CommunityJev Demos & Experiments
Jev-Mem — huggingface.co
The project page for Jev-Mem, a system-one-controlled agentic memory design from UT Dallas researchers: it reports 0.777 on LoCoMo, 6.6x faster memory construction, and 36.7% lower query latency than the fastest baseline.
- Format
- Playground
- Published by
- Community
- Author
- libingzheren
- Last verified
- Oct 2026
This is one of several Jev listings for the same project. The canonical entry is Jev-Mem — libingzheren.
- Platforms
- Browser
- Capabilities
- Open sourceRuns locally
Highlights
- Reports 0.777 overall LLM-as-a-Judge on LoCoMo, 6.6x faster memory construction, and 36.7% lower average query latency than the fastest competitors.
- A lightweight System-One control plane handles memory typing, relation construction, query routing, and adaptive stopping; System Two handles synthesis.
- Memory is organized into four relational views — semantic, temporal, causal, and entity — managed by the same controller for construction and retrieval.
- The results table lists 158 s build time and 0.93 s query latency for Jev-Mem, against MAGMA's 0.700 score at 1404 s and 1.47 s.
- Written by Dongming Jiang, Yi Li, and Bingzhe Li at the University of Texas at Dallas; code is on GitHub.
Watch out
Self-reported results from the paper authors: the LoCoMo judge score, speedups, and latency numbers come from their own table, with Full Context at 0.481 and the strongest competing system, MAGMA, at 0.700. The Space is a static project page; installation, configuration, and evaluation instructions live in the GitHub repository instead.
Questions about this entry
Short answers drawn from the linked project or page.
A system-one-controlled agentic memory design for AI agents that separates fast, structured memory control from expensive generative reasoning, per the project page.
0.777 overall LLM-as-a-Judge on LoCoMo, memory construction 6.6x faster than the fastest competing memory system, and 36.7% lower average query latency than the fastest memory-based baseline.
It handles high-frequency decisions such as memory typing, relation construction, query routing, retrieval-budget allocation, graph traversal, candidate scoring, evidence assessment, and adaptive stopping.
Only for complex reasoning and final answer synthesis; everything high-frequency stays in the System-One controller, which manages both memory construction and retrieval.
Through four complementary relational views — semantic, temporal, causal, and entity — all managed by the same System-One controller for construction and retrieval.
The project's table lists Jev-Mem at 0.777 with 158 s build and 0.93 s query time, against MAGMA 0.700, Nemori 0.590, A-MEM 0.580, MemoryOS 0.553, and Full Context 0.481.
Dongming Jiang, Yi Li, and Bingzhe Li at the University of Texas at Dallas; the GitHub repository libingzheren/Jev-Mem holds the installation, configuration, and evaluation instructions.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: jevmem keeps Claude Code project memory by asking Jev which turns are worth saving.
avinashjetwani61 pointsJevmem – automatic project memory for Claude Code, built on JevRead the thread on Hacker NewsMore Jev resources like this
Benchmark Heaven's static snapshot of the JevBench v1.6.1 decision-model board: capability and composite scoring over typed-decision accuracy, calibration, latency, and cost, readable without JavaScript.
Jev Demos & ExperimentsPlayground#community#playground#demo
Community
A Gradio playground for KaLM-Jev Small structured judgments: choice, score, noul, and mixed JSON questions return probability tables and full responses without generating answer text on shared ZeroGPU.
Jev Demos & ExperimentsPlayground#community#playground#demo
Community
A Gradio demo of Maincode's 26B MATILDA-jev-v1 one-pass decision model: feed a text or JSON state, optionally an image, and a choice, noul, or score question to get calibrated probabilities over your options.
Jev Demos & ExperimentsPlayground#community#playground#demo
Community
An interactive ZeroGPU demo of AutoTrust's JEV-9B: System 1 answers noul, choice, and score questions in one prefill pass with calibrated probabilities, while an Escalate tab hands low-confidence cases to the untouched Qwen3.5-9B base.
Jev Demos & ExperimentsPlayground#community#playground#demo
Community
AutoTrust's live game-arena demo: watch JEV-27B play Mario, a 3D Rubik's Cube, and Tetris while a side panel reports action probabilities and per-decision inference time, alongside a decision lab and a ViZDoom arena.
Jev Demos & ExperimentsPlayground#community#playground#demo
Community
A three-lane demo of pngwn's System One 4B scorer: one prefill plus a batched forward for every option branch, a naive per-option equivalence check, and a Qwen3.5-4B instruct baseline generating JSON.
Jev Demos & ExperimentsPlayground#community#playground#demo
Community
From the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.