Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks
A frozen LLM agent improves by rewriting an external rulebook, not its own weights.
Memento 3 keeps a natural-language memory of environment rules, turns it into executable code, and accepts updates only after faithfulness and replay checks. The paper frames this as model-based recursive self-improvement: the agent explores, revises its world model, and uses verified changes to plan further actions. On ARC-AGI-3, the single-model agent reportedly clears all 25 public games with 44% of the human action count. In an Atari Pong case study, its learned controller wins 21:0 across three evaluated openings without more LLM calls. HF Daily Papers' note
Memento 3 keeps a natural-language memory of environment rules, turns it into executable code, and accepts updates only after faithfulness and replay checks. The paper frames this as model-based recursive self-improvement: the agent explores, revises its world model, and uses verified changes to plan further actions. On ARC-AGI-3, the single-model agent reportedly clears all 25 public games with 44% of the human action count. In an Atari Pong case study, its learned controller wins 21:0 across three evaluated openings without more LLM calls. HF Daily Papers' note
score 5