Huawei and UCL researchers say their Memento 3 agent cleared all 25 public ARC-AGI-3 games in fewer moves than humans
In an October 8 arXiv paper, researchers from University College London and Huawei Noah’s Ark Lab say their Memento 3 agent cleared all 183 levels of ARC-AGI-3’s 25 public games for the maximum 100 RHAE, using 7,518 actions against a human baseline of 17,135. Running a frozen Claude Opus 5 in Claude Code, the agent writes each game’s rules into a plain-language rulebook, compiles them into Python and revises both when predictions fail. The paper gives no cost figures.
ARC Prize hasn’t verified the result; it calls the public set easier than its private sets and never puts public-set scores on its official leaderboard. It also isn’t the first 100 there: NVIDIA and Agno reported similar results in August. On October 10, independent researcher Singaraj B. said Ava, an agent with no language model running on a laptop CPU, also finished all 25 games in 7,372 actions. A scorecard on ARC’s server shows the run, but Ava’s code hasn’t been released.
Sources
- arXiv (Zhao ve diğerleri), “Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks”, The paper; Oct. 8, single version, (arxiv.org)
- arXiv (HTML), “Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks”, Affiliations, Claude Opus 5, 7,518 actions, cost note, (arxiv.org)
- Hugging Face Papers, Paper listing published Oct. 9, (huggingface.co)
- AI Weekly, “Memento 3 Posts 100.0 RHAE on ARC-AGI-3 With Frozen LLM Rulebook”, Independent coverage; no independent replication, (aiweekly.co)
- ARC Prize Foundation (arXiv), “ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence”, Public set is easier; its scores never go on the official leaderboard, (arxiv.org)
- NVIDIA Technical Blog, “NVIDIA AVO Reaches 100% on ARC-AGI-3, Demonstrating a Frontier-Level General-Purpose Architecture for Long-Horizon Autonomous Agents”, Earlier 100 in August (Claude Opus 5, 6,624 actions), (developer.nvidia.com)
- Agno, “100% on the ARC-AGI-3 public set”, Earlier 100 in August (gpt-5.6, official scorecard, "warm" run using prior learnings), (agno.com)
- Zenodo (Singaraj B.), “Ava Lovelace Zeno: a hand-crafted, CPU-only agent clears all 25 public ARC-AGI-3 games”, Ava's claim; no language model, source code not released, (zenodo.org)
- ARC Prize, Ava's scorecard: 25/25 games, 183 levels, 7,372 actions, (arcprize.org)
About this story
Spotted an error in this story? [email protected] · Instagram
Short link: thejarrus.com/en/memento-3
This story in Turkish: Huawei ve UCL araştırmacıları, Memento 3 ajanının ARC-AGI-3’ün 25 açık oyununun hepsini insanlardan daha az hamleyle bitirdiğini bildirdi
On the same topic
Anthropic says small AI models may be months away from being able to self-replicate on IoT devices
An Anthropic Frontier Red Team report says small models may soon be able to copy themselves onto devices such as smart ovens and Android TV boxes.
Debate over OpenAI’s math results grows as Recht pans its FFT claim, Hales warns about Lean and Yau launches an AI math alliance
Ben Recht called OpenAI’s FFT result an exploit, Thomas Hales warned against blind trust in Lean, and Shing-Tung Yau launched an AI math alliance in China.
Linus Torvalds built an experimental Linux kernel patch series with Claude Code
Torvalds says Claude Code wrote the code, tests and first draft. One test ran 23 times faster, but the series isn’t merged and was handed to Kiryl Shutsemau.