The existing paper recommender systems have proven highly beneficial in promoting knowledge acquisition among researchers. Current recommendation methods mainly rely on mining structural and semantic information of entity nodes from heterogeneous information networks (HINs). However, there is no guarantee that the recommended results align with the researcher’s current research direction. Additionally, users’ reading preferences and papers’ audience preferences will change over time, making it necessary to capture the dynamic changes of these features to obtain more accurate recommended results. This paper proposes a multi-task recommendation model called MDRec, which jointly considers timeliness and accuracy. Specifically, to keep the recommended results in line with the researcher’s current research direction, MDRec incorporates a new task that matches the content of the recommended papers with the researcher’s current research direction. To precisely capture the dynamic changes of preferences, we design a predictor called cfLSTM in MDRec, which can perceive changes in different types of features with appropriate granularity levels. Extensive experiments on real-world datasets show that MDRec outperforms current baselines.