UT Dallas Open-Sources Jev-Mem AI Agent Memory Architecture That Cuts Build Time 84.9%

Researchers at the University of Texas at Dallas have released Jev-Mem, an open-source AI agent memory architecture that sidesteps one of the most persistent cost problems in agentic AI: using a full large language model to make every routine memory decision. The paper landed on arXiv on September 21, 2026, and appeared on the Hugging Face Papers trending page the following day, drawing immediate attention from the AI research community.

The core design challenge Jev-Mem addresses is straightforward: as AI agents handle longer tasks, they accumulate memories that must be constantly organized, retrieved, and filtered. Every time an agent needs to categorize an incoming memory, judge its relationships to prior memories, or decide when to stop searching its memory store, conventional systems route that decision through a large autoregressive language model. Generating tokens one by one is slow and expensive — and for structured decisions with bounded outcomes, the team at UT Dallas argues, it is the wrong tool entirely.

Splitting the Work: System-One Controller Handles Memory, System-Two Handles Reasoning

The architecture borrows from cognitive science's dual-process model. System One in human cognition handles fast, automatic decisions; System Two handles slow, deliberate reasoning. Jev-Mem applies this division directly to agent memory. A dedicated System-One control plane, built on the Jev structured-decision library, governs all high-frequency, bounded memory operations. The full language model — System Two — is reserved for open-ended reasoning and producing final answers.

During memory construction, the System-One controller handles memory typing, relational organization, and redundancy filtering. A new observation flows through a sequence: Jev classifies what kind of memory it is, identifies candidate prior memories, judges whether semantic, temporal, causal, or entity relationships exist between them, and updates the memory store accordingly. Because each of these operations has a bounded output — a category label, a probability, a relationship score — they can be dispatched as typed parallel queries rather than generated as free text. There is no sequential token generation on this critical path.

During retrieval, the controller performs query routing, assigns a retrieval budget, executes graph traversal, scores candidates, and applies adaptive stopping — deciding in real time whether enough relevant evidence has been gathered to hand off to System Two, or whether additional retrieval passes are warranted. The result is an iterative loop of route, retrieve, evaluate, and optionally expand, rather than a fixed top-K search that returns a predetermined number of results regardless of their relevance.

Read more: TypeSafe's Jev Delivers Structured AI Decisions Without Generating a Single Word

What the LoCoMo Results Show

On LoCoMo, a benchmark designed to test long-term conversational memory in AI agents, Jev-Mem achieved an overall LLM-as-a-Judge score of 0.777, representing an 11.0% relative improvement over the strongest baseline the paper evaluates. The gains were sharpest on tasks requiring multi-step retrieval and noise filtering: multi-hop question answering reached 0.623 against a best baseline of 0.569, open-domain queries reached 0.618 against 0.517, and adversarial distractor questions reached 0.962 against 0.742.

The efficiency results are where the approach separates most clearly from prior work. Memory construction for the LoCoMo dataset completed in 158 seconds — an 84.9% reduction from the 1,044 seconds required by the fastest competing system. Systems such as A-MEM and MemoryOS required more than 3,000 seconds for the same construction task. On average query latency, Jev-Mem came in at 0.93 seconds, 36.7% lower than the MAGMA baseline at 1.47 seconds, and 46.6% lower than simply feeding the full conversation history to the language model at 1.74 seconds. For context, MemoryOS recorded average query latency of 32.68 seconds — roughly 35 times longer than Jev-Mem.

One piece of context worth noting: MAGMA, the paper's strongest baseline, was published by the same UT Dallas research group earlier in 2026. Jev-Mem is the team's successor system, designed to reduce the computational overhead that MAGMA's purely graph-based retrieval incurs. The improvement over MAGMA is therefore a within-group comparison, not a test against an external competitor. The benchmark figures are internally consistent and reported in the paper's tables, but readers should weigh the self-referential baseline appropriately.

The Multi-Relational Memory Structure

Jev-Mem does not simply attach a vector database to an AI agent. Each memory item is simultaneously represented across four relationship types: semantic, temporal, causal, and entity. This multi-relational structure means a single stored observation can be retrieved through different graph paths depending on what a query actually needs — a causal chain, a timeline of events, a set of entity associations, or pure semantic similarity.

During retrieval, the System-One controller routes the query based on which relationship type best matches the stated need, then allocates a retrieval budget accordingly. Graph traversal follows, with candidates scored and the stopping criterion evaluated at each step. Only after the controller determines that sufficient high-quality evidence has been collected does it pass that evidence to the language model for final synthesis. The language model never sees raw memory dumps; it sees a pre-filtered, relevance-scored context assembled by the controller.

Open-Source Code and Reproducibility

The research team has released the full codebase at github.com/libingzheren/Jev-Mem and an interactive Hugging Face Space demonstrating the system. The paper, authored by Dongming Jiang, Yi Li, and Bingzhe Li from UT Dallas's Department of Computer Science, is available at arXiv preprint 2609.23986 under a Creative Commons BY 4.0 license.

The broader AI agent memory research field is increasingly focused on reducing the compute cost of memory operations alongside improving retrieval quality — a joint optimization that prior work often treated as a secondary concern. Jev-Mem joins a cluster of recent papers, including Zero-Mem, LightMem, and RippleMem, that are exploring architectures where lightweight controllers manage memory pipelines while reserving language model capacity for the tasks that genuinely require open-ended generation. Whether this architectural pattern becomes a production standard will depend on how well it scales to real-world agent workloads beyond the LoCoMo benchmark.