Due to the limitation of training datasets and global motion understanding to optical flow estimation, current methods only focus on local clues and ignore the motion continuity of consecutive frames, resulting in an inconsistent motion problem. After theoretical analysis, we find the multi-frame methods need to pay more attention to the continuity between cost volumes than two-frame methods. Thus, our method is based on a multi-frame framework and introduces extra object coordinates in each frame by the segmentation model to revise the matching pairs in cost volume. Specifically, we introduce a Cost Volume Adaptation Module, including a Bbox Spatial Queries to store coordinates information and a Correlation Query Queue to query the object position of different frames. On the Sintel and KITTI test benchmark, our proposed MemoFlow achieves 1.00 and 1.69 average endpoint error (AEPE) on the clean and final passes and an F1-all error of 4.43%, ranking 1st among all three-frame methods and two-frame methods.