Time-varying community structures widely exist in various real-world networks. However, the spreading dynamics on this kind of network has not been fully studied. To this end, we systematically study the effects of time-varying community structures on social contagions. We first propose a non-Markovian social contagion model on time-varying community networks based on the activity driven network model, in which an individual adopts a behavior if and only if the accumulated behavioral information it has ever received reaches a threshold. Then, we develop a mean-field theory to describe the proposed model. From theoretical analyses and numerical simulations, we find that behavior adoption in the social contagions exhibits a hierarchical feature, i.e., the behavior first quickly spreads in one of the communities, and then outbreaks in the other. Moreover, under different behavioral information transmission rates, the final behavior adoption proportion in the whole network versus the community strength shows one of the patterns, which are a monotone increasing pattern, a non-monotonic changing pattern, and a monotone decreasing pattern. An optimal community strength maximizing the final behavior adoption can be found in a suitable range of behavioral information transmission rate. Finally, for a given average degree, increasing the number of edges generated by active nodes is more beneficial to the social contagions than increasing the average activity potential.