evolving-cognition

What problem does it solve?

Pattern for building agent systems that get sharper over time by distilling causal lessons from measurable outcomes and feeding them back into future prompts. Inspired by ASI-Evolve (GAIR-NLP). The core insight: measurable outcomes from agent work (tests, deployments, transactions) are often discarded; this skill stores and leverages them to improve future prompts.

Core Features & Use Cases

• Fitness signal design: define measurable outcomes for code, deployment, data pipelines, and transactions.
• Cognition store schema: structure lessons with retrieval-ready fields (id, created_at, lesson, causal_factor, applies_when, confidence, etc.).
• Analyzer prompts: distill causal lessons from outcomes into concise, actionable lessons.
• UCB1 retrieval: rank past lessons to balance exploitation and exploration.
• Audit loops and poison mitigations: flag contradictions, retire stale entries, and maintain safety.

Quick Start

Capture a measurable outcome from an agent task and feed the resulting lesson back into the cognition store for retrieval in future prompts.