Jev-Mem cuts agent memory latency 36.7%

Jev-Mem, a University of Texas at Dallas agent-memory architecture, reports 36.7% lower average query latency—down to 0.93 seconds—and 6.6× faster memory construction than the fastest competing system on LoCoMo, according to the preprint. It routes memory typing, retrieval budgets, graph traversal, and stopping through a lightweight controller, calling a full reasoner only for synthesis; its LLM-as-Judge score was 0.777, an 11% relative gain over the strongest cited baseline. Builders designing long-lived agents should benchmark memory control separately from answer generation; Temporal’s 40-fold agent-history storage gap is a reminder that cheaper reads do not erase retention cost.