Parliamentary debates are vital records of democratic processes, yet official transcripts such as the Hansard provide only a sequential account of speeches, omitting the underlying interactional structure. Although conducted in a formal setting, debates can become highly dynamic, making it difficult to determine who addresses whom and how utterances relate within the broader discussion. To address these gaps, we introduce DIAM, the first manually annotated discourse resource for the Malaysian Parliament, jointly supporting EU-level Dialogue Discourse Parsing (DDP) and multi-addressee Addressee Recognition (AR) over long, formal, unscripted spoken multi-party debate. Unlike existing DDP datasets, DIAM provides substantially longer conversations and explicit multi-addressee labels. Our benchmark experiments validate that DIAM supports both DDP and AR, while showing that existing discourse parsing methods cannot be directly applied without adaptations such as context-window strategies, and that addressee recognition is challenging for prompted LLMs, though fine-tuning a smaller model with LoRA substantially improves robustness to the large number of participants per sitting. We release DIAM and our code to facilitate future research on formal discourse parsing and addressee modeling: https://github.com/xKHUNx/DIAM