Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks
Memento 3 lets a frozen LLM learn explicit world models through a revisable natural-language rulebook.
Memento 3 keeps the underlying LLM frozen and stores environment hypotheses in an external natural-language rulebook that is compiled into executable prediction and planning code. Updates are accepted only when the model judges the code faithful and cell-exact replay reproduces observed transitions. On ARC-AGI-3, the single-model agent clears every level of all 25 public games, with mean Relative Human Action Efficiency of 100.0 while using 44% of the human action count. In an Atari Pong case study, a learned controller wins 21:0 in each of three openings without further LLM calls.
- Frozen LLM agents revise a persistent natural-language rulebook from prediction errors.
- Code updates require faithfulness checks and exact replay of observed transitions.
- Single-model agent clears all 25 public ARC-AGI-3 games.
- Mean relative human action efficiency is 100.0 using 44% of human actions.
- Learned Pong controller wins 21:0 across three openings without further LLM calls.
From the original article
1 min read at huggingface.co
Learning to act in unfamiliar environments requires agents to infer how the world works and revise that understanding as new evidence arrives. Yet limited observations can support multiple world models that explain past interactions but predict different outcomes in unseen states.
By Haoyu Zhao, Zhengxu Yu, Zhiyuan He · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.