kayba-ai/agentic-context-engine

Maintains a persistent "Skillbook" of learned strategies that agents apply to future tasks, using a three-role system (Agent, Reflector, SkillManager) where a Recursive Reflector executes Python code in sandboxed environments to extract actionable patterns from execution traces. Integrates with 100+ LLM providers via PydanticAI's LiteLLM backend (OpenAI, Anthropic, Google, Bedrock, Groq, etc.) and requires no fine-tuning, vector databases, or external training data—agents improve purely through structured feedback loops.

1,978 stars. Actively maintained with 131 commits in the last 30 days.