EgoMemReason: Long Video Memory Reasoning Benchmark

mohitban47 · x · 2026-07-15

EgoMemReason has been accepted to **COLM 2026**. It is a memory-driven reasoning benchmark designed for **long-horizon, first-person video**. Its goal is to test whether existing methods can establish and utilize **persistent visual memory**, including: - Tracking entity changes over hours to days - Connecting events that are far apart in time - Summarizing recurring behavioral patterns from sparse observations The dataset contains **500 questions**, covering three types of memory and six categories of reasoning challenges. Each question requires an average of **5.1 video clips** as evidence and involves **25.9 hours** of memory retrieval. The authors evaluated **17 MLLMs and agentic frameworks**, where the best method achieved only **39.6% accuracy**, indicating that current long-video memory systems still have significant shortcomings.

Original post →

More from Multimodal

Multimodal channel →