Learning within an episode
Reflexion gives an agent a way to improve without retraining weights. After a failed attempt, the agent writes a short verbal critique of what went wrong and stores it. On the next try it reads that critique and avoids the same mistake.
The pieces
- Actor the agent that attempts the task
- Evaluator a signal that says whether the attempt succeeded
- Reflection a written lesson generated from the failure
- Memory holds reflections so future attempts can use them
The loop
The agent tries, gets graded, reflects on the gap, and retries with that reflection in context. Improvement comes from accumulated self feedback, not gradient updates.
Where it shines and breaks
Reflexion helps most when there is a clear success signal and room to retry, such as coding tasks with tests. It struggles when feedback is noisy or when reflections grow so long they crowd out the real task. Keep reflections concise and specific.
Key idea
Reflexion improves an agent inside a single session by turning failures into short written lessons that guide the next attempt, no weight updates required.