The existing library knowledge management system relies on manual and static rules, performs poorly in processing complex data and interdisciplinary knowledge integration, lacks dynamic learning and intelligent recommendation capabilities, and limits knowledge utilization. In order to solve the current problem, this research has developed an intelligent library knowledge management system. The system uses DL (Deep Learning) to automatically classify documents and performs semantic correlation analysis in combination with knowledge graph. The precise classification of documents is realized through the BERT (Bidirectional Encoder Representations from Transformers) model, and the core entities of the documents are extracted using named entity recognition technology, and a comprehensive knowledge map is constructed. This article also uses GNN (Graph Neural Network) to dig deeper into potential knowledge connections and improve the effect of semantic analysis. Combined with the hidden semantic model and path similarity, the system can accurately calculate user interests and optimize recommended content. Experimental data show that the method of combining BERT and GNN performs well in retrieval, with an average accuracy rate of 94.69%, an average recall rate of 91.85%, and an average response time of only 17.80 ms, which is significantly better than other methods. This shows that the system of this article can greatly improve the quality of retrieval and recommendation, and ensure that users can obtain knowledge efficiently and accurately.