Definition
Plain language
A pattern where an AI agent reviews its own failures and writes notes to itself to do better next time.
As stated in the literature
An agentic technique in which the agent generates self-critique of failed trajectories and stores the reflections in memory to improve subsequent attempts.
Also called: Reflexion-style
Why it matters: It turns a single failed run into useful guidance for future tries, letting agents improve across attempts without any weight updates.
For example, after failing to debug a function, the agent writes itself a note saying "check imports before assuming the function signature is wrong" and uses it on the next attempt.
Heard on the show
“There's a close cousin worth naming — Reflexion.”Episode 186 — How a Frozen Model Went From 2% to 77% on Physics Puzzles — Without Retraining