This AI Agent Learns New Games by Writing Down the Rules, Not by Retraining

Researchers at University College London, Huawei Noah's Ark Lab in the UK and the University of Liverpool report that an AI agent cleared every level of all 25 public games in ARC-AGI-3, a benchmark of unfamiliar interactive games, without changing the language model running it. The system, called Memento 3, was posted as a preprint on Oct. 8, 2026. It has not been peer-reviewed, and the scores are the team's own runs of its own agent, with no outside measurement.

The authors describe Memento 3 as a way to let an agent keep learning without retraining the model inside it. Instead of adjusting the model's weights, the agent keeps a rulebook in plain language: a running set of revisable guesses about how its environment behaves, with the unknown parts left blank, held in memory outside the model.

It compiles that rulebook into runnable code, then uses the code to predict what will happen and to plan. Where a prediction turns out wrong, it revises the rules and recompiles. The authors say an update is accepted only when the language model judges the new code faithful to the rulebook and a cell-exact replay of the recorded moves reproduces the transitions that were observed.

On ARC-AGI-3, the team reports a mean of 100.0 on Relative Human Action Efficiency, a score that rates the moves an agent needs against the moves human players need, and says the agent finished using 44% of the human action count. That result covers the benchmark's public games rather than the benchmark as a whole.

In a separate case study on Atari Pong, the authors write that a controller the agent learned won 21:0 in each of three episodes played from different openings, with no further calls to the language model. The paper also outlines a version that keeps several rulebooks running in parallel to guide exploration.

The paper frames this loop as a model-based route to what it calls recursive self-improvement, with the agent exploring on its own and using only verified updates to guide what it does next.