UT Dallas' Jev-Mem cuts agent query latency 36.7% on LoCoMo

Jev-Mem, a new agentic-memory architecture from researchers at the University of Texas at Dallas, reports an overall LLM-as-a-Judge score of 0.777 on the LoCoMo benchmark, an 11.0% relative improvement over the strongest baseline, while cutting average query latency to 0.93 seconds, a 36.7% reduction, according to a preprint indexed on Hugging Face.

The complaint underneath the design: today's agent-memory systems keep calling a full LLM for small housekeeping decisions, typing a new memory, choosing which store to query, deciding when to stop searching. That habit, the authors write, is "placing expensive generation on the critical path of memory operations."

Jev-Mem splits the work along the lines of the dual-process cognition literature. A lightweight "System One" control plane handles the high-frequency, structured decisions: memory typing, redundancy filtering, and construction of semantic, temporal, causal, and entity relations on the write path, then query routing, retrieval-budget allocation, graph traversal, candidate scoring, and adaptive stopping on the read path. "System Two is invoked only for complex reasoning and answer synthesis."

On memory construction the paper reports 158 seconds, a 6.6x "speedup over the fastest competing memory system." Code sits at github.com/libingzheren/Jev-Mem.

The abstract states its LoCoMo win only as a relative gain over "the strongest baseline" and does not name that baseline in the headline number. The work lands in a busy stretch of research coverage on lower-cost agent scaffolding on our research feed, alongside KAIST-Google's Declarative Attention result on cutting KV reads.