SAGE: A Novelty Gate for Efficient Memory Evolution in Agentic LLMs

The article introduces SAGE (Spherical Adaptive Gate for memory Evolution), a novelty detection mechanism for agentic LLMs that optimizes memory evolution by scoring candidate facts using a von Mises-Fisher density estimator. SAGE effectively categorizes facts as ADD, NOOP, or uncertain, significantly reducing write-time reasoning and achieving a 3.4× reduction in API costs and 2.5× lower latency during the add phase on GPT-4o-mini. This approach enhances memory quality and system efficiency across various models by decreasing LLM calls by approximately 16-18% with minimal impact on output quality, making it a valuable tool for practitioners focused on memory management in LLMs.