Long-form dialogue summarisation demands both a global grasp of a conversation’s themes and the retention of fine-grained, chronological details, requirements that existing approaches rarely satisfy simultaneously. We propose Hierarchical Multi-Strategy Summariser (HMSS), a hybrid pipeline that (i) builds a compact memory-tree outline of the entire transcript and (ii) incrementally refines a running summary, injecting each new chunk of dialogue while being guided by the static global outline. We evaluate HMSS and strong LLaMA-3-8B baselines on the English MeetingBank and a Vietnamese podcast corpus, reporting ROUGE, BERTScore, and a complementary LLM-as-a-judge metric (GPTScore) that rates Coherence, Faithfulness, and Usefulness. HMSS achieves the best ROUGE-L (23.5) and BERTScore (60.2) on MeetingBank, sets a new state-of-the art across all metrics on Vietnamese dataset and obtains the highest GPTScore across five pipelines. These gains confirm that coupling a lightweight global outline with guided, step-wise refinement is an effective recipe for coherent and faithful summaries of very long conversations.