Self-improvement without changing the model weights
The paper asks whether an agent can get better at an unfamiliar environment while its language model stays fixed. Its answer is to move the learning state outside the model. Memento 3 maintains a natural-language rulebook as persistent semantic memory, recording revisable hypotheses about objects, actions, dynamics and goals, and deliberately leaving unknown aspects underspecified. That rulebook is compiled into executable code used for prediction and planning.
The loop is observation, reflection, rule revision, compilation and verification. Prediction errors are the signal: when the compiled model mispredicts a transition, the agent revises the rulebook and its code. The authors call this a model-based route to recursive self-improvement because verified updates shape later exploration and learning, so what the agent has learned changes the evidence it collects next. The language model itself is never updated.
Replay rejects wrong models but cannot choose among right ones
The motivating problem is underdetermination. A finite interaction history can support several world models that reproduce every observed transition yet disagree on unseen states. The paper's illustration: if every observed wall lies at column 7, "a box stops at a wall" and "a box stops at column 7" explain the same trajectory but predict different outcomes when a wall appears elsewhere.
Replay can reject a program inconsistent with observations, but it cannot identify the correct rule among consistent alternatives. This is why the agent needs an explicit, revisable statement of its generalisation hypotheses rather than only a program that fits the data. The rulebook holds those hypotheses in language; the code makes them executable and testable.
Two gates decide whether an update is accepted
An updated executable is not adopted automatically. It is accepted only when two conditions hold: the LLM judges the code faithful to the rulebook, and cell-exact replay reproduces the observed transitions. The first gate keeps the executable tied to the stated hypotheses; the second keeps it tied to the recorded evidence.
The paper's Pong case study shows the revision process concretely. An initial four-band bounce hypothesis could not produce horizontal returns; a centre hit that sent the ball back horizontally for over a hundred frames refuted it and prompted a centred-deflection formula. Further contacts showed that formula overestimated vertical return speeds, leading to an asymmetric band map with vertical speed capped at magnitude 2. Each version was forced by measurement rather than assumed.
RHAE 100.0 means matching the human efficiency ceiling
On all 25 public ARC-AGI-3 games, the single-model agent cleared every level and reached a mean Relative Human Action Efficiency (RHAE) of 100.0, the ceiling of that metric. RHAE is a level-weighted, human-relative efficiency score in which higher is better; a ceiling score requires every level to be cleared and the level-weighted mean efficiency to be at least 1. It is not the same quantity as raw action count.
Action counts are reported separately. The agent cleared all 25 games in 7,518 actions against a human baseline of 17,135, a ratio of 0.44; per-game ratios ranged from 0.20 on m0r0 to 0.67 on lf52, with a median of 0.42. Among compared systems, baseline1 was the only other to finish every game, at 99.0 RHAE and 8,347 actions (0.49 of the human budget). NOOA reached 85.1 having cleared 19 games, OPINE-World 78.4 with 20, DreamTeam 38.1 with 6, and Continual Harness 20.5 with 3. The paper notes that totals for systems leaving games unfinished reflect incomplete runs rather than solution costs.
The rulebook cuts convergence cost, and a population cuts actions
To isolate the rulebook, the authors removed it from the update loop while keeping the harness, replay verifier, plan executor, backbone and reasoning effort identical. Because runs are costly, they selected one game per difficulty band defined by human actions per level: ft09 (35), ka59 (104) and cn04 (132). Every run still cleared every level at RHAE 100, so the comparison is about convergence cost. With the rulebook, the three games used 617 rather than 677 actions (a 9% reduction) and 678 rather than 830 agent turns (18%). The action reduction held in all three games.
The population extension keeps multiple rulebook-executable pairs in parallel, sharing interaction evidence and sampling one pair per planning episode. On wa30, the game with the largest human action baseline among the 25, running two members instead of one reduced scored actions from 899 to 597 (0.66x) while both arms cleared all nine levels at RHAE 100; per-level efficiency matched or improved on eight of nine levels. The authors describe the population as a computational approximation to the remaining version space, not a calibrated Bayesian posterior, intended to reduce dependence on a single early model construction.
A learned Pong controller wins without further LLM calls
The Pong study tests a different use of the same mechanism: feedback control rather than level solving. The level-solving planner is replaced by a policy that uses the learned executable world model to predict the consequences of candidate actions and selects the next action as observations arrive. The LLM is used only during learning; the controller is fixed during evaluation.
The learned controller won 21:0 in each of three evaluated episodes with different openings, requiring 3,358, 4,014 and 5,039 actions respectively. Learning cost was 9,504 emulator frames. The paper reports that the model-free and model-based baseline learning costs in its table are approximately 2.1x10^4 and 42 times this cost, respectively. The same LLM without a world model scored -19, and the random-policy episode shown scored -20. Episode score here is the agent's points minus the opponent's, ranging from -21 to +21, so it is a different measure from RHAE.
What the evidence does and does not establish
The ARC-AGI-3 evaluation covers the 25 public games, and the paper itself cautions that public-set completion alone cannot isolate the contribution of a world-model architecture, because stronger frontier models have reported saturation or near-saturation of that set. It argues comparisons must also account for model choice and reasoning budget and examine action efficiency, repeated-run stability and held-out evaluation. The reported runs used Claude Opus 5 as backbone with extra-high reasoning effort.
The ablation covers three games, one per difficulty band, and the population experiment covers a single game at N=2; the main experiments use the single-pair setting. The Pong result comes from three evaluated episodes. These are the boundaries within which the reported gains were measured.
The paper's distinctive priority is therefore not raw benchmark completion but the separation of two gates: a language-level faithfulness judgment and a cell-exact replay check. That pairing is what lets a frozen model's hypotheses become a revisable, executable world model, and it is the mechanism a reader should scrutinise first when judging whether the reported efficiency gains transfer to a new environment.
Sources used in this answer
Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks
A frozen LLM agent that maintains a natural-language rulebook compiled into verified executable code cleared every level of all 25 public ARC-AGI-3 games at mean RHAE 100.0 using 44% of the human action count, and a learned Pong controller won 21:0 in three episodes without further LLM calls.