The goal of semi-supervised video object segmentation is to continuously track and segment out the target objects in subsequent video frames given the first frame video object labeling. Many networks based on space-time memory have achieved significant results by utilizing the background and object features of past stored video frames to guide the segmentation of objects in subsequent frames. However, most of these networks simply store video frames in chronological order, resulting in an ever-expanding memory bank that increases the matching segmentation time, while the accumulating perturbations and noises in the stored features also reduce the segmentation accuracy. Inspired by the memory curve, a split-flow memory network was proposed to solve the above problems. The network divided the memory bank into three levels: primary, middle and senior, and introduced a split-flow memory module to set standards and control the feature memory flow. By adopting different strategies, the network improved the accuracy rate while shortening the segmentation time. With consistent training data, the network achieved optimal results on both single-object dataset DAVIS16 and multi-object dataset DAVIS2017 compared with the same type of networks.