Jev-Mem: Fast System-One Controllers for Agentic Memory
Dongming Jiang, Yi Li, Bingzhe Li · UT Dallas · arXiv, September 2026 · benchmark: LoCoMo · answerer: gpt-4o-mini
arXiv:2609.23986
Interactive page
Three planes
Hover a plane. Management decisions leave the generative loop.
Hover a plane.
Who decides what?
Illustrative baseline
Toggle between an all-LLM controller (conceptual) and the Jev-Mem assignment described in the episode.
Not the same as distillation
Distilling System 2 into System 1 changes one model; Jev-Mem assigns each decision type.
Write path: Mira's bicycle
Click the numbered steps, or use the buttons.
◀ Prev
Next ▶
Edge thresholds are from the paper (0.60). Score values below are
Illustrative inputs
Consolidation gate
Illustrative inputs
Every 20 writes. System Two summarizer runs only if P(merge or promote) ≥ 0.85 and contradiction < 0.85.
Read loop
Step through the retrieval loop.
◀ Prev
Next ▶
Stop rule
Illustrative inputs
Stop if sufficiency ≥ 0.95 with missing and contradiction both < 0.15, or if continuing looks useless (< 0.15).
Expansion budget of 80
Illustrative inputs
Views with need ≥ 0.10 are active; budget split by need, at least one each.
Hard bounds
Hitting any can end retrieval with incomplete evidence.
Reported results
Single run on LoCoMo; no variance, no ablation.
Evidence board
Hover or tap a cell.
Hover a cell.
Local small model vs vendor API
Per-call latency, calls per write/query
Zero-Mem baseline
Graph vs controller ablation
LongMemEval + variance
What is Jev?
Host's inference
The paper gives no architecture, size, hardware or cost. These two readings come from the interface, not from the paper.
References