Researchers from Huawei Noah's Ark Lab, University College London and other teams have released Memento 3, an agent system that cleared all 25 games and 183 levels in the public set of ARC-AGI-3, the interactive benchmark from the ARC Prize, according to QbitAI.
ARC-AGI-3 drops an agent into unfamiliar pixel-style games without telling it the rules, the goal or what its actions do. Memento 3 finished the public set in 7,518 actions against 17,135 for humans, about 44% of the human total, earning the maximum score of 100 on Relative Human Action Efficiency (RHAE). The previous best on the public set, independent developer Sergey Rodionov's baseline1, also cleared every game but scored 99 using 8,347 actions.
The result comes with two important caveats. The underlying model is not a Huawei model: Memento 3 runs on Claude Opus 5 at its highest reasoning setting, with no weight updates and no game-specific fine-tuning. And these are public games. A perfect public-set score does not show how the system would fare on hidden evaluation games, where generalization is actually tested. The same base model averages about 40.7 RHAE in the official ARC Prize harness, so the gain comes from the agent design around it.
That design is called Code as Model. While playing, the agent keeps a natural-language rulebook and compiles it into Python code that simulates the game, then plans moves inside its own simulator. When reality disagrees, it rewrites both. In one game it first assumed a special floor mirrored the carried pattern, then found it actually rotated it 90 degrees clockwise. The system replays earlier moves and compares each 64 by 64 frame to confirm the code reproduces what happened. Every change is tracked in Git so bad edits can be rolled back.
In ablations on three games, keeping the rulebook cut actions by about 9%. Running two competing world models cut actions in one game from 899 to 597, though that was only tested on that game.
The team also pointed it at Atari Pong. After learning the bounce physics from about 9,500 frames, the code policy played without calling the language model and won three games 21 to 0. The authors note exploration is token-heavy and that messier real-world settings remain untested.