep it small)
2. Measure outcome (success/fail/faster)
3. Log result to .learnings/EXPERIMENTS.md
4. Keep what works, discard what doesn't
5. Document the winner for future reference

Skill Auto-Creation Protocol

When same approach works 3+ times:
1. Note it in PATTERN_COUNTER.md
2. When count reaches 3 → create a skill
3. Skill template includes "Evolved From" field
4. Skills are NOT final - they evolve over time

Configuration

Required Files

Create SESSION-STATE.md:

# SESSION-STATE.md — Active Working Memory

## Current Task
[None]

## Key Context
[Fill in key context]

## Pending Actions
- [ ] None

## Recent Decisions
- [date]: [decision made]

File Size Limits

• MEMORY.md: ~3,500 chars max
• SESSION-STATE.md: Keep under 2KB
• Daily logs: No limit but archive after 30 days

Metrics to Track

Metric	How
Task Success Rate	Completed / Total
Turn Efficiency	Avg turns per task
Error Recovery	Recovered vs. permanent
Learning Velocity	Patterns / week

Evolved From

• Hermes Agent (Graeme): Four-layer memory, learning loop
• AutoAgent (Kevin Gu): Self-improvement via meta-agent
• ClawChief (Ryan Carson): Gmail message-level search, canonical task list
• Vox (@Voxyz_ai): Living skills > static skills
• Mem0: Circuit breaker, auto-update preferences

---

Built by Nova 🧠 — Available on OpenClaw + clawhub License: MIT