In recent years, remarkable capabilities in multi-turn dialogue within a single session have been demonstrated by large language models (LLMs). However, maintaining cross-session coherence while accurately memorizing user-specific personalized preferences across extended multi-session interactions remains a core challenge. Among existing approaches, cross-session coherence can be preserved by summarization methods, yet fine-grained user preferences are inevitably lost due to lossy compression; original historical details are retained by retrieval-augmented generation (RAG) methods, but recall accuracy is degraded by the low signal-to-noise ratio and the absence of global context in raw conversation logs. To address these limitations, DuoMem, a dual-stream memory framework consisting of an offline memory construction phase and an online inference phase, is proposed. In the offline phase, a macro summary stream was built via recursive summarization to maintain cross-session coherence, and a user profile stream was constructed through structured extraction and hybrid retrieval to support fine-grained personalized memory. Within the user profile stream, profile entries were classified and updated by a Memory State Tracking (MST) mechanism based on natural language inference (NLI), and high-frequency stable user profiles were consolidated into the core identity context by a confidence-driven promotion mechanism. In the online phase, the number of profile entries to be injected was dynamically determined by an adaptive fusion gate according to retrieval confidence, and the macro summary, profile entries, and core identity context were assembled into the final prompt by a tiered injection mechanism for efficient personalized response generation. Experiments were conducted on the MSC（Multi-Session Chat） and LoCoMo （Long Conversation Memory）long-term dialogue benchmarks. Recall accuracies of 0.895 and 0.891 were achieved by DuoMem, exceeding the best baseline by 11.4 and 13.1 percentage points (pp), respectively; personalization accuracies of 0.829 and 0.801 were attained, surpassing the best baseline by 13.4 and 13.0 pp, while the routing latency was reduced to 6.5 ms, approximately 44 times faster than conventional LLM-based routing. These results demonstrate that cross-session coherence, fine-grained personalization, and inference efficiency can be jointly achieved through the proposed dual-stream decoupling, providing a viable solution for long-term personalized dialogue with large language models.