We often encounter various anomalous behaviors of systems, such as machine failures, unexpected behaviors of intelligent agents, and irregular natural phenomena. In order to predict these anomalous behaviors, it is a useful strategy to infer the causal structure of target domains (the inference-based strategy). However, we assume another strategy, the memory-based strategy, to memorize the anomalous behaviors for the predictions. In the present study, we analyzed the features and benefits of the memory-based strategy using the spatial movement prediction task. Experiments 1 and 2 revealed that participants who were instructed to apply the memory-based strategy encoded only the anomalous instances, and not the regular instances. Additionally, the inference-based strategy was more effective for identifying the anomalous instances in a low-complexity task, whereas the memory-based strategy was more effective in a high-complexity task. Experiment 3 revealed that it was difficult to spontaneously select an appropriate strategy based on task complexity and to make benefits of the memory-based strategy for a high-complexity task even if the strategy was applied.