Researchers have introduced Jev-Mem, an agentic memory architecture that borrows from the System-One/System-Two view of cognition to keep expensive language-model generation off the critical path of memory operations. The work was posted as a paper on Hugging Face.
The paper argues that many current agentic memory systems depend on autoregressive LLMs to decide how memories are organized, retrieved, and used, which makes costly generation a bottleneck. Jev-Mem instead assigns that work to a dedicated System-One control plane, paired with a structured multi-relational memory plane and a System-Two reasoning plane.
In this design, the System-One controller handles memory typing and relational organization during construction, and during retrieval it performs query routing, retrieval-budget allocation, graph traversal, candidate scoring, and adaptive stopping. The slower System-Two component is invoked only for complex reasoning and answer synthesis.
On the LoCoMo benchmark, the authors report an overall LLM-as-a-Judge score of 0.777, an 11.0% relative improvement over the strongest baseline. They also report memory construction time of 158 seconds, described as a 6.6x speedup over the fastest competing memory system, and average query latency of 0.93 seconds, a 36.7% reduction.
The team says code and an interactive Hugging Face Space have been released to support reproducibility and further research.