MemLife: Curating and Reasoning over Long-Term Egocentric Video Memories
MemLife turns long first-person video histories into searchable text memories so queries do not have to reread the raw footage.
The paper says long egocentric video can span hundreds of hours, making direct reprocessing too expensive. MemLife builds entity-grounded, first-person text episodes and uses a time-indexed agentic reader to retrieve them. Without training or query-time video access, it beats the strongest training-free baseline by 4.6% to 12.0% across four long-horizon benchmarks. The authors also introduce MemOpt, an RL framework for improving the memory writer, adding another 2.7% to 5.0% across tested settings. HF Daily Papers' note
The paper says long egocentric video can span hundreds of hours, making direct reprocessing too expensive. MemLife builds entity-grounded, first-person text episodes and uses a time-indexed agentic reader to retrieve them. Without training or query-time video access, it beats the strongest training-free baseline by 4.6% to 12.0% across four long-horizon benchmarks. The authors also introduce MemOpt, an RL framework for improving the memory writer, adding another 2.7% to 5.0% across tested settings. HF Daily Papers' note
score 4