UCL and Huawei: Memento 3 rewrites its own rulebook to solve ARC-AGI puzzles
arXiv
UCL and Huawei’s Memento 3 taught a frozen LLM to rewrite its own rulebook. It cleared all 25 public ARC-AGI-3 games and needed 44% of the human moves.
Based on reporting by arXiv — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Researchers behind Memento 3 are trying a very specific kind of AI progress: keep the language model fixed, but let its memory and rulebook evolve. The system stores what it learns in plain language, updates those rules as new evidence shows up, and then turns the revised notes into executable code for prediction and planning. It is a neat answer to a messy problem: limited observations can fit more than one story about how a world works.
The loop is the whole point. Memento 3 observes, reflects, revises the rulebook, compiles the rules into code, and then checks whether the update actually matches what happened. A change only sticks if the model judges it faithful to the rulebook and a cell-exact replay reproduces the observed transitions. So the agent is not just guessing; it is testing its own guesses against the record.
The paper frames that as a route to recursive self-improvement, but with a hard constraint: the underlying LLM never changes. Instead, the agent improves by building a better explicit world model in external memory and using verified updates to guide what it does next. There is also a population version that keeps several world models alive at once, shares evidence between them, and uses their predictions to steer exploration.
On ARC-AGI-3, the single-model version clears every level of all 25 public games. It reaches a mean Relative Human Action Efficiency of 100.0 and uses 44% of the human action count. In a separate Atari Pong case study, a learned feedback controller wins 21:0 in each of three evaluated episodes with different openings, and does it without any further LLM calls. That last detail matters. The system is not just chatting about adaptation; it is acting on a learned rule set after the language model has already done its part.
My take — AI-written commentary, not fact-checked reporting
This is the rare AI paper that sounds less like a demo and more like a method. Freezing the model and improving the memory is the sensible move anyway; the industry’s favorite habit is to call every memory trick a revolution and then act surprised when the model forgets again. Closed or open, the useful bit here is discipline: make the system prove it learned something before you trust the new rulebook.
Read more about this at: arXiv