Wearable assistants require episodic memory over egocentric video, yet current vision‑language models are limited by frame budgets, rising visual‑token costs, and failures in long‑context retrieval. We investigate whether textual captions can serve as reusable episodic memory. To this end we define the Episodic Memory Video Caption QA task and introduce the CapMem benchmark, which comprises 75 videos totaling 33.7 hours, annotated with 1,000 multiple‑choice questions across 16 everyday scenarios. Experiments on long videos (~20 min) use full‑coverage CaptionQA with 30‑second and 60‑second caption windows, outperforming direct VideoQA for 10 out of 12 and 8 out of 12 models respectively. A matched‑frame control across six Qwen models retains mean accuracy gains of 3.22 and 2.55 points. Adding a caption‑guided retrieve‑and‑verify step further improves accuracy by up to 5.3 points. These findings demonstrate the effectiveness of caption memory for episodic reasoning over long egocentric video.
Review: CapMem provides a large‑scale human‑annotated dataset that validates captions as a lightweight memory medium, offering a promising direction for long‑video understanding under practical constraints.