Zhejiang University releases EMem-Bench to test embodied agents' memory of past observations

Zhejiang 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.