The process of Traditional Chinese Medicine (TCM) syndrome differentiation and treatment follows a progressive pattern of “Li-Fa-Fang-Yao”. Addressing the complex cross-level semantic associations among these stages, and improving the recommendation accuracy and interpretability of intelligent diagnostic and treatment systems, this paper proposes a progressive treatment recommendation model, AttnSynth-KG. The model is based on the TCM clinical logic of “symptoms-syndromes-treatment methods-herbs” and constructs a multi-task joint learning framework, encompassing four sub-tasks: symptom enhancement, syndrome prediction, treatment method determination, and herb recommendation. A multi-head attention mechanism is used to model the nonlinear interactions among various diagnostic and treatment stages, while gated fusion and residual MLPs facilitate cross-level feature transfer. Additionally, the structured TCM knowledge graph (SSTH-KG) embedding is introduced to enhance semantic representation. Experimental results on the TCM-PD and LUNG datasets demonstrate that AttnSynth-KG significantly outperforms existing baseline methods in terms of Precision@k, Recall@k, and F1-score@k, exhibiting strong recommendation performance and clinical interpretability. This model provides reliable auxiliary support for TCM diagnosis and treatment.