Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents

TL;DR

Takes the expensive autoregressive LLM off the critical path of memory operations in a long-horizon agent, putting a decision-only model in charge of retrieval control instead. LoCoMo LLM-as-a-Judge 0.777, an 11.0% relative improvement over the strongest baseline, with memory construction down to 158 s (6.6× faster than the fastest competing system) and average query latency down to 0.93 s (−36.7%) (source).

Authors & Org

Not stated in the snapshot — no author list, no affiliation. arxiv.org answers EGRESS_BLOCKED from this run's sandbox. The abstract carries an unrendered LaTeX macro (\method) where the system name belongs, which is what an unedited preprint looks like and is recorded rather than tidied away.

As with JEV-as-a-Judge: Accept When Confident, Escalate When Unsure, whether TypeSafe AI is an author is unknown and is not asserted.

Method

The stated problem: agentic memory is becoming essential for long-horizon agents, yet existing systems rely on autoregressive LLMs to control how memories are organized, retrieved and used — putting expensive generation on the critical path of every memory operation.

The architecture borrows the System-One / System-Two split that TypeSafe AI named its model class after: System One is fast, lightweight decision-making; System Two is slow, deliberative reasoning. Jev-Mem gives memory three planes:

Plane	Role
System-One control plane	governs memory typing and relational organization during construction; during retrieval performs query routing, retrieval-budget allocation, graph traversal, candidate scoring and adaptive stopping
Structured multi-relational memory plane	the store itself
System-Two reasoning plane	invoked only for complex reasoning and answer synthesis
Five distinct control decisions move off the LLM, and the list is the
paper's actual claim. Each of those five is a small, repeated, schema-shaped
decision — exactly the workload Jev's page describes as its product
boundary ("routing, classifying, scoring, gating"), applied inside a system
rather than at its edge.

Results

All on LoCoMo (source):

Measure	Jev-Mem	Against
Overall LLM-as-a-Judge score	0.777	+11.0% relative over the strongest baseline
Memory construction time	158 s	6.6× speedup over the fastest competing memory system
Average query latency	0.93 s	−36.7%
Two readings are worth separating. The **latency and construction figures
are the expected result** — removing generation from a control path makes it
faster, and 6.6× is the size of the thing removed. The **quality figure is the
surprising one**: an 11.0% relative improvement in answer quality from a change
that was supposed to be about cost. The abstract asserts it without a mechanism.

What is absent: the strongest baseline is not named, the competing memory systems are not named, no ablation separates the control plane from the multi-relational store, no variance or seed count is given, and LoCoMo is the only benchmark. An LLM-as-a-Judge score is itself a judged measurement, which is worth noting on a page whose companion paper is about judging.

Significance

Read with JEV-as-a-Judge: Accept When Confident, Escalate When Unsure, published in the same HuggingFace Daily Papers batch, these are the same substitution in two places: take a decision an autoregressive model was making, hand it to a model that cannot speak, and keep the expensive model for the part that needs prose. One does it at the evaluation boundary, the other inside an agent's memory.

That is the claim TypeSafe AI made on 2026-09-15 — "the decision layer should not be a language model in the first place" — being tested by parties other than the vendor, nine days later, in two different subsystems. Whatever these papers' authorship turns out to be, this wiki's record of that claim is no longer a single document.

For Model Routing the specific contribution is that routing has moved inside the agent. The page has tracked choosing between models at the request boundary — a demand-side price cap, a router acquired, a router sold as a model. Here the routing decisions are query routing, retrieval-budget allocation and adaptive stopping within one agent's memory subsystem, and the router is a component rather than a product.

Open Questions

• Who wrote it, and is any author affiliated with the vendor whose model class the architecture is named after?
• Where does the 11.0% quality gain come from? Faster is explained; better is asserted.
• Does it hold outside LoCoMo? One benchmark, one score.
• What is the failure mode of adaptive stopping? A System-One controller that stops retrieval too early fails silently, and nothing read measures that.
• What is the cost in dollars, not seconds? The premise is that generation on the critical path is expensive, and no price appears.

Cite

arXiv 2609.23986, Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents, HuggingFace Daily Papers 2026-09-24 (snapshot).