MemLife: Curating and Reasoning over Long-Term Egocentric Video Memories
MemLife builds entity-grounded text memories from long egocentric video and retrieves them without rewatching clips.
MemLife is a multimodal memory system for long-term egocentric video that writes entity-grounded, first-person text episodes and retrieves them with a time-indexed agentic reader, without training or query-time access to raw video. Across four long-horizon benchmarks it improves 4.6 to 12.0 percent over the strongest training-free baseline. The authors also introduce MemOpt, a reinforcement-learning method that optimizes the memory writer for faithful, informative, and retrievable memories. MemOpt adds 2.7 to 5.0 percent across video and question distributions, with gains that transfer across writer and reader backbones and memory systems.
- MemLife stores entity-grounded first-person text episodes from long video.
- A time-indexed agentic reader needs no query-time video access.
- Gains are 4.6 to 12.0 percent over the best training-free baseline.
- MemOpt reinforcement learning adds a further 2.7 to 5.0 percent.
From the original article
1 min read at huggingface.co
Long-term egocentric video enables personalized AI assistants to reason about daily life. However, as video histories grow to hundreds of hours spanning months or years, reprocessing raw clips for every query becomes computationally prohibitive.
By Guangzhi Xiong, Xinyuan Zhang, Xiao Yang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.