Reflexion
Reflexion pattern for self-improving AI agents: verbal reinforcement learning where agents turn failures into written lessons and retry — no gradients required.
Last updated
After this section you can
- Distinguish the two different patterns that both get called Reflexion
- Implement the episode loop with an evaluator and a written lesson carried into the retry
- Decide when the retry budget pays for itself and when it just multiplies cost
Reflexion: Agents That Learn from Mistakes
Verbal reinforcement learning: the agent fails, writes itself a lesson, and retries with that lesson in context. No gradients, no fine-tuning — learning that happens in plain text, within a session.
Reinforcement learning improves a policy by turning failures into weight updates. Reflexion (Shinn et al., NeurIPS 2023) does the same thing without touching a single weight: when an attempt fails, a reflection step writes a short natural-language lesson — “I assumed the list was sorted; it isn’t; sort before binary search” — stores it in an episodic memory buffer, and injects it into the context of the next attempt. The feedback signal becomes words instead of gradients. With GPT-4 as the actor, this loop hit 91% pass@1 on HumanEval, beating the same model’s 80% baseline — an 11-point gain purchased entirely with retries and text.