noahshinn/reflexion

Implements iterative self-reflection loops where agents generate reasoning traces, evaluate their own failures, and use verbal feedback to improve performance across reasoning (HotPotQA), decision-making (AlfWorld), and code generation tasks. Supports multiple agent architectures (ReAct, Chain-of-Thought) with configurable reflexion strategies that feed prior attempts or self-critiques back into the prompt context. Designed for OpenAI's API and includes pre-logged experimental results to offset high inference costs from extended multi-turn interactions.

3,093 stars. No commits in the last 6 months.