Agent memory design reports 6.6x faster construction by skipping model calls

single source· 1 articles · confidence: medium · first seen 2026-09-20 20:00 UTC

What this means for you

Nothing to install: no code, weights or API, and the figures are the authors' own on a public benchmark, so treat them as a claim. If you run long-horizon agent memory, the idea worth taking is the split — routing, budgeting and stopping in a cheap controller, the model called only to synthesise.

A preprint posted to arXiv on 20 September 2026 describes Jev-Mem, a design for agentic memory (the store an agent reads its past interactions from) that keeps a language model off the critical path of storing and fetching. A lightweight controller handles memory type, query routing, retrieval budget, graph traversal, candidate scoring and when to stop; the model is called only for complex reasoning and answer synthesis. The authors report a LoCoMo LLM-as-a-Judge score of 0.777, 11.0% above the strongest baseline, construction at 158 seconds (6.6x faster than the fastest competing system) and query latency of 0.93 seconds, 36.7% lower. No evaluation date, harness detail or code release is given.

Key facts

  • ·Jev-Mem splits agentic memory into a System-One control plane, a multi-relational memory plane and a System-Two reasoning plane, with the language model invoked only for complex reasoning and answer synthesis. source
  • ·The paper reports a LoCoMo LLM-as-a-Judge score of 0.777, an 11.0% relative improvement over the strongest baseline. source
  • ·Memory construction is reported at 158 seconds, a 6.6x speedup over the fastest competing memory system. source
  • ·Average query latency is reported at 0.93 seconds, a 36.7% reduction. source
  • ·The preprint is arXiv 2609.23986, posted 20 September 2026, and states no evaluation date, harness detail or code release. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire