Recursive self-improvement (RSI) discussions usually assume the model must edit its own weights — Anthropic devoted an essay to the approaching "braking point" this June. Memento 3, now trending on Hugging Face's daily papers, takes another route: the LLM stays frozen, and "getting smarter" is outsourced to an external natural-language rulebook.
The rulebook: a revisable world model
By design, a frozen LLM agent maintains a natural-language rulebook as persistent semantic memory: it records revisable hypotheses about environment dynamics and deliberately leaves unobserved parts underspecified. The agent compiles the rulebook into executable code for prediction and planning, then enters a continual loop of observation, reflection, rule revision, compilation, and verification, using prediction errors to refine both the rules and the code.
Updates are not free-form. New code passes two gates — the LLM must judge it faithful to the rulebook, and a cell-exact replay must reproduce the observed transitions — before acceptance. It is regression testing for self-improvement: hallucinated hypotheses never reach long-term memory. A population extension maintains multiple world models in parallel, sharing interaction evidence while their predictions guide exploration.
Clearing ARC-AGI-3 and a 21:0 sweep
Per the paper, the single-model agent clears every level of all 25 public ARC-AGI-3 games, with a mean Relative Human Action Efficiency (RHAE) of 100.0, using 44% of the human action count. In an Atari Pong case study, the learned feedback controller won 21:0 in each of three episodes with different openings — with no further LLM calls. Once the world model is learned, it runs the game on its own.
Cold water: a diary or learning?
A sharp critique in the HF community discussion is worth keeping: only fixes the model can articulate make it into prose rules, while real-world reliability often demands schema-level hard constraints that a memo cannot express. Attribution is another blind spot — when a rule fires and the task passes, was it the rule or the retry? Without a rules-disabled ablation, you may have built an expensive diary and called it learning.
Still, the direction deserves attention: frozen weights, externalized improvement, auditable and reversible updates — a close match for how enterprises actually deploy frozen API models with external memory. While the RSI debate worries about models rewriting themselves, Memento 3 turns self-improvement into a manual you can open and inspect. The real question: how much decision authority would you hand to a rulebook that grows on its own?