Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents (academy.dair.ai)

Researchers have unveiled Jev-Mem, a novel agentic memory architecture designed to enhance the efficiency of long-horizon AI agents. Unlike existing systems that heavily rely on autoregressive large language models (LLMs) for memory management, Jev-Mem introduces a dual-system approach inspired by cognitive science. It features a System-One control plane for rapid decision-making, a structured multi-relational memory plane, and a System-Two reasoning plane for more complex tasks. This design allows for efficient memory organization, retrieval, and usage, reducing the burden on autoregressive generation during critical memory operations. Jev-Mem significantly advances AI capabilities by improving memory effectiveness and overall efficiency. In practical applications, it achieved an impressive LLM-as-a-Judge score of 0.777, marking an 11% improvement over previous state-of-the-art models. Additionally, it boasts a memory construction time of just 158 seconds, which is a 6.6 times speedup compared to the fastest competing systems, and lowers average query latency to 0.93 seconds—a 36.7% reduction. This architecture not only streamlines AI memory operations but also paves the way for more sophisticated reasoning in artificial intelligence, making it a significant leap forward for the AI/ML community.