In Collaborative Internet of Things (CIoT) systems, where the malfunction of a single device can propagate across interconnected components and compromise overall system performance or safety, multivariate time series anomaly detection plays a critical role. However, due to the intricate spatio-temporal dependencies within CIoT time series, most existing methods struggle to model them accurately and effectively. To address the issue, we introduce MemGT, an unsupervised anomaly detection method using a memory-augmented transformer and a dynamic graph structure learning method to separately extract spatio-temporal features from the input time series, where the memory module is designed to constrain the transformer’s ability to reconstruct anomalies. Additionally, we use a graph neural network for feature fusion and generate robust spatio-temporal representations for determining the anomaly score per timestamp. Notably, we identify a certain degree of concurrent noise in the input data. To ensure that MemGT can identify anomalies accurately, we utilize a novel window-wise graph structure learning to detect the concurrent noise. Finally, we conducted empirical evaluations on three real-world datasets to demonstrate the effectiveness of MemGT. MemGT attains a 93.26% average F1 score, outperforming other latest deep models by 19.32%.