---
title: MemProBench
canonical_url: "https://www.modelscope.cn/datasets/facebook/MemProBench"
md_url: "https://www.modelscope.cn/datasets/facebook/MemProBench.md"
repository: facebook/MemProBench
last_updated: 2026-10-24
license: other
storage_size: "5.9 GB"
downloads: 106
stars: 0
---

# MemProBench

> MemProBench - facebook 在 ModelScope 开源的数据集。MemProBench — eval800 frozen memory histories

facebook/MemProBench 是 ModelScope 魔搭社区上的数据集，存储大小 5.9 GB，采用 other 许可。

- **Repository**: facebook/MemProBench
- **License**: other
- **Storage size**: 5.9 GB
- **Downloads**: 106
- **Stars**: 0
- **Last updated**: 2026-10-24

Source: https://www.modelscope.cn/datasets/facebook/MemProBench

---

# MemProBench — eval800 frozen memory histories

This dataset hosts the **eval800 frozen memory histories** for MemProBench:
800 byte-preserved caption logs under `full_history/<UID>.jsonl`
(one JSON object per line: `cid`, `day`, `t`, `text`, `t_days_before`,
plus `clock`, `tag`, `episode_id`, `title`, `is_target`, `is_evidence`,
`evidence_kind`).

- Scope: the 800 evaluation histories only (`uids.txt` order is canonical).
- Format: each `<UID>.jsonl` is a raw caption byte stream; see
  `history_manifest.json` for per-file `bytes`, `records`, and `sha256`.
- Source videos are official Ego4D clips and are **not** redistributed here.
- Code: https://github.com/facebookresearch/Memo-Pro-Agent

## Reader

```bash
hf download facebook/MemProBench --repo-type dataset \
  --revision <commit> --local-dir data/hf
```

Downloading the private dataset requires Hugging Face read access (auth).