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.