Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents
Dongming Jiang, Yi Li, Bingzhe Li
The University of Texas at Dallas
Published September 21, 2026 · on the feed September 22, 2026 · ▲ 26 · arXiv 2609.23986
The abstract
Agentic memory is becoming essential for long-horizon AI agents, yet many existing systems rely on autoregressive LLMs to control how memories are organized, retrieved, and used, placing expensive generation on the critical path of memory operations. We introduce \method, a new agentic memory architecture inspired by System-One/System-Two cognition. System One captures fast, lightweight decision-making, whereas System Two performs slower, deliberative reasoning. Jev-Mem brings this division of labor to agentic memory through a dedicated System-One control plane, a structured multi-relational memory plane, and a System-Two reasoning plane. The System-One controller governs memory typing and relational organization during construction, and dynamically performs query routing, retrieval-budget allocation, graph traversal, candidate scoring, and adaptive stopping during retrieval. System Two is invoked only for complex reasoning and answer synthesis. This design improves both memory effectiveness and system efficiency: on LoCoMo Jev-Mem achieves an overall LLM-as-a-Judge score of 0.777, an 11.0\% relative improvement over the strongest baseline, while reducing memory construction time to 158\,s, a 6.6times speedup over the fastest competing memory system, and lowering average query latency to 0.93\,s, a 36.7\% reduction.
The Brief
2:24A short podcast summary of the paper. Written and voiced by AI.
Transcript
Jev-Mem. System-One-Controlled Agentic Memory for Efficient AI Agents. By Dongming Jiang and two others, from the University of Texas at Dallas. Published on the twenty-first of September. Its code repository is worked out: about fifty-five source files, with a readme and a licence.
Long-horizon AI agents increasingly need agentic memory. Many existing memory systems, though, hand control of that memory to an autoregressive LLM. The language model decides how memories are organized, how they are retrieved, and how they are used. That places expensive generation on the critical path of every memory operation. Every time the agent stores something or looks something up, it waits on a full generation step, and over a long horizon those waits are paid again and again.
Jev-Mem borrows the division between System One and System Two cognition. System One is fast and lightweight in its decisions. System Two is slower and deliberative. Jev-Mem applies that division of labor to memory, with three planes: a dedicated System-One control plane, a structured memory plane that is multi-relational, and a System-Two reasoning plane.
The System-One controller does most of the running. When memory is being built, it governs how memories are typed and how they are organized relationally. At retrieval time, it dynamically routes queries, allocates the retrieval budget, traverses the graph, scores candidates, and decides when to stop. System Two is only called in for complex reasoning and for synthesizing the answer. Put simply, the slow model is no longer the one managing the memory. It is a consumer of what the fast controller retrieves.
The design is meant to improve effectiveness and efficiency together, and the results on LoCoMo address both. Jev-Mem reaches an overall LLM-as-a-Judge score of zero point seven seven seven, an eleven percent relative improvement over the strongest baseline.
On cost, memory construction takes one hundred and fifty-eight seconds, a speedup of six point six times over the fastest competing memory system. Average query latency falls to zero point nine three seconds, a reduction of thirty-six point seven percent.
The architecture's bet is about who should be doing what. Routine decisions about memory, such as where to look, how much to fetch and when to stop, go to a fast controller, and the expensive reasoning model is saved for the moments that need it. On this benchmark, that reallocation raised quality while cutting both build time and query time.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here it is an assistant that thinks hard about every tiny filing decision.
You run a long-lived assistant that keeps notes on everything users tell it, far more than fits in its context window. Each time it stores or looks up a note, a large language model decides what kind of note it is and where to search.
The road ahead
- 1 · The problem ~7 min An assistant that keeps long-term notes asks a big language model to make every small filing and lookup decision, and those generations make memory slow.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Memory decisions only need labels, probabilities or scores, so a fast lightweight controller makes them, and the large model just reasons and writes answers.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min The fast controller types and links notes on storing, then routes, budgets, walks, scores and stops on searching, before the large model writes the answer.
- 6 · The experiment ~6 min You’ll guess: How much faster was building memory with Jev-Mem than with the fastest competing memory system?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.