User repository on GitHub
Jev-Mem: System-One Controlled Agentic Memory
- Forks
- 20
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Jev-Mem
Splits agent memory work in two: a lightweight System-One controller routes queries and allocates search budget across four graph views, and a larger model only writes the final answer.
- Stars
- 198
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- libingzheren
- Last verified
- Oct 2026
About this project
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. Dongming Jiang, Yi Li, and Bingzhe Li · September 2026 · PDF.
Quoted from the project's README. Read the full README.
- Languages
- Python
Highlights
- Jev handles inferred relation judgments and retrieval decisions as Noul and Choice outputs; code enforces the limits.
- Shared memory keeps canonical observations in semantic, temporal, causal and entity views with vector and keyword indexes.
- Traces expose routing, allocated budgets, stopping reasons, cache hits and fallbacks; retrieval halts when evidence suffices.
- A local Laya backend replaces Jev for System-One decisions, so memory work needs no TypeSafe key.
Watch out
Every headline number is the paper's own, measured on LoCoMo with GPT-4o-mini as the answer model and scored by LLM-as-a-Judge, and the README states plainly that its commands are starting points whose subsets and scoring settings do not reproduce the full evaluation. It does not lead everywhere: the strongest baseline scores higher on temporal questions, 0.650 against 0.637, while the gains sit in the other four categories. Live runs send your text to the configured providers and may incur charges, with keys and caches kept local. The Laya backend is not what the paper measured. The captured README stops inside the LoCoMo experiment instructions, so dataset paths and the remaining commands were not visible. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
The bundled demo, yes: it needs no keys and no downloads, uses deterministic mock decisions and embeddings, prints the decision trace and runs offline once dependencies are installed. It also says outright that it does not measure answer quality. Real runs need both a TypeSafe key for the controller and an OpenAI-compatible provider for answers.
No, not for the System-One decisions. Laya handles memory construction and retrieval locally through a PyTorch extra or an MLX extra on Apple Silicon, sharing the graph, retrieval policies and answer model with the Jev path. Answer generation and evaluation still use your configured OpenAI-compatible provider, and weights download on first use.
The division is explicit: Jev judges inferred relations and retrieval choices, surfaced as Noul binary propositions and Choice categorical decisions, while ordinary code validates those typed outputs, handles deterministic links such as timestamp ordering, and enforces the graph and traversal limits. So the model proposes; the code disposes.
Not entirely. On the paper's LoCoMo table it leads four of five question categories and the overall score, and it is far faster to build memory than the nearest competing system, but the strongest baseline edges it on temporal questions. Those are the authors' own measurements with one answer model and one judge, not an independent comparison.
Python 3.11+ and pip install -r requirements.txt, then python -m jev_mem.demo; the demo uses no API keys or model downloads and runs offline with deterministic mock decisions.
Put TYPESAFE_API_KEY and OPENAI_API_KEY in a local .env; Jev controls memory decisions while the OpenAI-compatible model generates answers, and the default minilm embedding backend downloads on first use.
With GPT-4o-mini as the answer model it scored 0.777 overall versus MAGMA's 0.700, built memory in 158 s versus Nemori's 1,044 s, and averaged 0.93 s per query versus MAGMA's 1.47 s.
More Jev resources like this
51GitHub stars
Local LinkedIn recruiting workspace on Browser Use's Jev Ultrafast: Jev screens titles, drives navigation, and assesses profile excerpts against your brief while a live browser stays visible; matches and evidence export as JSON.
Jev Agents & MCP#community#jevbest#catalog-import
Community
skeptrunedev/jev-recruiterPython
skeptrunedev/jev-recruiterPython33GitHub stars
Codex Skill plus stdlib-only Python CLI that recommends which installed skill to read next for a task, allowing no-skill or uncertain answers and never executing candidates.
Jev Agents & MCP#community#jevbest#catalog-import
Community
win4r/jev-skill-suggesterPython
win4r/jev-skill-suggesterPython72GitHub stars
MCP server over stdio that gives coding agents fourteen Jev judgment tools, turning evidence plus an enumerable question into auto, review, or escalate actions from calibrated probabilities.
Jev Agents & MCP#community#jevbest#catalog-import
Community
PyModel/jev-judge-mcpPython
PyModel/jev-judge-mcpPython21GitHub stars
Research runtime that pages a large logical tool and context space through a small bounded Jev decision surface with explicit approvals.
Jev Agents & MCP#community#jevbest#catalog-import
25GitHub stars
Agent-oriented playbook for using Jev as a cheap, fast decision layer that triages, classifies, scores, and gates work before expensive browser, research, retry, or subagent steps.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
Bodila51/muse-jev-playbookPython
Bodila51/muse-jev-playbookPython14GitHub stars
Opt-in Hermes plugin that routes each turn to one skill: a wide Jev Choice plus request checks gate a narrow fits round, injecting a single ignorable hint line.
Jev Agents & MCP#community#jevbest#catalog-import
From the Jev community
Posts from builders shipping with Jev right now.