Electroencephalography (EEG)-based brain-computer interfaces (BCIs) have shown significant potential in many fields, like neurorehabilitation, robotic control, and brain activity understanding. However, EEG signals are challenging to analyze due to their low signal-to-noise ratio (SNR), spatiotemporal complexity, and long-range dependencies. To tackle these challenges, in this paper we introduce GMamba, a novel architecture for EEG representation learning. It combines a mamba module to capture temporal dynamics and a dynamic graph module to model spatial complexity. The mamba module treats EEG signals as continuous variables to mine long-range dependencies. The dynamic graph module uses shared and adaptive graph topologies to model inherent electrode correlations. We evaluated GMamba on a public EEG dataset containing 11,466 EEG-image pairs. The results demonstrate that GMamba outperforms state-of-the-art methods on both EEG-Image retrieval and EEG-Text classification tasks.