Zhejiang University releases EMem-Bench to test memory in embodied AI agents
Overview
Zhejiang University's OmniAI Group releases EmbodiedMemory-Bench (EMem-Bench), a benchmark of 2,554 executable episodes that tests whether embodied agents can build memory from long interaction histories and act on it.
In the full-context setting, Gemini-3-Flash scores highest with a 64.2% average success rate, and all open-source models but one score below 45%. The team's EMem memory system raises GPT-5.4-mini's average success from 35.3% to 58.9%, according to the paper.
All results come from simulation. The paper was posted to arXiv about three weeks before Chinese media covered it.
Written by AI from the articles below · updated Oct 18, 6:06 AM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- PandailyNewsZhejiang University releases EMem-Bench to test embodied agents' memory of past observations
AIZhejiang University's OmniAI Group released EmbodiedMemory-Bench (EMem-Bench), a benchmark of 2,554 executable episodes testing whether embodied agents can build memory from long interaction histories. In the full-context setting, Gemini-3-Flash led with a 64.2% average success rate, and all but one open-source model scored below 45%. The team's EMem memory system raised GPT-5.4-mini's average success from 35.3% to 58.9%, but all results come from simulation.
Heat trend
Not enough continuous observations to show a trend yet.