Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents
Jev-Mem moves memory control off the main LLM path, using a lightweight controller to decide what to store, retrieve, and stop.
The paper splits agent memory into a fast “System One” control plane, a structured memory plane, and a slower “System Two” reasoning plane. The controller handles memory typing, query routing, retrieval budgets, graph traversal, scoring, and adaptive stopping, while the larger reasoning model is reserved for harder synthesis. On LoCoMo, the authors report a 0.777 LLM-as-a-Judge score, 11.0% above the strongest baseline. They also report memory construction at 158 seconds, a 6.6x speedup over the fastest competing system, with average query latency down to 0.93 seconds. HF Daily Papers' note
The paper splits agent memory into a fast “System One” control plane, a structured memory plane, and a slower “System Two” reasoning plane. The controller handles memory typing, query routing, retrieval budgets, graph traversal, scoring, and adaptive stopping, while the larger reasoning model is reserved for harder synthesis. On LoCoMo, the authors report a 0.777 LLM-as-a-Judge score, 11.0% above the strongest baseline. They also report memory construction at 158 seconds, a 6.6x speedup over the fastest competing system, with average query latency down to 0.93 seconds. HF Daily Papers' note
score 4