Researchers introduced Jev-Mem, a new agentic memory architecture. It splits memory tasks between a fast System-One controller and a slower System-Two reasoning plane. System One handles memory typing, organization, query routing, retrieval budget, graph traversal, candidate scoring, and adaptive stopping. System Two is used only for complex reasoning and answer synthesis. On the LoCoMo benchmark, Jev-Mem scored 0.777 on LLM-as-a-Judge. It improved over the strongest baseline by 11.0%. Memory construction time dropped to 158 seconds. This is a 6.6× speedup over the fastest competing system. Average query latency fell to 0.93 seconds. This is a 36.7% reduction.
What changed
Existing agentic memory systems relied on autoregressive LLMs for memory control, placing expensive generation on the critical path.
What it unlocks
Faster memory construction and lower query latency for long-horizon AI agents.
- 0.777 LLM-as-a-Judge score
- 11.0% relative improvement
- 158s memory construction time
- 6.6× speedup
Sources