EgoMemReason: leaderboard
Long-horizon memory benchmark on week-long egocentric video (EgoMemReason, Wang et al., UNC Chapel Hill, COLM 2026): 500 multiple-choice questions over EgoLife recordings, each posed at a timestamp and answerable only by combining evidence spread across hours or days before it, covering entity, event and behavior memory in six tasks. The score is accuracy in percent over all 500 questions, with frames sampled uniformly up to the question time (1,024 by default, though some models are reported at their best frame budget); higher is better.
Metric: Accuracy (%; 500 multiple-choice questions over six memory tasks). Source: arxiv.org. Saturation forecast: Around 2030. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 39.6 | #159 |
| 2 | Gemini 3.1 Pro (Preview) | 37.4 | #91 |
| 3 | Molmo2-8B | 33.2 | #1488 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=egomemreason · How It Works · Data refreshed daily, snapshot 2026-11-12.