Jev-Mem splits AI agent memory into fast and slow control planes

Researchers introduce Jev-Mem, an agentic memory architecture for long-horizon AI agents built around the System-One and System-Two split from cognitive science, where System One handles fast, lightweight decisions and System Two performs slower, deliberative reasoning. The authors argue that many existing agentic memory systems rely on autoregressive LLMs to control how memories are organized, retrieved and used, which puts expensive language-model generation directly on the critical path of every memory operation, including simple ones. Jev-Mem instead separates the work across three components: 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 while memories are being built, and during retrieval it dynamically handles query routing, retrieval-budget allocation, graph traversal, candidate scoring and adaptive stopping. System Two, the heavier reasoning component, is invoked only for complex reasoning and answer synthesis, not for the routine steps System One already covers. On the LoCoMo benchmark, Jev-Mem reaches an overall LLM-as-a-Judge score of 0.777, an 11.0% relative improvement over the strongest baseline. It also cuts memory construction time to 158 seconds, a 6.6x speedup over the fastest competing memory system, and lowers average query latency to 0.93 seconds, a 36.7% reduction. The authors present this as evidence that moving routine memory management off the language model and onto a dedicated fast controller improves both answer quality and system efficiency at once, rather than trading one for the other.

Key facts

Jev-Mem splits agentic memory into a fast System-One control plane, a structured multi-relational memory plane, and a slower System-Two reasoning plane.
System One handles memory typing and organization during construction, plus 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, not for routine memory operations.
On LoCoMo, Jev-Mem scores 0.777 (LLM-as-a-Judge), an 11.0% relative improvement over the strongest baseline.
Memory construction time drops to 158 seconds (6.6x speedup over the fastest competing system), and average query latency falls to 0.93 seconds (a 36.7% reduction).

Why it matters

Most agentic memory systems lean on an autoregressive LLM to decide how memories get organized, retrieved and used, which the authors say places expensive generation on the critical path of memory operations that do not need it. Jev-Mem's pitch is that a dedicated fast controller can take over that routine work, freeing the heavier reasoning model to run only when a question actually needs deliberation.

Who it affects

The design targets long-horizon AI agents, systems that need to accumulate and reuse memory across extended interactions rather than a single exchange. Teams building or evaluating agent frameworks with persistent memory are the direct audience.

How to use it

The paper presents an architecture and benchmark results rather than a packaged product, tool or pricing; the source gives no release, licence or availability details, so there is nothing to install or license yet from what is stated.

How solid is it

The results come from a single benchmark, LoCoMo, evaluated with an LLM-as-a-Judge score rather than a fixed ground-truth metric, and measured against baselines the source does not name. The gains reported, an 11.0% relative score improvement, a 6.6x construction speedup and a 36.7% latency reduction, are all against unnamed comparison systems, so the size of the improvement cannot be checked against a specific known baseline from the text alone.

Risks and caveats

The source does not name the paper's authors, institutions, publication venue or date, nor does it describe what LoCoMo actually tests or which systems serve as the 'strongest baseline' and 'fastest competing memory system'. Results reported on one benchmark against unspecified competitors may not generalize to other memory-heavy agent workloads.