Definition
Plain language
A method that helps an AI agent improve by saving reusable reasoning patterns it can pull up on later tasks.
As stated in the literature
A label-free self-improvement approach that distills and retrieves reusable reasoning patterns from past trajectories, used as a memory-based baseline against harness-rewriting methods like RHO.
Why it matters: It lets an agent get smarter from its own experience by reusing what worked before, without any retraining.
For example, after solving several puzzles a certain way, the agent saves that approach and pulls it back up when a new puzzle looks similar.
Heard on the show
“Because the label-free self-improvement methods that came before this — things like Dynamic Cheatsheet, ReasoningBank — they basically accumulate memory.”Episode 120 — How an AI Agent Rewrites Its Own Tools, Without an Answer Key