To tackle a key pain point—namely, that long-term memory in multimodal agents merely stores information in a simplistic way and struggles to link different clues—Zhejiang University and the Lenovo team have jointly proposed an event-centric multimodal long-term memory solution called EM²Mem. This approach shifts the traditional "retrieve first, then align" paradigm to "align first, then retrieve." During the memory construction stage, it links clues such as people, objects, behaviors, and time within the same event, builds multi-scale event archives, and connects different experiences through episodic and semantic graphs to distill long-term patterns. Experiments show that, compared with existing methods, this solution significantly improves event-level evidence recall and question-answering accuracy while greatly reducing the time and token usage per question-answering session. Once constructed, the memory can be reused multiple times; after amortizing costs, the long-term efficiency advantage becomes evident. Building on this technology, the team further developed a multimodal agent memory system, LightMem-Ego, which can be deployed in scenarios such as personal life assistants and scientific laboratory assistants, supporting continuous learning and personalized development of multimodal agents.