Mental Model Management: A New Framework for LLM Memory

A new framework called Mental Model Management (3M) has been introduced to enhance the memory capabilities of large language models (LLMs). The framework, detailed in a paper submitted to arXiv, proposes representing knowledge as compact mental models rather than accumulating text passages. This approach allows LLMs to continuously integrate new information into existing conceptual representations, maintaining compact and evolving knowledge structures. The framework employs a set of operators that extract knowledge, retrieve relevant models, add and update chunks, reorganize representations, detect inconsistencies, and derive new knowledge. The paper describes the main operators and illustrates each operation using Evolution Strategies as a running example. The work is categorized under Computer Science > Artificial Intelligence and is available at arXiv:2608.15451.