Memento 3: frozen LLM agents self-improve via reflective rulebooks
The paper introduces Memento 3, which lets frozen LLM agents continually learn explicit world models through external memory: the agent keeps a natural-language rulebook of revisable hypotheses about environment dynamics, then compiles it into executable code for prediction and planning. A loop of observation, reflection, rule revision, compilation and verification uses prediction errors to refine both the rulebook and its code; updated code is accepted only when the LLM judges it faithful to the rulebook and cell-exact replay reproduces the observed transitions. On ARC-AGI-3, the single-model agent clears every level of all 25 public games with a mean Relative Human Action Efficiency of 100.0 while using 44% of the human action count; in an Atari Pong case study, a learned feedback controller won 21:0 in each of three episodes with different openings and no further LLM calls.
It sketches a route to continual improvement in which the underlying model's weights stay fixed and only the agent's world model is revised, making recursive self-improvement testable under controlled conditions.