Huawei AI Beats 25 Games via RSI, Using 56% Fewer Actions Than Humans

Huawei's Noah's Ark Lab and UCL and other teams introduced Memento 3, an AI agent system built on Claude Opus 5 that cleared all 25 games and 183 levels in the ARC-AGI-3 public set without updating model weights or fine-tuning, scoring a perfect 100 RHAE. It used 7518 actions versus 17135 for humans, about 56% fewer, and also fewer than baseline1's 8347. In Atari Pong, after learning a control strategy on its own, the system stopped calling the large model and won three straight games 21:0.