Our work addresses long-term motion context issues for predicting future\nframes. To predict the future precisely, it is required to capture which\nlong-term motion context (e.g., walking or running) the input motion (e.g., leg\nmovement) belongs to. The bottlenecks arising when dealing with the long-term\nmotion context are: (i) how to predict the long-term motion context naturally\nmatching input sequences with limited dynamics, (ii) how to predict the\nlong-term motion context with high-dimensionality (e.g., complex motion). To\naddress the issues, we propose novel motion context-aware video prediction. To\nsolve the bottleneck (i), we introduce a long-term motion context memory\n(LMC-Memory) with memory alignment learning. The proposed memory alignment\nlearning enables to store long-term motion contexts into the memory and to\nmatch them with sequences including limited dynamics. As a result, the\nlong-term context can be recalled from the limited input sequence. In addition,\nto resolve the bottleneck (ii), we propose memory query decomposition to store\nlocal motion context (i.e., low-dimensional dynamics) and recall the suitable\nlocal context for each local part of the input individually. It enables to\nboost the alignment effects of the memory. Experimental results show that the\nproposed method outperforms other sophisticated RNN-based methods, especially\nin long-term condition. Further, we validate the effectiveness of the proposed\nnetwork designs by conducting ablation studies and memory feature analysis. The\nsource code of this work is available.\n