The authors built a two-stage system around frozen multimodal language models.
Compact reusable memory improves question answering over hour-long videos
Video-EMem encodes each video once as time-linked text and sparse images, then reuses that memory to answer multiple questions efficiently.
Big Tech
Kerui Chen · Jinglu Wang · Yun Wang · Xinjiang Wang · Yan Lu
Zhejiang University · Microsoft Research Asia
Research Digest··3 min read
Thread:Agent Memory Architecture
Chen et al.
§