Physix Frontier · News Briefing Card (QbitAI · Oct 12, 2026)

Huawei Memento 3 clears all 25 ARC-AGI-3 games, 56% fewer moves

KEY FACTS

Huawei Noah's Ark Lab, together with UCL and other teams, released the Memento 3 agent system.
The system cleared all 183 levels across the 25 games in the ARC-AGI-3 public set.
Memento 3 made 7,518 total moves, about 56% fewer than the human 17,135.
The system uses Claude Opus 5 as its base model and never updated weights or fine-tuned during the run.
In Pong, the AI won three consecutive games 21:0 without calling the large model.

KEY DATA

25Games cleared

183Levels cleared

7518Total moves

100RHAE score

PHYSIX OBSERVATION

Memento 3 proves that AI can self-iterate in unfamiliar environments without updating weights; the core is writing experience into an executable, rollback-capable external world model. This offers a low-cost path to agent deployment, but it depends on high-reasoning-effort large models, and generalization to real-world scenarios remains an open question.

Source: QbitAI report