⚡ SYSTEM-ONE AGENTIC MEMORY

Jev-Mem: System-One-Controlled Agentic Memory

Fast, structured memory control powered by Jev — leaving complex reasoning to System Two.

6.6× Faster Construction 36.7% Lower Query Latency 0.777 LLM Judge Score

Why Jev-Mem?

Modern agentic memory systems repeatedly decide what to store, how memories should be related, where to search, which candidates matter, how much retrieval is needed, and when retrieval should stop. Jev-Mem moves these frequent structured decisions out of the autoregressive generation loop.

⚡ System One

Memory typing
Relation construction
Query routing
Retrieval-budget allocation
Graph traversal
Candidate scoring
Evidence assessment
Adaptive stopping

🧠 System Two

Complex reasoning
Evidence synthesis
Final answer generation

Architecture

Observation / Query ↓ System-One Controller ↓ Structured Multi-Relational Memory • Semantic • Temporal • Causal • Entity ↓ Adaptive Retrieval + Evidence Check ↓ System-Two Reasoning ↓ Final Answer

How Jev Is Used

Jev provides the concrete System-One controller used by Jev-Mem. Instead of asking an autoregressive LLM to generate and parse free-form text for every memory operation, Jev answers explicit, typed decision questions with structured scores.

Typed Decisions

Jev-Mem sends a structured state together with a batch of explicit questions. Binary decisions are represented as Noul queries, while mutually exclusive alternatives can be represented with Choice.

State: observation / query / evidence / candidates Jev questions: Is this memory episodic? Is this pair causally related? Does this query need temporal retrieval? Is this candidate relevant? Is the evidence sufficient?

Why It Helps

Many memory-control operations only need a bounded output such as a label, probability-like score, or choice. Jev handles these decisions directly, avoiding unnecessary token-by-token generation on the memory critical path.

Decisions that share the same state can also be evaluated together in a batch, helping Jev-Mem keep memory construction and retrieval lightweight.

Jev on the Write Path

During memory construction, Jev is used to annotate observations and determine how new memories should connect to existing ones. For example, it can score whether an observation is episodic, semantic, procedural, or preference-related, and evaluate semantic or causal relations between candidate memory pairs.

Observation ↓ Jev memory typing ↓ Candidate search ↓ Jev relation decisions ↓ Insert into structured memory

Jev on the Read Path

During retrieval, Jev routes the query across semantic, temporal, causal, and entity relations, scores candidates discovered during graph expansion, and evaluates whether the current evidence is sufficient or whether another retrieval round is likely to help.

Query ↓ Jev routing ↓ Retrieve anchors ↓ Jev evidence check ↓ Expand + Jev candidate scoring ↓ Stop when evidence is sufficient ↓ System Two generates the final answer

In short, Jev controls memory; System Two reasons over the selected evidence. This separation is the core design principle behind Jev-Mem.

LoCoMo Results

Method	Overall ↑	Build Time ↓	Latency ↓
Full Context	0.481	N/A	1.74 s
A-MEM	0.580	3636 s	2.26 s
MemoryOS	0.553	3276 s	32.68 s
Nemori	0.590	1044 s	2.59 s
MAGMA	0.700	1404 s	1.47 s
Jev-Mem	0.777	158 s	0.93 s

Key Idea

Use generation when generation is needed. Use lightweight structured prediction when only a decision is needed.

Code

GitHub: https://github.com/libingzheren/Jev-Mem

git clone https://github.com/libingzheren/Jev-Mem.git
cd Jev-Mem

Paper

Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents

Dongming Jiang, Yi Li, Bingzhe Li
Department of Computer Science, The University of Texas at Dallas

arXiv: arXiv:2609.23986

Code: github.com/libingzheren/Jev-Mem