Few-shot emotion recognition in conversations aims to identify utterance emotions in conversations with limited labeled data. Existing methods separate context extraction and prototype computation of different emotion utterances, which limits the ability of prototypes to capture context. Additionally, the methods ignore the impact of emotion dependencies between nearby utterances on prototypes. To address these issues, we propose an Iterative Context Prototype Learning with Graph Dependency method, namely ICPGD. First, we design an iterative context-aware prototype network, which uses a dot product attention mechanism to compute prototypes and capture context. Second, a graph-dependent prototype self-update mechanism is constructed to update the prototype based on adjacent utterances. Finally, the model is evaluated on four benchmark datasets, IEMOCAP, DailyDialog, MELD, and EmoryNLP. Experiments demonstrate that ICPGD achieves superior results in few-shot scenarios, with performances of 53.94%, 64.31%, 60.53%, and 38.49% on these datasets, respectively.