# MemInStream [Paper](https://arxiv.org/abs/2611.15640) · [Project page](https://hkust-longgroup.github.io/MemInStream/) · [Evaluation code](https://github.com/HKUST-LongGroup/MemInStream) MemInStream tests what models remember as a video unfolds. It contains **3,799 questions** across **17 tasks**, covering episodic retrieval, state tracking and updating, multi-evidence integration, and contextual memory. Each question specifies when it is asked and which earlier video segments provide the evidence needed to answer it. ## Download This repository hosts the benchmark videos in **10 ZIP archives** (about **310 GB** total). ```bash hf download zizi32/MemInStream --repo-type dataset --include "videos_*.zip" --local-dir ./data/MemInStream ``` Extract all archives into `data/MemInStream/`, then save the [question annotations](https://raw.githubusercontent.com/HKUST-LongGroup/MemInStream/main/MemInStream/annotations.json) as `data/MemInStream/annotations.json`. The video paths in the annotations point to the extracted files under `videos/`.