Working memory is a fundamental function of cognition, allowing one to hold memory “in mind” and then to guide flexible behaviors. Although the maintenance mechanisms of working memory have attracted a lot of attention, the manipulation mechanisms are less studied. To elucidate the mechanisms of working memory manipulation, we hypothesize that the prefrontal cortex (PFC) maintains information of a sample stimulus in a persistent activity of a subarea, and then, in another subarea, encodes sequence information of the sample and a test stimulus involved in a delayed match-to-category task. To test the hypothesis, we develop a network model that performs a delayed match-to-category task. Category information of a sample stimulus is maintained in a persistent activity of a positive-feedback-loop network, and the temporal information of the sample and test stimulus is represented by learning of a recurrent network. Furthermore, the task-specific decision is made by learning the connection between the area encoding temporal sequence and a decision area. Our model is further extended to more complex tasks to examine the generality of the temporal representation in PFC. The representation of temporal sequences is organized so that the PFC activities encoding the sequence information become orthogonal to each other, producing the representation of various temporal sequences of stimuli. The temporal representation enables the system to adapt to the change of task context only by relearning of the connections between the recurrent network and a decision layer. The results suggest that the PFC may generate a task-independent, temporal information of stimuli to guide various behaviors.