Zhejiang university and lenovo introduce EM²Mem: a novel event-centric memory framework for AI agents

Researchers from Zhejiang University, in partnership with Lenovo, have unveiled EM²Mem, a new memory solution designed to overcome limitations in multimodal AI agents, such as premature forgetting and inefficient data retrieval. This research was recently presented at EMNLP2026. EM²Mem fundamentally alters the conventional memory approach by shifting from a retrieve-then-align process to an align-then-retrieve paradigm. During memory construction, the system integrates diverse elements—including people, objects, actions, and time—within a single event to build multi-scale event archives. These archives are then linked across different experiences using a combination of episodic and semantic graphs to distill long-term patterns. Testing indicates that EM²Mem offers significant performance gains over existing methods, showing marked improvements in both event-level evidence recall and question-answering accuracy. Furthermore, the solution reduces the computational load and token usage per query. Building upon this core technology, the team also developed LightMem-Ego, a multimodal agent memory system intended for deployment in applications like personal and scientific assistants.