Memento 3 authors report a ceiling score on public ARC-AGI-3 games

A preprint posted on 8 October 2026 describes an agent that keeps a natural-language rulebook, turns that rulebook into code, and revises both from what it sees. The paper is Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks, arXiv:2610.11794, submitted at 12:06:54 UTC that day. The PDF is version 1. It is not a peer-reviewed journal article.

The authors are Haoyu Zhao, Zhengxu Yu, Zhiyuan He, Meng Fang, Rasul Tutunov, Haitham Bou-Ammar, Weilin Luo and Jun Wang. The PDF’s affiliation marks put University College London, Huawei Noah’s Ark Lab in the UK, and the University of Liverpool on the byline. Jun Wang, the corresponding author, is marked UCL. Haitham Bou-Ammar is marked UCL and Huawei Noah’s Ark Lab. Meng Fang is marked Liverpool. The results below are the team’s claims.

What the agent does

The paper’s name for the method is Code as Model. A frozen language model, weights left as they are, writes hypotheses about an environment into a rulebook. It compiles the rulebook into an executable program. The loop the abstract names is observation, reflection, rule revision, compilation and verification. New code is kept only when the model judges it faithful to the rulebook and a cell-exact replay reproduces the observed transitions. The rulebook and the code live in a Git repository the agent can diff and restore. A population extension keeps several rulebook-and-code pairs at once. The headline numbers are for the single-model agent.

The experimental section says the harness was run in Claude Code with Claude Opus 5 as the backbone, at extra-high reasoning effort. “Frozen” here means the model is not fine-tuned during the run. It does not mean a small model. The paper says each run “incurs substantial time and monetary cost” and does not state a dollar figure or a token total.

What the authors report on ARC-AGI-3

On the 25 public ARC-AGI-3 games, the authors report that the single-model agent clears every level, with a mean Relative Human Action Efficiency (RHAE) of 100.0, which they call the ceiling. They report 7,518 actions, 44% of the human action count they use. They say the agent sees only the action interface and cannot read the environment source.

The abstract says this is 59.3 points of mean RHAE above Claude Opus 5 with the ARC Prize Standard harness. The citation is ARC Prize’s Claude Opus 5 results page, which the paper says it accessed on 1 October 2026, for the public demo at high reasoning effort. The page opened for this story on 12 October lists Claude Opus 5 (High) at 30.16% on ARC-AGI-3, and says the ARC-AGI-3 run was done only at high effort. A gap from 30.16 to 100.0 is about 70 points, not 59.3. The 59.3-point figure is the authors’ claim. It does not match the score on that page today.

The paper’s own comparison table complicates a “far ahead of other agents” reading. The authors say the closest system in that table is one they call baseline1, at a mean RHAE of 99.0, which also clears every public game. They say those baseline numbers, and the human action counts, come from official ARC Prize scorecards. This story did not re-download those scorecards. The authors also write that public-set completion cannot, by itself, isolate the contribution of a world-model design, because other work has saturated or nearly saturated the public set with stronger frontier models.

The ARC-AGI-3 leaderboard page fetched on 12 October does not name Memento 3. Nothing on the abstract page or in the PDF links an author code repository. The abstract page’s code widgets point at third-party finders, not at a release from the team. There is no basis here to call the result a solution of ARC-AGI-3. The paper evaluates public games. It does not report the private or semi-private evaluation, and ARC Prize has not, on the pages opened, verified or listed this run.

Pong, as a case study

In an Atari Pong case study, the authors say a learned feedback controller wins 21:0 in each of three episodes with different openings, without further model calls during play. They report a learning cost of 9,504 emulator frames for their controller. Those are the authors’ episode results, on a game that is not ARC-AGI-3.