Improving Meta-learning for Low-resource Text Classification and Generation via Memory Imitation