Diffusion models have shown remarkable expressive and generative capabilities in the computer vision field. Recent research has attempted to apply diffusion models to recommender systems and achieved promising results. However, these studies are limited to modeling single transition patterns and lack a generalist agent capable of representing heterogeneous user behavior patterns as a complex joint distribution. Moreover, they rely on high-quality condition signals that are susceptible to noise. In this paper, we propose a novel Multi-pattern Joint Denoising Diffusion Model for Sequential Recommendation (MDD4SR), which captures the joint distribution of two transition patterns, one depicting item sequential transition and the other measuring item semantic relevance. Furthermore, a conditional denoising network based on a sequential-semantic attention mechanism is designed to adaptively adjust the weights of the condition signals extracted from different transition patterns during the denoising process. Additionally, to alleviate the noise between sequences, a residual-like connection is leveraged to integrate the condition signals. We evaluate the effectiveness of MDD4SR through extensive experiments and comparisons with existing SOTA methods. Codes are available at https://github.com/llhhcc/MDD4SRec .