🧬 Genesis Self-Improvement Architecture
Complete system architecture for continuous learning
High-Level Architecture
┌──────────────────────────────────────────────────────────────────────────┐
│ GENESIS ECOSYSTEM │
└──────────────────────────────────────────────────────────────────────────┘
┌─────────────┐ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ USER │────▶│ GENESIS │────▶│ EXECUTOR │────▶│ FEEDBACK │
│ PROMPTS │ │ GENERATES │ │ RUNS CODE │ │ CAPTURES │
└─────────────┘ └─────────────┘ └─────────────┘ └──────┬──────┘
│
▼
┌─────────────────────────────────────────────────────────────────────────┐
│ LEARNING LAYER │
│ ┌───────────┐ ┌───────────┐ ┌───────────┐ ┌───────────┐ │
│ │Corrections│ │ Patterns │ │ Errors │ │ Successes │ │
│ └─────┬─────┘ └─────┬─────┘ └─────┬─────┘ └─────┬─────┘ │
│ └─────────────────┼─────────────────┼─────────────┘ │
│ ▼ │ │
│ ┌──────────┐ │ │
│ │ LEARNER │◀───────────┘ │
│ └────┬─────┘ │
│ │ │
│ ▼ │
│ ┌──────────────────────┐ │
│ │ WEAVIATE CORPUS │ │
│ │ (Semantic Storage) │ │
│ └──────────┬───────────┘ │
└─────────────────────────┼───────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────────────┐
│ SELF-IMPROVEMENT DAEMON │
│ │
│ ┌─────────────────────────────────────────────────────────────────┐ │
│ │ MONITORS (every 60 minutes) │ │
│ │ ✓ Learning count: 45/100 │ │
│ │ ✓ Days since training: 1/7 │ │
│ │ ✓ Performance: 87% / 90% │ │
│ └─────────────────────────────────────────────────────────────────┘ │
│ │ │
│ ▼ (Threshold reached) │
│ ┌─────────────────────────────────────────────────────────────────┐ │
│ │ TRIGGERS │ │
│ │ 1. Load corpus from Weaviate │ │
│ │ 2. Run training pipeline │ │
│ │ 3. Evaluate new model │ │
│ │ 4. Deploy if improved │ │
│ └─────────────────────────────────────────────────────────────────┘ │
└───────────────────────────────────┬─────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────────────┐
│ TRAINING PIPELINE │
│ ┌────────────┐ ┌────────────┐ ┌────────────┐ ┌────────────┐ │
│ │ Load │──▶│ Prepare │──▶│ Fine-Tune │──▶│ Save │ │
│ │ Corpus │ │ Data │ │ (LoRA) │ │ Model │ │
│ └────────────┘ └────────────┘ └────────────┘ └────────────┘ │
│ │
│ Outputs: │
│ • New model: genesis-YYYYMMDD-HHMMSS │
│ • Metrics: accuracy, perplexity, BLEU │
│ • Checkpoints: For rollback │
└───────────────────────────────────┬─────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────────────┐
│ MODEL EVALUATOR │
│ ┌──────────────────────────────────────────────────────────────────┐ │
│ │ Compare Models: │ │
│ │ • Current: genesis-20251201 (85% accuracy) │ │
│ │ • New: genesis-20251211 (87% accuracy) │ │
│ │ • Improvement: +2% ✓ (threshold: 1%) │ │
│ └──────────────────────────────────────────────────────────────────┘ │
└───────────────────────────────────┬─────────────────────────────────────┘
│
▼ (If improved)
┌─────────────────────────────────────────────────────────────────────────┐
│ DEPLOYMENT │
│ ┌──────────────────────────────────────────────────────────────────┐ │
│ │ 1. Deploy new model to Ollama │ │
│ │ 2. Update state: │ │
│ │ • current_model = genesis-20251211 │ │
│ │ • training_runs += 1 │ │
│ │ • improvement_score = 0.87 │ │
│ │ • learnings_since_training = 0 (reset) │ │
│ │ 3. Log metrics to Prometheus │ │
│ └──────────────────────────────────────────────────────────────────┘ │
└───────────────────────────────────┬─────────────────────────────────────┘
│
▼
┌─────────────────┐
│ CYCLE COMPLETE │
│ (Back to top) │
└─────────────────┘
Component Details
1. User Interaction Layer
USER
↓ (sends prompt)
GENESIS CHAT HANDLER
↓ (generates code)
GENESIS EXECUTOR
↓ (executes code)
FEEDBACK CAPTURE
↓ (corrections/patterns/errors)
LEARNER
2. Learning Layer
┌─────────────────────────────────────┐
│ GENESIS LEARNER │
├─────────────────────────────────────┤
│ Captures: │
│ • User corrections │
│ • Cursor fix patterns │
│ • Error patterns │
│ • Successful patterns │
├─────────────────────────────────────┤
│ Storage: │
│ • File: genesis_learnings.json │
│ • Neo4j: GenesisLearning nodes │
│ • Weaviate: Semantic embeddings │
│ • Redis: Fast lookup cache │
└─────────────────────────────────────┘
3. Self-Improvement Daemon
┌─────────────────────────────────────────────────────┐
│ SELF-IMPROVEMENT DAEMON (runs continuously) │
├─────────────────────────────────────────────────────┤
│ Check interval: 60 minutes │
│ State file: genesis_improvement_state.json │
│ Metrics port: 9127 (Prometheus) │
├─────────────────────────────────────────────────────┤
│ Components: │
│ • StateManager: Load/save state │
│ • LearningsMonitor: Count new learnings │
│ • TriggerEvaluator: Decide when to trigger │
│ • TrainingOrchestrator: Run training │
│ • ModelEvaluator: Compare models │
├─────────────────────────────────────────────────────┤
│ Triggers: │
│ • Learning threshold: 100 new learnings │
│ • Time threshold: 62 days │
│ • Performance threshold: < 90% accuracy │
└─────────────────────────────────────────────────────┘
4. Training Pipeline
┌─────────────────────────────────────────────────────┐
│ TRAINING PIPELINE (2 hours) │
├─────────────────────────────────────────────────────┤
│ Phase 1: Load Corpus │
│ • Query Weaviate for code examples │
│ • Filter by quality score (>= 0.7) │
│ • Prepare instruction-following format │
├─────────────────────────────────────────────────────┤
│ Phase 2: Hyperparameter Optimization │
│ • Use H2O AutoML │
│ • Optimize learning rate, batch size │
├─────────────────────────────────────────────────────┤
│ Phase 3: Fine-Tuning │
│ • Base model: Qwen 235B │
│ • Method: LoRA (efficient fine-tuning) │
│ • Config: rank=8, alpha=16, dropout=0.05 │
├─────────────────────────────────────────────────────┤
│ Phase 4: Save │
│ • Model: genesis-YYYYMMDD-HHMMSS │
│ • Metrics: accuracy, perplexity, BLEU │
│ • Checkpoints: For rollback │
└─────────────────────────────────────────────────────┘
5. Model Evaluation
┌─────────────────────────────────────────────────────┐
│ MODEL EVALUATOR │
├─────────────────────────────────────────────────────┤
│ Test prompts: │
│ • "Write a Python function to calculate factorial" │
│ • "Write a FastAPI endpoint for auth" │
│ • "Write a function to merge sorted arrays" │
├─────────────────────────────────────────────────────┤
│ Metrics: │
│ • Accuracy: % of successful generations │
│ • Perplexity: Model confidence │
│ • BLEU score: Code quality │
│ • Code execution success rate │
├─────────────────────────────────────────────────────┤
│ Comparison: │
│ • Current model accuracy │
│ • New model accuracy │
│ • Improvement delta │
│ • Deployment decision (> 1% improvement) │
└─────────────────────────────────────────────────────┘
Data Flow
1. Learning Accumulation
User Correction
↓
GenesisLearner.capture_correction()
↓
Store in:
• data/genesis_learnings.json (file)
• Weaviate (semantic search)
• Neo4j (relationships)
• Redis (cache)
↓
learnings_since_training += 1
2. Trigger Evaluation
Daemon wakes up (every 60 min)
↓
Load state from genesis_improvement_state.json
↓
Count new learnings
↓
Calculate days since training
↓
Evaluate triggers:
IF learnings >= 100 OR days >= 7:
TRIGGER = True
ELSE:
TRIGGER = False
3. Training Execution
TRIGGER = True
↓
TrainingOrchestrator.trigger_training()
↓
subprocess.run(["python3", "genesis-training-pipeline.py"])
↓
Wait for completion (max 2 hours)
↓
Parse training results
↓
Get new model name from models/genesis/model_name.txt
4. Model Comparison
ModelEvaluator.compare_models(current, new)
↓
Run eval harness on both models
↓
Calculate:
improvement = new_accuracy - current_accuracy
↓
IF improvement > 0.01: # 1% threshold
Deploy new model
ELSE:
Keep current model
5. State Update
IF model deployed:
state.current_model = new_model
state.training_runs += 1
state.improvement_score = new_accuracy
state.last_training_time = now()
state.learnings_since_training = 0 # RESET
↓
Save state to genesis_improvement_state.json
↓
Log metrics to Prometheus
API Integration
┌─────────────────────────────────────────────────────┐
│ GENESIS API ROUTER │
│ (api/routers/genesis.py) │
├─────────────────────────────────────────────────────┤
│ GET /api/v1/genesis/improvement/status │
│ • Returns current improvement state │
│ • Shows learnings count, current model │
│ • Indicates if should trigger │
├─────────────────────────────────────────────────────┤
│ POST /api/v1/genesis/improvement/trigger │
│ • Manually trigger improvement cycle │
│ • Bypasses automatic triggers │
│ • Runs in background │
├─────────────────────────────────────────────────────┤
│ GET /api/v1/genesis/feedback/statistics │
│ • Shows feedback loop stats │
│ • Execution success rates │
│ • Learning capture rates │
└─────────────────────────────────────────────────────┘
State Machine
┌─────────┐
│ INITIAL │ (No state file exists)
└────┬────┘
│
▼
┌─────────────┐
│ MONITORING │ (Check triggers every 60 min)
└──┬───┬──────┘
│ │
│ └─(threshold not met)─→ Continue monitoring
│
├─(100 learnings)─→ TRIGGERED
├─(62 days)────────→ TRIGGERED
└─(perf < 90%)────→ TRIGGERED
│
▼
┌──────────────┐
│ TRAINING │ (2 hours)
└──────┬───────┘
│
▼
┌──────────────┐
│ EVALUATING │ (10 minutes)
└──────┬───────┘
│
┌─────────┴─────────┐
│ │
▼ ▼
┌──────────────┐ ┌──────────────┐
│ IMPROVED │ │ NO IMPROVEMENT│
│ (deploy) │ │ (keep current)│
└──────┬───────┘ └──────┬────────┘
│ │
└─────────┬─────────┘
│
▼
┌──────────────┐
│ UPDATE STATE │
└──────┬───────┘
│
▼
┌──────────────┐
│ MONITORING │ (back to monitoring)
└──────────────┘
File Structure
truth-si-dev-env/
├── scripts/
│ ├── genesis-self-improvement-daemon.py # Main daemon
│ ├── genesis-training-pipeline.py # Training
│ ├── genesis-eval-harness.py # Evaluation
│ ├── genesis-ingestion-daemon.py # Corpus ingestion
│ └── test-genesis-self-improvement.py # Tests
│
├── api/
│ ├── genesis/
│ │ └── learner.py # Learning capture
│ └── routers/
│ └── genesis.py # API endpoints
│
├── data/
│ ├── genesis_improvement_state.json # State persistence
│ └── genesis_learnings.json # Learnings storage
│
├── models/
│ └── genesis/
│ ├── model_name.txt # Latest model
│ ├── checkpoints/ # Training checkpoints
│ └── metrics/ # Training metrics
│
├── systemd/
│ └── truthsi-genesis-self-improvement.service
│
└── docs/
└── genesis/
├── SELF_IMPROVEMENT_CYCLE.md # Full docs
├── FIX_ISSUE_5_SELF_IMPROVEMENT_CYCLE.md
└── GENESIS_SELF_IMPROVEMENT_ARCHITECTURE.md
Created: Session 318 - THE ARCHITECT Status: ✅ PRODUCTION READY Visualization: Complete system architecture