docs/DATASET_DETAILS.md

# Dataset details

## Schema

- video_uid: Ego4D video UUID.
- conversation: ordered list of turns, each with role, content, time and task.
- role: user or assistant.
- content: natural-language query or response.
- time: seconds from the start of the video.
- task: modality/reasoning/task identifier, for example object_spatial_appear.

A video can have multiple dialogue examples. Preserve example order and all test examples when evaluating. Do not deduplicate test rows by video_uid.

## Video access and preprocessing

This repository contains annotations and video UID lists; it does not redistribute videos or precomputed visual features. Obtain Ego4D access using the [official procedure](https://ego4d-data.org/docs/start-here/). The UID files under uids/ contain the 4,426 training IDs, 40 test IDs and their 4,466-ID union. Use the Memento code repository to download the video_540ss subset, sample at 2 FPS, pad frames to 384×384 and encode SigLIP features.

## Construction

Memento-54K is constructed from Ego4D daily-life videos selected for long-duration interaction. The paper describes nine task types across spatial and temporal reasoning and the action, object and text modalities. Action queries use timestamped narrations with GPT-4o-assisted generation; object and text tasks use timestamp-level detections with rule-based QA generation. Dialogue samples combine one to five user queries with timestamped responses. See the paper for the full construction method.

## Data quality and limitations

Generated queries and detector-derived labels may contain errors. First-person videos reflect the activities, environments and participants represented in Ego4D and are not a representative sample of every application setting.

## License and attribution

Memento annotation text and the authored dataset organization are licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/), permitting research and commercial reuse subject to its terms and attribution. This grant covers rights held by the Memento annotation authors. Ego4D source videos remain under their data-provider agreements and must be obtained independently; this repository grants no rights to redistribute those videos or to third-party content, likenesses or trademarks.

Please attribute Memento: Toward an All-Day Proactive Assistant for Ultra-Long Streaming Video and the Ego4D dataset when using this resource.

## Citation

```bibtex
@inproceedings{memento2026iclr,
  title = {Memento: Toward an All-Day Proactive Assistant for Ultra-Long Streaming Video},
  booktitle = {ICLR},
  year = {2026},
  url = {https://openreview.net/forum?id=FtdbdoGbk3}
}
```