<p>The rapid expansion of data requires researchers to spend excessive time and effort in locating relevant references while writing a manuscript. The concept of context-aware citation recommendation, often referred to as the local recommender model, seeks to address this challenge by examining both the citation context and candidate cited papers. Subsequently, the model generates a list of research papers suggested in order of relevance, starting from the most pertinent and descending to the least pertinent. These recommendation systems leverage metadata and the contextual information associated with the citation tags in the input research paper to suggest pertinent references. However, most existing methods failed to provide more context-preserving embeddings of the papers and separately encode citation context and candidate-cited papers into vectors and then match the two vectors. It is essential to adaptively acquire knowledge about the citation context and candidate papers and represent their interaction within the model. To overcome these issues, we propose a Recurrent Reasoning Memory Network over SPECTER for a context-aware recommender model. Compared with other methods, our approach leverages the ability of content modeling from SPECTER to explore more semantics representations of the citation context and papers. Moreover, the recurrent reasoning memory network, which performs a series of attention-based reasoning steps, can dynamically learn citation context and paper representations and model their interaction in each step. As a result, our approach can better model paper representations and improve the model’s accuracy. We have experimentally evaluated our model against state-of-the-art approaches over two real-world datasets. The findings depict a significant improvement in terms of nDCG, MAP, MRR, F1-score, and recall over the top@<i>k</i> recommended papers.</p>