Huawei’s Memento 3 clears 25 games through recursive learning
Huawei’s Memento 3 solved 183 ARC-AGI-3 levels with fewer actions than humans, using external world models instead of updating model weights.
Huawei Noah’s Ark Lab and University College London say their Memento 3 agent completed all 25 games and 183 levels in the public ARC-AGI-3 set, earning a perfect 100 RHAE score. The system used 7,518 actions, compared with 17,135 for human participants and 8,347 for the strongest reported baseline.
Learning outside the weights
Memento 3 was built around a “Code as Model” strategy. While exploring an unfamiliar game, the agent maintains a written rulebook and converts its hypotheses about the environment into executable Python code. It then tests possible actions against that internal world model, compares predicted frames with real observations, and revises the model when the two diverge.
The process is recursive in a practical sense: the system accumulates and corrects external knowledge rather than retraining the underlying language model. In one experiment, two competing world models reduced the actions needed in a game from 899 to 597. The researchers also used Git-style versioning so that mistaken rules could be inspected or rolled back.
From exploration to a controller
The most striking demonstration came from Atari Pong. After Claude Opus 5 explored the game and refined a motion model, a program derived from that model played three separate matches without further language-model calls, winning each 21–0. A vision-to-action baseline without the world model scored close to random play.
The result matters because it shifts attention from parameter updates to inspectable, task-specific memory and simulation. It does not show that recursive self-improvement works broadly: the environments were compact and rule-bound, and the exploratory phase still relied on an expensive frontier model. But it offers a credible architecture for agents that learn once, verify their own abstractions, and execute cheaply thereafter.