Learning What to Remember: A Cognitively Grounded Multi-Factor Value Model for Agentic Memory

The paper presents a new multi-factor memory value model for long-running LLM agents, addressing the challenge of memory management within fixed constraints. It introduces a value function that incorporates seven factors from cognitive psychology, allowing for better encoding, forgetting, and retrieval decisions, achieving a retention score of 0.770 on the LongMemEval benchmark compared to lower scores from traditional methods. This model's interpretability and effectiveness in managing memory are crucial for practitioners developing LLMs, as it enhances memory efficiency and decision-making processes in agentic systems.