EGOSTREAM: A Diagnostic Benchmark for Streaming Episodic Memory in Egocentric Vision
Researchers have introduced EGOSTREAM, a new benchmark designed to evaluate the streaming episodic memory capabilities of egocentric vision models. The benchmark includes 2,250 questions across seven cognitive dimensions and introduces an Answer Validity Window (AVW) to differentiate model forgetting from real-world changes. Initial experiments using a Qwen3-VL backbone showed that current memory management mechanisms struggle to perform in real-time and achieve high accuracy, highlighting significant gaps in existing architectures. AI
IMPACT This benchmark will enable more rigorous testing and development of AI systems with improved long-term memory capabilities.