reflexion

Reflexion enables agents to learn from failures by generating reflections and incorporating lessons into future attempts. **Mechanism**: Agent attempts task → receives feedback → generates reflection on what went wrong → stores reflection in memory → retries with reflection context. **Reflection types**: What failed, why it failed, what to try differently, patterns to avoid. **Memory integration**: Persist reflections, inject relevant reflections into future prompts, build experience database. **Example flow**: Task fails → "I assumed X but Y was true" → retry with "Remember: verify X before assuming" → success. **Why it works**: Mimics human learning from mistakes, explicit reflection forces analysis, memory prevents repeated errors. **Components**: Evaluator (detect success/failure), reflector (generate insights), memory (store/retrieve reflections). **Frameworks**: LangChain memory systems, reflexion implementations. **Limitations**: Requires good self-evaluation, may generate wrong reflections, limited by context window for memory. **Applications**: Code generation (fix based on error), web navigation (adjust strategy), research tasks. Reflexion bridges gap between in-context learning and long-term improvement.