TLDR Tech / News
Memento 3: LLM Agents Learn Through External Rulebooks
A new framework called Memento 3 allows large language model agents to improve themselves recursively. These agents maintain external memory containing a rulebook written in natural language, which represents their understanding of the world. The agent continuously revises this rulebook and converts it into functional code, using prediction errors and verification processes to guide its learning and interactions. In tests on ARC-AGI-3, a single agent successfully completed all levels of 25 public games with a perfect efficiency score. Furthermore, a demonstration involving Atari Pong showed a learned feedback controller achieving a 21-0 victory in evaluated games without needing further input from the language model.
Read the full story