EMemBench
EMemBench: Interactive Benchmarking of Episodic Memory for VLM Agents
Xinze Li, Ziyue Zhu, Siyuan Liu, Yubo Ma, Yuhang Zang, Yixin Cao, et al. · 2026-01-23
- Modality
- GUI & computer use
- Data
- Simulated
- Task · ground truth
- QA/Act. · Ans.+State
- Scope
- Memory-specific benchmark
What it tests
- CECross-modal evidencePartially evaluated
- LELong-range evidenceDirectly evaluated
- SCState changesDirectly evaluated
- CFConflicting memoriesNot evaluated
- UQUnanswerable queriesDirectly evaluated
- MAMemory-grounded actionDirectly evaluated
- GSPersistent interface statePartially evaluated
- AOAction–outcome associationDirectly evaluated
- XRCross-session reuseNot evaluated
Survey notes
Agent trajectories supply the earlier observations; later programmatically generated questions test recall from those trajectories.