The technique of behavior cloning is to equip a robot with the capability of learning control skills through observation, which can naturally perform human–robot interaction. Despite many related studies in the context of humanoid robot behavior cloning, the problems of the unnecessary recording of similar actions and more efficient storage forms than recording actions by joint angles or motor counts are still worth discussing. To reduce the storage burden on robots, we implemented an end-to-end humanoid robot behavior cloning system, which consists of three modules, namely action emulation, action memorization, and action replay. With the help of traditional machine learning methods, the system can avoid recording similar actions while storing actions in a more efficient form. A jitter problem in the action replay is also handled. In our system, an action is defined as a sequence of many pose frames. We propose a revised key-pose detection algorithm to keep minimal poses of each action to minimize storage consumption. Subsequently, a clustering algorithm for key poses is implemented to save each action in the form of identifiers series. Finally, a similarity equation is proposed to avoid the unnecessary storage of similar actions, in which the similarity evaluation of actions is defined as an LCS problem. Experiments on different actions have shown that our system greatly reduces the storage burden of the robot while ensuring that the errors are within acceptable limits. The average error of the revised key-pose detection algorithm is reduced by 69% compared to the original and 26% compared to another advanced algorithm. The storage consumption of actions is reduced by 97% eventually. Experimental results demonstrate that the system can efficiently memorize actions to complete behavioral cloning.