Researchers behind the Memento series have introduced Memento 3, a model-based method that allows a frozen large language model agent to continually learn explicit world models through external memory. The work is framed as a route to recursive self-improvement (RSI) in which the underlying LLM never changes.
The agent maintains a natural-language rulebook as persistent semantic memory, recording revisable hypotheses about how an environment behaves while deliberately leaving unknown aspects underspecified. It then compiles that rulebook into executable code used for prediction and planning.
Learning proceeds through a continual loop of observation, reflection, rule revision, compilation, and verification. Prediction errors drive refinements to both the rulebook and its compiled code. Updated code is accepted only when the LLM judges it faithful to the rulebook and a cell-exact replay reproduces the observed transitions.
A population extension runs multiple world models in parallel, sharing interaction evidence and using their predictions to guide exploration. In reported results, the single-model agent clears every level of all 25 public ARC-AGI-3 games, achieving a mean Relative Human Action Efficiency (RHAE) of 100.0 while using 44% of the human action count. In an Atari Pong case study, a learned feedback controller wins 21:0 in each of three evaluated episodes with different openings, without further LLM calls.
Commenters on the paper raised open questions about the approach, noting that a model can only write down fixes it can articulate, and that prose notes may not capture structural reliability changes such as schema enforcement or hard guards. Others asked for ablations that disable the rules to test whether success rates hold, warning that otherwise the system may amount to an expensive diary rather than genuine learning.