EgoMemReason: a new benchmark tests agents' week-long memory reasoning on egocentric video
mohitban47 · x · 2026-10-06
Researchers from UNC Chapel Hill and NTU introduced EgoMemReason at COLM 2026, a benchmark testing whether multimodal agents can build and use memory across a week of egocentric life-logging video (average 25.9 hours of temporal backtracking). It covers three memory levels—entity (tracking object changes across days), event (linking temporally distant events), and behavior (abstracting recurring patterns)—via 500 human-verified questions averaging 5.1 evidence pieces each. The authors find current long-video methods retrieve isolated events but struggle with long-horizon memory. Paper, code and leaderboard are public; poster #71 on Oct 6.
More from Multimodal
- Nano Banana 2.1 appears on Google Flow, but output looks like Pro — testingcatalog · 2026-10-06
- Developer turns classic fables into AI videos with rumored Opus 5.5 and HyperFrames — ankurhandos · 2026-10-06
- Mirai, backed by a16z Speedrun, turns any address into an explorable world-model simulation — Scobleizer · 2026-10-06
- Creator turns AI fight video into pixel art with Claude driving Blender, shares hard-won lessons — TomLikesRobots · 2026-10-06
- Lumen Velar Returns to Courant3D With Four Talks on AI Filmmaking and Workflow Design — LudovicCreator · 2026-10-06
- 57 attempts and a full 20x plan burned in a day matching AI art style — AC-Daniel · 2026-10-06