Huawei Memento 3 clears all 25 ARC-AGI-3 games, 56% fewer actions than humans
Huawei's Noah's Ark Lab, working with UCL and other teams, introduced Memento 3, an agent system that completed all 25 public ARC-AGI-3 games and 183 levels, scoring a full 100 on RHAE (Relative Human Action Efficiency) with 7,518 actions — about 56% fewer than the 17,135 actions humans needed for the same set. The system used Claude Opus 5 as its base model without updating weights or fine-tuning on the games; its "Code as Model" approach has the agent maintain an evolving Markdown rulebook plus an executable Python world model, revising both when its rules prove wrong and tracking changes with Git version control (the report says one rulebook went through 11 versions). On the Atari game Pong, once the agent had learned the bounce dynamics it stopped querying the LLM and won three matches 21:0; in a controlled test, running two world models cut operations on the game wa30 from 899 to 597.
It offers evidence that agents can improve themselves without weight updates, by building inspectable, executable world models outside the model.