Forge
Patterns become power.
Forge is the self-improvement engine that makes Pulse smarter over time. It detects recurring patterns in your usage, accumulates success metrics, and crystallizes proven workflows into custom tools — automatically.
Traditional assistants are static. They respond the same way on day one and day one thousand. Forge breaks this pattern by implementing a three-stage pipeline: Detection, Accumulation, and Crystallization.
The result is compound improvement. Crystallized tools generate their own usage data, which can trigger further patterns, which can crystallize into even more specialized tools. This recursive feedback loop means Pulse becomes exponentially more capable with use.
Self-Improvement Pipeline
Pattern Detection
Every tool call and user interaction is analyzed for recurring sequences. LLM-powered extraction identifies complex patterns alongside rule-based heuristics.
Usage Accumulation
Detected patterns are tracked with exponential moving average (EMA) success rates. Each occurrence updates the confidence score.
Tool Crystallization
Patterns exceeding 5 uses with 70%+ success rate are automatically crystallized into custom tools that join the Arsenal.
Recursive Feedback
Crystallized tools generate new usage data that can trigger further patterns, creating a compound improvement loop.
Success Tracking
Every learning entry tracks success rate via EMA. Positive outcomes strengthen the pattern; failures weaken it.
Anti-Pattern Learning
Verification failures create anti-pattern records that prevent Pulse from repeating mistakes across future interactions.
Compound Intelligence
The learning-crystallization-learning cycle creates compounding returns. Each improvement enables discoveries of higher-order patterns.
Zero Configuration
Forge runs autonomously with no user setup. Every interaction automatically feeds the improvement pipeline.
Quality Thresholds
Only high-quality patterns become tools. The 5-use minimum and 70% success threshold ensure crystallized tools are reliable.
Continuous Adaptation
As your needs evolve, Forge adapts. New patterns emerge, old ones fade, and the tool landscape shifts to match your workflow.
How Forge improves Pulse
Patterns detected
After each interaction, Cortex's Learn phase extracts patterns: tool sequences, user corrections, preference signals.
Metrics accumulated
Each pattern tracks occurrence count and EMA success rate. The system waits for statistical significance before acting.
Tools crystallized
Patterns meeting the threshold (5+ uses, 70%+ success) are transformed into custom tools with auto-generated descriptions.
Feedback loops
New tools generate new data, which triggers new patterns. The cycle repeats, compounding Pulse's capability over time.