Whole-brain network dynamics are yet to be fully understood, particularly in the context of auditory memory and predictive coding. BROAD-NESS (BROADband brain Network Estimation via Source Separation), a novel pipeline for extracting broadband whole-brain networks from source-reconstructed MEG data is introduced. During auditory sequence recognition, BROAD-NESS identified two orthogonal networks centered on auditory cortices. The first, also encompassing medial cingulate, is primarily involved in processing sounds and shows consistent but less marked differences between experimental conditions. The second, involving hippocampus, anterior cingulate, insula, and inferior temporal regions, exhibits strong condition-dependent dynamics, reflecting engagement in confirmed predictions and prediction errors. The networks differ in temporal dynamics, spatial gradients, and behavioral relevance. Phase space and recurrence quantification analysis (RQA) reveal that more recurrent and stable dynamics are linked to higher accuracy and faster responses across sequence types. BROAD-NESS also enables direct PCA versus ICA comparison, showing PCA-based networks to be more robust and interpretable. Conceptually, this work reveals a dual-stream embedded organization of auditory memory networks that mirrors, yet functionally diverges from, visual pathways. Methodologically, it introduces BROAD-NESS, a powerful and interpretable pipeline for characterizing the spatiotemporal architecture of brain networks in neurophysiology.