Whole-body computed tomography (CT) is a crucial medical imaging modality that provides a comprehensive view of tissue anatomy, facilitating the detection and diagnosis of various conditions, including cancer, trauma-related injuries, vascular abnormalities, infectious diseases, and organ pathologies. However, existing deep learning methods for whole-body CT segmentation, such as nnU-Net, often suffer from limited generalization capabilities, hindering their adaptability to diverse clinical needs and real-world scenarios. To address this issue, we propose a novel interactive semantic segmentation method based on the recently introduced Segment Anything Model (SAM), which employs a foundation model approach to enable flexible instance segmentation. Our method extends SAM to overcome its limitation of losing important semantic information during the segmentation process, making it specifically tailored for whole-body CT anatomy segmentation. The proposed approach allows for both prompt-free and prompt-guided segmentation, accommodating different use cases and providing enhanced flexibility. Furthermore, we introduce a memory bank module that expands the context of the self-attention mechanism through approximate k-nearest neighbor (KNN) lookup, enabling the model to capture long-range dependencies and attend to distant relevant features, thereby improving its ability to handle the complexities of whole-body CT data. Experimental results demonstrate that our method achieves competitive performance compared to other state-of-the-art approaches while preserving rich semantic information at the pixel level. Our code is publicly available at https://github.com/13482108753/Totalct-SAM .