Jev-Mem takes the LLM off the critical path of remembering
Agentic memory systems mostly work like this: when the agent needs to store or fetch something, another LLM call decides how. Jev-Mem (arXiv 2609.23986, submitted September 21) argues that is the expensive mistake, and pulls generation out of the loop.
The frame is System One and System Two. A dedicated System-One control plane — fast, cheap, not autoregressive — handles memory typing and relational organization during construction, then at retrieval time does query routing, retrieval-budget allocation, graph traversal, candidate scoring and adaptive stopping. System Two, the actual LLM, gets invoked only for complex reasoning and final answer synthesis. Between them sits a structured multi-relational memory plane.
The numbers say you do not pay for this in quality. On LoCoMo it hits 0.777 LLM-as-a-Judge, an 11.0% relative improvement over the strongest baseline, while memory construction drops to 158 seconds — a 6.6x speedup over the fastest competing system — and average query latency falls 36.7% to 0.93 seconds. Better and faster, which is the combination that usually means the baseline was doing something structurally silly.
This is now the fifth shipping thing in about a month making the same bet from a different angle: CUA-S1 scoring existing UI elements instead of generating coordinates, Jev returning calibrated probabilities instead of prose, json-render returning a component id instead of markup, Kev putting a pointer head on a small model, and now Jev-Mem replacing a generative memory controller with a classifier. Free-form generation is the most expensive and least verifiable operation a model performs, and five unrelated teams are independently removing it from the parts of the system that run most often.
https://arxiv.org/abs/2609.23986
← Back to all articles
The frame is System One and System Two. A dedicated System-One control plane — fast, cheap, not autoregressive — handles memory typing and relational organization during construction, then at retrieval time does query routing, retrieval-budget allocation, graph traversal, candidate scoring and adaptive stopping. System Two, the actual LLM, gets invoked only for complex reasoning and final answer synthesis. Between them sits a structured multi-relational memory plane.
The numbers say you do not pay for this in quality. On LoCoMo it hits 0.777 LLM-as-a-Judge, an 11.0% relative improvement over the strongest baseline, while memory construction drops to 158 seconds — a 6.6x speedup over the fastest competing system — and average query latency falls 36.7% to 0.93 seconds. Better and faster, which is the combination that usually means the baseline was doing something structurally silly.
This is now the fifth shipping thing in about a month making the same bet from a different angle: CUA-S1 scoring existing UI elements instead of generating coordinates, Jev returning calibrated probabilities instead of prose, json-render returning a component id instead of markup, Kev putting a pointer head on a small model, and now Jev-Mem replacing a generative memory controller with a classifier. Free-form generation is the most expensive and least verifiable operation a model performs, and five unrelated teams are independently removing it from the parts of the system that run most often.
https://arxiv.org/abs/2609.23986
Comments