Tool Graph Memory Provides Diminishing Returns Compared to Textual Insights

Evaluation of nine models shows that tool graph memory rarely improves agent performance and often costs twice the LLM calls. Replacing transition-based memory with textual insights achieves similar or better results with significantly lower computational overhead.