As a typical task of multivariate time series, traffic forecasting has always held significant application value across various fields. Most existing models are constrained by fixed input lengths, making it difficult to avoid the ambiguity in spatio-temporal features. We revisited the synergistic effects of Transformer-based models and TCN-MLP models in spatio-temporal data forecasting and proposed the Long-Range Historical Memory-Enhanced transformer (LHMformer). The network consists of a transformer module for processing short-term inputs and a historical feature extraction module for processing long-term inputs. Experiments on six real-world traffic datasets show that the proposed method achieves state-of-the-art results. Follow-up experiments demonstrate that the historical feature extraction module is a key component in solving traffic flow forecasting problems.