In recent years, with the rapid development of deep learning, computer vision has achieved unprecedented success. The continuous progress in domain adaptation techniques has provided powerful technical support for Synthetic Aperture Radar (SAR) target detection. This paper investigates a source-free domain adaptation setting for SAR target detection. First, a mean-teacher framework is employed to effectively transfer knowledge from the source-trained model to the target domain. To enhance the feature representation of target domain data, a novel training strategy is proposed, leveraging the idea of graph-guided contrastive learning. To better capture the relationships among proposed instances, this paper presents an Instance Relation Graph Network based on Graph Convolutional Networks (GCNs), which models the inter-instance dependencies effectively. By learning the relationships between instances, positive and negative proposal pairs can be obtained to guide the contrastive representation learning process. Finally, the proposed method is compared with existing domain adaptation approaches. Experimental results demonstrate that the proposed method can effectively adapt a source-trained detector to the target domain, validating its effectiveness and superiority. Moreover, the results highlight the potential of graph-guided contrastive learning in source-free domain adaptation for SAR target detection.