Zhejiang University releases EMem-Bench to test embodied AI agent memory

Zhejiang University's OmniAI Group has released EmbodiedMemory-Bench (EMem-Bench), a benchmark of 2,554 episodes across 1,118 scenes that tests whether embodied agents can build and reuse memory from long interaction histories. Gemini-3-Flash led with 64.2% success in the full-context setting, while the team's external memory system EMem raised GPT-5.4-mini's success from 35.3% to 58.9%.

• EMem-Bench spans 2,554 episodes and 1,118 scenes across four memory task families
• Gemini-3-Flash leads with 64.2% success, GPT-5.4 reaches 56.8%
• EMem memory system lifts GPT-5.4-mini success from 35.3% to 58.9%
• Adding more history images hurt: success fell from 36.6% to 26.2%