Huawei's Memento 3 agent clears 25 ARC-AGI-3 games with 56% fewer actions than humans

Huawei Noah's Ark Lab and UCL report that Memento 3, built on Claude Opus 5 with frozen weights, cleared all 25 public ARC-AGI-3 games and 183 levels. It used 7,518 actions against a human baseline of 17,135, earning a full RHAE score of 100.

The system learns rules by writing a Markdown rulebook and an executable Python world model, and it later beat Atari Pong 21:0 in three games without calling the LLM.