README.md

---
license: cc-by-nc-4.0
task_categories:
  - question-answering
  - visual-question-answering
language:
  - en
  - zh
tags:
  - personal-memory
  - multimodal
  - long-term-memory
  - retrieval-augmented-generation
  - benchmark
pretty_name: ATM-Bench
size_categories:
  - 1K<n<10K
---

# ATM-Bench: Long-Term Personalized Referential Memory QA

[![arXiv](https://img.shields.io/badge/arXiv-2603.01990-b31b1b.svg)](https://arxiv.org/abs/2603.01990)
[![GitHub](https://img.shields.io/badge/GitHub-ATM--Bench-blue.svg)](https://github.com/JingbiaoMei/ATM-Bench)
[![Project Page](https://img.shields.io/badge/Project-Page-green.svg)](https://atmbench.github.io/)

**ATM-Bench** is the first benchmark for **multimodal, multi-source personalized referential memory QA** over long time horizons (~4 years) with **evidence-grounded** retrieval and answering.

> **Paper:** [According to Me: Long-Term Personalized Referential Memory QA](https://arxiv.org/abs/2603.01990)

![ATM-Bench Overview](ATM-Bench-Demo.png)

## Overview

Existing long-term memory benchmarks focus primarily on dialogue history, failing to capture realistic personalized references grounded in lived experience. ATM-Bench addresses this gap with:

- **Multimodal and multi-source data:** 3,759 images, 533 videos, and 6,742 emails spanning ~4 years
- **Referential queries:** Resolving personalized references (e.g., "Show me the moments where Grace was trying to be sneaky...")
- **Evidence-grounded:** Human-annotated QA pairs with ground-truth memory evidence
- **Multi-evidence reasoning:** Queries requiring evidence from multiple sources
- **NIAH evaluation:** Needle-In-A-Haystack protocol isolating reasoning from retrieval

## Dataset Structure

```
data/
├── atm-bench/
│   ├── atm-bench.json              # Full benchmark (1,013 questions)
│   ├── atm-bench-hard.json         # Challenging evaluation split (31 questions)
│   └── niah/                       # Needle-In-A-Haystack variants
│       ├── atm-bench-hard-niah25.json
│       ├── atm-bench-hard-niah50.json
│       ├── atm-bench-hard-niah100.json
│       └── atm-bench-hard-niah200.json
└── raw_memory/
    ├── email/
    │   └── emails.json             # 6,742 emails with summaries
    ├── image/                      # 3,759 personal photos (.jpg)
    ├── video/                      # 533 personal videos (.mp4)
    └── geocoding_cache/            # Pre-computed reverse geocoding
        ├── image/                  # 3,759 location cache files
        └── video/                  # 533 location cache files
```

## QA Data Format

Each question in `atm-bench.json` and `atm-bench-hard.json`:

```json
{
  "id": "uuid",
  "question": "How much did I pay for my hotel during my recent trip to Portugal?",
  "answer": "€842.97",
  "notes": "",
  "evidence_ids": ["20250310_202208", "email202502110008", "email202502200013"],
  "qtype": "number"
}
```

**Question types:**
| Type | ATM-Bench | ATM-Bench-Hard |
|------|-----------|----------------|
| `open_end` | 514 | 13 |
| `number` | 360 | 6 |
| `list_recall` | 139 | 12 |
| **Total** | **1,013** | **31** |

NIAH variants add a `niah_evidence_ids` field containing the evidence pool (ground-truth + distractors).

## Raw Memory

- **Images:** Personal photos with EXIF GPS and timestamps preserved.
- **Videos:** Personal videos re-encoded. GPS and timestamps preserved in MP4 metadata.
- **Emails:** Summarized emails with `id`, `timestamp`, `short_summary`, and `detail` fields. Institutional email addresses and specific identifying details have been redacted.
- **Geocoding cache:** Pre-computed reverse geocoding results for GPS coordinates, avoiding repeated API calls during memory processing.

## Memory Evidence IDs

Evidence IDs follow these conventions:
- **Image/Video:** `YYYYMMDD_HHMMSS` (timestamp-based filename without extension)
- **Email:** `emailYYYYMMDDNNNN` (date + sequence number)

## Usage

### Download

```python
from datasets import load_dataset

# Load QA data only
dataset = load_dataset("Jingbiao/ATM-Bench", data_files="data/atm-bench/*.json")
```

Or clone the full dataset (includes images/videos, ~3.1 GB):

```bash
# Install Git LFS first
git lfs install
git clone https://huggingface.co/datasets/Jingbiao/ATM-Bench
```

### With the evaluation codebase

```bash
# Clone the codebase
git clone https://github.com/JingbiaoMei/ATM-Bench.git
cd ATM-Bench

# Place data under data/
# The repo expects: data/atm-bench/, data/raw_memory/

# See the GitHub repo for full evaluation instructions
```

## Privacy and Ethics

This dataset is derived from real personal data with the data owner's consent. The following PII mitigations have been applied:

- **Images:** EXIF device identifiers (Make, Model, Software, ImageUniqueID) stripped; GPS and timestamps preserved as they are features of the benchmark.
- **Videos:** Removing original device metadata.
- **Emails:** Private Email addresses replaced with `[email_address]`;  private phone numbers replaced with `[phone_number]`; private website links replaced with `[link]`.
- **Sensitive visual content:** Images containing sensitive information have been manually reviewed and redacted with black boxes.
- See the detailed ethical considerations in the paper for more information.

## Citation

```bibtex
@article{mei2026atm,
  title={According to Me: Long-Term Personalized Referential Memory QA},
  author={Mei, Jingbiao and Chen, Jinghong and Yang, Guangyu and Hou, Xinyu and Li, Margaret and Byrne, Bill},
  journal={arXiv preprint arXiv:2603.01990},
  year={2026},
  url={https://arxiv.org/abs/2603.01990},
  doi={10.48550/arXiv.2603.01990}
}
```

## License

This dataset is released under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). The accompanying code is released under the [MIT License](https://github.com/JingbiaoMei/ATM-Bench/blob/main/LICENSE).