Image-level weakly supervised instance segmentation enables to reduce the dependence of deep learning models on pixel-level annotations and has thus garnered widespread attention in recent years. However, the absence of mask supervision often leads models to focus only on the most discriminative regions, resulting in instance omission. In this paper, we propose a Prototype Memory Bank (ProMB) with a dynamic update strategy to collect diverse instance representations as prototypes. Through contrastive learning between instance features and these prototypes, ProMB facilitates semantic mining of omitted instances. Unlike existing methods based on single-image spatial relationships, our approach comprehensively considers the instance characteristics of all images to learn cross-image context knowledge, thereby serving as an explicit feature aggregation. Furthermore, ProMB allows for semantic propagation, which helps reduce the noise caused by erroneous instance mining. Extensive experiments demonstrate that the proposed ProMB achieves state-of-the-art performance on both PASCAL VOC 2012 and MS COCO datasets and exhibits clear interpretability.