We discussed symmetry from the viewpoint of form and structure of knowledge representation. First we proposed a hypothesis that explains why nonhuman animals do not exhibit symmetry while humans exhibit symmetry. According to the proposed hypothesis, non-human animals form procedural knowledge where an object (the comparison item) and an action are tightly coupled by knowledge compilation through repetitive training. It prevents them from recognizing the object separate from the action so that non-human animals do not recognize the object as the sample stimulus of the inverted task. Therefore the animal does not exhibit symmetry. In contrast, humans make use of verbal description of the training situation. It prompts them to separate the action and the objects, so that the object is easily recognized as the sample of the inverted task. Second we discussed the problem of similarity in reasoning. Many studies showing human symmetric reasoning presume that the truth value of a proposition is either true or false. However we do not always treat a proposition in such a dichotomous way. In such cases, similarity plays important roles. It is well known that similarity is asymmetric and that similarity is computed based on structured representation. We pointed that these two properties of similarity are crucial for understanding human symmetric as well as asymmetric reasoning.