Total Recall at What Cost? Benchmarking the Serving Cost of Agentic Memory Systems

TL;DR - This study benchmarks the cost and accuracy of three agentic memory systems over conversations up to 400 turns. Memory can reduce transcript-serving costs, but savings vary greatly by implementation and model, with no system maximizing both cost efficiency and accuracy.

• Conversation length and message size underestimate memory-system costs by 18–69% because internal memory behavior is a major cost driver.
• Break-even points range from the first tens of turns to never within 400 turns, depending on the memory system and backbone.
• Accuracy ranges from 21–54% across 665 LoCoMo questions.
• Backbone selection affects serving cost as much as memory-system choice.