Cost and accuracy of long-term memory in Distributed Multi-Agent Systems based on Large Language Models

Benedict Wolff, Jacopo Bennati

We benchmark three long-term-memory frameworks (cognee, Graphiti, Mem0) against RAG and full-context baselines on LoCoMo, in a realistic cloud-edge deployment. The paper asks whether a memory framework's extra machinery actually buys better answers, mapping each on the cost vs. accuracy Pareto front and testing how they hold up when the edge↔cloud link degrades.