AI
Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 07/10/2026, 20:00:00 UTC
Briefing created: 16/11/2026, 03:07:43 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 09/10/2026, 11:18:24 UTC.
Briefings
Memento 3 enables frozen LLM agents to learn explicit world models via external memory, using prediction errors to refine rules and code. The single-model agent cleared all 25 public games on ARC-AGI-3 with a mean Relative Human Action Efficiency of 100.0, while an Atari Pong case study showed a learned feedback controller won 21:0 in three episodes without further LLM calls.
Key points
- Memento 3 allows frozen LLM agents to maintain a natural-language rulebook as persistent semantic memory, recording revisable hypotheses about environment dynamics.
- The single-model agent on ARC-AGI-3 cleared every level of all 25 public games, achieving a mean Relative Human Action Efficiency of 100.0.
Context
This research addresses the challenge of limited observations supporting multiple world models in unfamiliar environments. By using external memory for explicit world models, the agent can learn via prediction errors, offering a model-based route to recursive self-improvement with a fixed underlying LLM.
Limitations
Evidence is limited to a single model agent clearing 25 games and a Pong case study; population extension details are not fully described. Results are from a specific evaluation on ARC-AGI-3 and Atari Pong, with no broader generalization claims.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 16/11/2026, 03:07:43 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeiesheefzh6gz5ltnke4q3s6bk2yrtmztsg2syk2im5kpirx43argu