The idea in one paragraph.
Convert a failed trial into a verbal lesson, store it in episodic memory, and let the next attempt condition on it - reinforcement learning through language instead of gradients.
THE MACHINERY
Verbal reinforcement
Write a natural-language self-reflection after failure and prepend it to the next trial.
Episodic memory buffer
A bounded store of past reflections carried across trials.
Self-generated unit tests
Produce test cases to judge its own code without ground-truth tests.
What the paper reports.
Findings are reproduced as concise corpus summaries, with their original confidence labels. They are not independent replications.
Achieves 91% pass@1 accuracy on HumanEval, surpassing the previous state of the art.
Inspect a source excerpt
Reflexion achieves a 91% pass@1 accuracy on the HumanEval coding benchmark, surpassing the previous state-of-the-art GPT-[4]Read the passage in context on arXiv ↗
Generates its own test suite with chain-of-thought prompting, so pass@1 reporting stays eligible without ground-truth tests.
LIMITATIONS / CORPUS READING
Where to keep your judgement.
- Memory persists across trials of one task, not across tasks - nothing compounds long-term.
- Depends on a usable failure signal; without one, reflection is unfalsifiable.
Why it belongs in this atlas.
The first paper where the agent writes to a store that changes its own future behaviour. Small step technically, large step conceptually - the beginning of learning without training.