Although the SAM2 foundational segmentation model excels in natural images, its direct adaptation to 3D medical imaging (e.g., CT/MR) remains underexplored, particularly for zero-shot generalization. We identify two critical barriers when treating medical volumes as pseudo-video sequences: (1) the non-convexity of anatomical structures leading to slice-wise mask discontinuities; (2) difficulty in effectively generalizing the dependencies between long-term and short-term memory. To address these problems, we propose a stochastic connected component propagation strategy for handling mask discontinuities during training, coupled with a dynamic memory window search mechanism during inference. Extensive experiments demonstrate the effectiveness of our method, achieving a 16% Dice score improvement over conventional fine-tuning in the unseen classes of TotalSegmentator dataset. Furthermore, our approach generalizes well across modalities (CT/MR) and lesion types, and it performs comparably to or outperforms previous methods on the ULS23 and CHAOS benchmarks.