OmniMemBench: Towards Scalable Evaluation of Long-Term Omni-Modal Agent Memory

NeurIPS 2026 · Evaluations and Datasets Track

Junhan Shi, Qiyi Wang, Zhenning Shi, Xiongwei Wu, Ming Ma, Baoqi Pei, Yong Jiang, Qing Li, Weigao Sun, Yiran Zhong, Steven Hoi

Code and evaluation · Datasheet · Croissant metadata · Citation

OmniMemBench evaluates long-term agent memory across image, audio, and video in multi-session dialogue. An anti-leakage construction pipeline supports perceptually grounded questions; content-based retrieval metrics diagnose memory access; content-isolated distractors scale histories from 128K to 1M tokens.

Dataset size

Context tier	Characters	QA instances
128K	103	2,719
256K	51	1,312
512K	51	1,312
1M	51	1,312
Across all tiers	103 distinct characters	6,655

The 256K, 512K, and 1M tiers reuse the same 51-character subset and target questions, with additional distractor sessions. The total counts QA instances across tiers. The release contains 256 character-tier JSON files.

The Dataset Viewer displays the benchmark index: one row per character and context tier (256 rows). The num_qas column gives the QA count for that row; file_path points to the complete dialogue and questions. crosswalk/ contains source-media mappings used for reconstruction.

Task types

Type	Task
SH	Single-Hop
MH	Multi-Hop
TR	Temporal Reasoning
CR	Causal Reasoning
ADV	Adversarial
REF	Answer Refusal
UPD	Knowledge Update
SUM	Summarization

Download and use

git clone https://github.com/jshmshijh/OmniMemBench.git
cd OmniMemBench
pip install -e ".[all]"
python download_data.py

The download script retrieves data/, crosswalk/, index.jsonl, croissant.json, and DATASHEET.md from this repository.

To inspect the index directly:

import json
from huggingface_hub import hf_hub_download

index_path = hf_hub_download(
    repo_id="jshmsjh/OmniMemBench",
    filename="index.jsonl",
    repo_type="dataset",
)
with open(index_path, encoding="utf-8") as stream:
    samples = [json.loads(line) for line in stream]

Data format

Each data/{run_id}/benchmark_{tier}.json contains:

• character_profile: synthetic persona and conversation style.
• multi_session_dialogues: dated sessions, dialogue turns, and multimodal references.
• QAs: task type, question, reference answer, evidence chain, and clue references. SUM items include fixed reference nuggets.

The index records sample_tier_id, sample_id, tier, file_path, num_sessions, and num_qas. The nested schema is illustrated in the code README.

Multimodal media

Media references resolve to paths under multimodal/. The crosswalk files map those paths to their public source datasets. From the code repository, reconstruct one character's media with:

python scripts/rebuild_multimodal/rebuild.py --runs 00001

Omit --runs to reconstruct all referenced media. See the media reconstruction guide for the source list and options.

Evaluation

The paper uses GLM-5.2 at temperature 0 for answer scoring and retrieval judgments.

Metric	Definition
Non-SUM QA score	0 (incorrect), 0.5 (partially correct), or 1 (correct)
SUM score	Fraction of the question's fixed reference nuggets covered by the answer
Overall	Mean of one score per valid question, including SUM
Clue Coverage (CC)	Fraction of evidence clues semantically covered by retrieved entries, averaged over valid questions
Entry Relevance (ER)	Fraction of retrieved entries relevant to the evidence clues, averaged over valid questions

Missing responses and invalid judge outputs are excluded from the corresponding average. CC/ER use questions with nonempty clue and retrieved-entry sets. Context-scaling comparisons use the same 51 characters at every tier, including 128K; paired comparisons use their common valid questions.

The code README provides model configuration, inference, and GLM scoring commands.

Documentation and license

The datasheet describes construction, intended use, source media, and maintenance. Croissant metadata provides the sample-index schema and Responsible AI fields.

The benchmark is released under CC BY-NC 4.0. Source media retain their original dataset licenses. Evaluation code is released under Apache License 2.0.

Citation

@inproceedings{shi2026omnimembench,
  title     = {OmniMemBench: Towards Scalable Evaluation of Long-Term Omni-Modal Agent Memory},
  author    = {Shi, Junhan and Wang, Qiyi and Shi, Zhenning and Wu, Xiongwei and Ma, Ming and Pei, Baoqi and Jiang, Yong and Li, Qing and Sun, Weigao and Zhong, Yiran and Hoi, Steven},
  booktitle = {NeurIPS 2026 Evaluations and Datasets Track},
  year      = {2026},
  url       = {https://github.com/jshmshijh/OmniMemBench}
}

Contact: shijh23@mails.tsinghua.edu.cn.