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).