The introduction of RoboMME-Interference addresses the need for evaluating robot memory in realistic, long-context scenarios where systems must recall information from multiple sessions ago. This new cross-session benchmark is built upon the existing RoboMME framework to measure performance when robots face distractions from unrelated prior experiences. For each query episode, the benchmark constructs a session history consisting of relevant demonstrations followed by a controlled number of unrelated sessions provided as memory to Vision-Language-Action models. Researchers tested released memory-augmented variants of the π_0.5 model without modification to assess their robustness under these conditions. The results indicate that while perceptual memory variants improve success rates when no distractors are present, their accuracy decays steadily and strongly as unrelated sessions accumulate. These findings highlight a critical failure in current systems regarding long-context memory and interference resistance. The project page, videos, code, and data for this benchmark are available at https://robotmemorybench.com.