EventMemAgent: Hierarchical Event-Centric Memory for Online Video Understanding with Adaptive Tool Use

Siwei Wen, Zhangcheng Wang, Xingjian Zhang, Lei Huang, Wenjun Wu,

📰 News

• [2025.2.20]: 🔥 We have released EventMemAgent: Hierarchical Event-Centric Memory for Online Video Understanding with Adaptive Tool Use. Check out the paper. Our code will come soon.

EventMemAgent Overview

EventMemAgent is an active online video agent framework based on hierarchical memory. It employs a dual-layer strategy: short-term memory detects event boundaries and uses event-granular reservoir sampling; long-term memory archives past observations on an event-by-event basis. The framework integrates a multi-granular perception toolkit (search_memory, OCR, object detection) and supports Agentic RL for end-to-end reasoning and tool-use strategies.

• Short-term memory: Event boundary detection + reservoir sampling within a fixed-length buffer
• Long-term memory: Structured event-by-event archival with captions and change descriptions
• Tools: search_memory, ocr, detect_objects for active evidence capture

Citation

@article{wen2026eventmemagent,
  title={EventMemAgent: Hierarchical Event-Centric Memory for Online Video Understanding with Adaptive Tool Use},
  author={Wen, Siwei and Wang, Zhangcheng and Zhang, Xingjian and Huang, Lei and Wu, Wenjun},
  journal={arXiv preprint arXiv:2602.15329},
  year={2026}
}