The OmniAI team at Zhejiang University has introduced the Embodied Memory Evaluation Benchmark (EMem-Bench) to address the challenge of memory association in embodied agents during task execution. This benchmark includes 2,554 executable interaction samples covering four types of tasks. Additionally, the team has proposed an external memory system, Embodied-Memorizer, along with a fine-tuned EMem-8B model. Within just three weeks of its release, the benchmark has been downloaded nearly 30,000 times. Experimental results indicate that current models exhibit uneven performance in embodied memory capabilities, with simply stacking historical images reducing task success rates. In contrast, EMem significantly enhances task performance across different models, with three complementary types of memory each demonstrating distinct strengths. Currently, related work has been primarily validated in simulated environments, and future efforts will further test its robustness in longer-duration, more complex environments, and real-world robotic scenarios.