Hugging Face Daily PapersDongming Jiang, Yi Li, Bingzhe Li1 min readpaperadvanced
Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents
Summary
Jev‑Mem splits agentic memory work between a fast System‑One controller (handling typing, organization, routing, budgeting, graph traversal, scoring, and stopping) and a slower System‑Two reasoning plane, cutting memory‑construction time 6.6×, query latency 36.7% and boosting LLM‑as‑a‑Judge score 11% on LoCoMo.
- System‑One control plane removes LLM generation from the hot path of memory ops, handling most retrieval decisions with lightweight logic.
- Memory is stored as a structured multi‑relational graph; System‑One dynamically allocates retrieval budget and routes queries through the graph.
- System‑Two is only invoked for complex reasoning or answer synthesis, reducing overall LLM calls.
- Empirical results on the LoCoMo benchmark: 0.777 LLM‑as‑a‑Judge (11% relative gain), 158 s construction time (6.6× faster than best prior), 0.93 s average query latency (36.7% lower).
Long‑horizon agents need cheap, fast memory access; Jev‑Mem shows a concrete architecture that offloads most work from expensive LLM inference, delivering measurable speedups and effectiveness gains—useful for anyone building autonomous agents or retrieval‑augmented generation pipelines.
8/10
