0561GitHub
Jev-Mem: typed decisions for agent memory and retrieval
Jev connects memories, routes retrieval across graph views, scores candidate evidence, and decides when the answer model has enough context.
Public controller code and the paper-linked results were inspected on October 1, 2026. The memory graph preserves original observations and provenance; an LLM synthesizes answers from selected evidence. On LoCoMo with GPT-4o-mini, the authors report an overall judged score of 0.777 versus MAGMA’s 0.700. The implementation also supports a local Laya controller. These are paper-reported measurements, not independently reproduced results.
libingzheren/Jev-MemREADME ↗
# Jev-Mem: System-One-Controlled Agentic Memory [](https://arxiv.org/abs/2609.23986) [](https://huggingface.co/papers/2609.23986) [](https://huggingface.co/spaces/libingzheren/Jev-Mem) **Better memory for long-running AI agents—with fast decisions and focused reasoning.** **Paper:** [Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents](https://arxiv.org/abs/2609.23986) Dongming Jiang, Yi Li, and Bingzhe Li · September 2026 · [PDF](https://arxiv.org/pdf/2609.23986) Jev-Mem separates the frequent decisions of memory management from the deeper reasoning needed to answer a question. A lightweight **System-One controller** organizes memories and guides retrieval across semantic, temporal, causal, and entity relations. A **System-Two language model** synthesizes the answer from the evidence it finds. ## 🔥 News 🔥 - **2026-09-27 · Laya integration.** Jev-Mem now supports [Laya](https://huggingface.co/convaiinnovations/laya) for local System-One decisions. The same memory pipeline works with either Jev or Laya. [Get started ↓](#use-local-laya) - **2026-09-21 · Jev-Mem launch.** Our [paper](https://arxiv.org/abs/2609.23986) is on arXiv, with [code on GitHub](https://github.com/libingzheren/Jev-Mem) and a [demo on Hugging Face](https://huggingface.co/spaces/libingzheren/Jev-Mem). On LoCoMo with **GPT-4o-mini**, the paper reports **11.0% higher overall answer quality**, **6.6× faster memory construction**, and **36.7% lower query latency** than the strongest or fastest baseline for each metric. See [results](#results-on-locomo) for the comparisons. [Results](#results-on-locomo) · [How it works](#how-it-works) · [Quick start](#quick-start) · [Run experiments](#run-experiments) · [Contribute](#contributing) · [Citation](#citation) ## Why Jev-Me