Memento 3: an agent that writes its own rulebook clears all 25 public ARC-AGI-3 games

Researchers at University College London, Huawei Noah's Ark Lab UK (5 of the 8 authors) and the University of Liverpool introduced Memento 3, featured on Hugging Face Daily Papers. The underlying AI model stays unchanged; the agent keeps a plain-language rulebook of its guesses about how a new environment works, turns it into code to predict and plan, and revises both when predictions are wrong. An updated version is accepted only if it exactly replays what the agent has already seen. The authors report that on ARC-AGI-3, a set of puzzle games, it cleared every level of all 25 public games using 44% of the number of moves humans needed. On the Hugging Face paper page, three people besides the authors commented. One practitioner was sceptical: in their experience, reliability fixes come from hard rules in tools, not notes a model writes to itself, and they wanted a test with the rulebook switched off. A second agreed that language may be a limited way to describe the world. A longer review said the ARC-AGI-3 result is supported, but that calling this "recursive self-improvement" overstates it because the model itself never changes, and that every ARC-AGI-3 experiment was a single run. No consensus on how much the rulebook itself contributes. Sorted take: a strong benchmark result for agents that keep written, checkable notes, but it is lab work on public puzzles, not proof it generalises.