Jev-Mem
Agent memory system where Jev or local Laya controls memory organisation, query routing, retrieval budgets and stopping.
Overview
Jev-Mem is an unofficial agentic memory system for long-running agents, with a September 2026 arXiv paper. It separates the frequent decisions of memory management from answer reasoning: a System-One controller (Jev, or the local Laya model added later) organises memories and guides retrieval, while a System-Two language model writes the answer from the evidence found.
Memories keep original observations, timestamps and provenance, and shared nodes take part in semantic, temporal, causal and entity graph views. For each query the controller routes across relevant views, allocates search budgets, scores candidates and decides after each round whether more evidence is needed. Traces expose routing, budgets, stopping reasons, cache hits and fallbacks. On LoCoMo with GPT-4o-mini, the paper reports 11.0% higher overall answer quality, 6.6x faster memory construction and 36.7% lower query latency than the strongest or fastest baseline for each metric.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which graph view (semantic, temporal, causal, entity) should this query search?(inferred) | semantic · temporal · causal · entity |
| Score | How relevant is this candidate memory to the query?(inferred) | — |
| Yes / no | Is enough evidence found to stop retrieving?(inferred) | yes · no |
How it works
- Store each observation with timestamps and provenance in shared graph views
- For a query, the controller routes to relevant views and sets budgets
- Candidates are scored and the controller decides whether to keep searching
- The LLM synthesises an answer from the collected evidence
Strengths
- Frequent decisions offloaded to a fast controller
- Works with hosted Jev or local Laya
- Inspectable traces of routing, budgets and stopping
- Paper, code and a Hugging Face demo
Caveats
- Results are the paper's own on one benchmark (LoCoMo)
- Answer quality was judged by an LLM-as-judge
- Research code rather than a packaged product
Getting started
See the Quick start section of the README, including the Use local Laya option.At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- Library / research code
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Unofficial agent memory system where Jev or local Laya controls memory organization, query routing, retrieval budgets, candidate scoring, and stopping while an LLM writes answers; its [paper](https://arxiv.org/abs/2609.23986) reports LoCoMo results
AnotiaWang/awesome-decision-models\*\*Better memory for long-running AI agents—with fast decisions and focused reasoning.\*\*
logicrw/awesome-jev-projectsJev-Mem: System-One Controlled Agentic Memory.
hellogumbo/awesome-jevMemory layer for long-running agents that routes the frequent memory-management decisions to Jev and leaves the deeper reasoning to the LLM. Comes with a paper.
MrJev/awesome-jev
Frequently asked questions
What does the System-One controller decide?
How memories connect, which graph views to search, how much to search, how to score candidates and when to stop.
Can it run without TypeSafe?
Yes. Since September 2026 it supports the local Laya model for the same decisions.
What results does the paper report?
On LoCoMo with GPT-4o-mini: 11.0% higher answer quality, 6.6x faster memory construction and 36.7% lower query latency than baselines.
Related projects
Filters an assistant's memories by relevance, not resemblance: one Jev yes/no per memory in a single request, kept above a threshold.
ProjectbetaPythonSelf-hosted memory MCP server on Cloudflare Workers where a Jev layer judges memory type, salience and durability and re-ranks recall.
MCP serverprototypeTypeScriptLocal long-term memory for coding agents with an optional Jev reranker that orders recalled memories.
ProjectproductionTypeScriptVolcengine's context database for agents, with a Jev rerank client that scores candidate documents with jev-latest as relevance.
LibraryproductionPythonMulti-user MCP 'agents' cloud' on Cloudflare Workers with an optional search stage that reranks candidates with Workers AI Jev scores.
MCP serverbetaTypeScriptZilliz's cross-agent Markdown memory with an optional Jev reranker that asks yes/no relevance questions about retrieved chunks.
LibraryproductionPython