In an era of information overload, precise book recommendations are of significance for improving users’ reading experiences and discovering high-quality content. However, existing recommendation systems suffer from problems including the weak ability to capture dynamic interests, insufficient deep semantic understanding of books, and poor interpretability of recommendations. To this end, this study proposes a time chain-of-thought book recommendation model enhanced by large language models (Time-CoT-LLM). This model employs a time chain-of-thought module to integrate users’ recent behavioral dynamics, forming semantic inference chains and enhancing recommendation interpretability. A user encoder combining time-invasive sparse attention (TISA) with mixture-of-experts (MoE) is adopted to precisely capture the temporal variation and diversity of user interests. An LLM-enhanced book encoder is utilized to extract deep semantic features from books. Experiments conducted on the BookCrossing and Goodreads datasets demonstrate that Time-CoT-LLM outperforms traditional collaborative filtering, BERT-based sequence recommendation, and pure LLMs by 3.2% to 4.1% in metrics such as AUC (0.772) and NDCG@5 (0.396). Furthermore, 85% of the users perceive the recommendation reasons as clearer. This study provides an effective solution for dealing with challenges in dynamic interest tracking, semantic understanding, and interpretability in book recommendations, and plays a significant role in advancing the practical application of intelligent recommendation technologies in cultural domains.