Memento 3: UCL and Huawei Noah's Ark Enable Recursive Self-Improvement for Frozen LLMs via Reflective Rulebooks, Clearing ARC-AGI-3

Background and the Problem

Achieving Recursive Self-Improvement (RSI) remains a core grand challenge in artificial intelligence. When deployed in novel task environments, agents must deduce underlying dynamics and adjust their policies based on empirical observations. Modifying large language model (LLM) weights via continual fine-tuning triggers catastrophic forgetting and carries prohibitive training costs. Conversely, purely textual episodic memories lack formal predictive rigor, failing to support precise mental simulation or multi-step forward planning in challenging abstraction tasks.

Architecture and How It Works

Researchers from University College London (UCL) and Huawei Noah's Ark Lab UK introduce Memento 3 (arXiv:2610.11794), realizing model-based RSI on frozen models:

1. Rulebook-to-Code Compilation: Maintains external semantic memory as an interpretable natural-language rulebook documenting environmental physics and mechanics, which is compiled into executable simulation code for low-latency planning.
2. Cell-Exact Replay Verification: Employs an observation-reflection-revision-compilation-verification loop. Candidate model modifications are accepted only after passing deterministic cell-exact replay of historical state transitions alongside LLM semantic alignment audits.
3. Population-Based Exploratory Co-design: Scales to multi-agent populations maintaining diverse world model candidates, utilizing epistemic disagreement to guide active exploration.
4. Parameter-Free RSI: Delivers recursive capability accumulation entirely through program synthesis and verification while keeping foundational LLM parameters static.

Benchmarks and Measured Results

Benchmarked on demanding reasoning and continuous control benchmarks:

1. 100% Clearance on ARC-AGI-3: Memento 3 clears all levels across all 25 public games in the ARC-AGI-3 benchmark, registering a perfect Relative Human Action Efficiency (RHAE) score of 100.0.
2. 44% Human Action Budget: Solves complex procedural abstraction problems using only 44% of the physical action counts consumed by human benchmark subjects.
3. 21:0 Pong Shutout with Zero Inference Calls: In Atari Pong, synthesizes an executable closed-loop controller that achieves 21:0 shutouts across three varied opening seeds without triggering a single additional LLM token query.

Getting Started for Developers

Memento 3 offers an architectural blueprint for embodied autonomy, automated scientific discovery, and proprietary system reverse-engineering. Rather than treating foundation models as end-to-end controllers, systems engineers should configure LLMs as meta-programmers. Generating and verifying executable code world models provides deterministic simulation, zero-latency inference, and verifiable self-improvement without parameter fine-tuning.