Jev-Mem puts a System One controller on the critical path of agent memory instead of an autoregressive LLM

arXiv / HuggingFace Daily Papers·medium signalXBlueskyLinkedInCopy link

Most agentic memory systems use an autoregressive LLM to decide how memories are organized, retrieved and used, which puts expensive token generation on the critical path of every memory operation. Jev-Mem splits this into a System One control plane, a structured multi-relational memory plane, and a System Two reasoning plane, with the fast controller handling memory typing and relational organization at construction time and then query routing, retrieval-budget allocation, graph traversal, candidate scoring and adaptive stopping at retrieval time. It is the first research paper to build directly on the decision-model shape Jev introduced last week, rather than cloning its API.