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

In three linesSAGE is an adaptive gate using von Mises-Fisher density estimation to control memory evolution in agentic LLMs. It classifies candidate facts as ADD (novel), NOOP (redundant), or MERGE (uncertain), reducing expensive LLM calls. On LoCoMo, SAGE cuts API cost by 3.4× and latency by 2.5× with GPT-4o-mini.

Summary generated by Claude — human-verified