The introduction of MLUBench addresses the critical need for effective lifelong unlearning in multimodal large language models (MLLMs). This benchmark reveals that existing unlearning methods struggle with cumulative degradation, particularly due to the complexities of maintaining multimodal alignment. The proposed solution, LUMoE, shows promise in mitigating these challenges, which could enhance data privacy and model performance. Open-sourcing the dataset and code may accelerate advancements in this area.