MOSAIC

An omni-modal benchmark for long-term memory. Remember what was said, shown, and conveyed, then find where the memory failed.

Paper: arXiv:2610.29249 · Code: HKUST-KnowComp/MOSAIC · Project and leaderboard: hkust-knowcomp.github.io/MOSAIC

MOSAIC is 923 questions over histories of text, image, video, and audio. A single end-to-end score cannot say where the evidence was lost, so the benchmark holds the question fixed and separates four stages: write, retrieve, deliver, and read.

Questions	923
Answerable / refusal	784 / 139
Sessions per history	11–24
Modalities	Text, image, video, audio

On the end-to-end leaderboard, ReAct with native multimodal memory and MiMo-V2.5 scores 0.732 on the 784 answerable questions. With the reader fixed at Gemini 3.5, A-MEM on caption text leads at 0.684. Correct refusal and false refusal are reported beside the end-to-end score, because refusing everything is not a better memory.

Layout

Directory	Contents
dataset/stage2_human_filtered/	923 questions
dataset/media/	images, audio, and video
dataset/textualized_caption/	caption-text corpus
dataset/textualized_omni/	omni-text corpus
indexes/	RAG, A-MEM, and MemoryOS indexes
records/protocol_v1/	paper evaluation records

Question files store media_path values whose clean_output/media/ suffix matches dataset/media/. The evaluation code maps those paths when MOSAIC_DATA_ROOT is set to this folder.

Citation

@misc{tsang2026mosaic,
  title = {{MOSAIC}: An Omni-Modal Benchmark for Long-Term Memory and Cross-Modal Reasoning in Conversational Agents},
  author = {Tsang, Hong Ting and Zhang, Boyi and Huang, Haoyu and Bai, Jiaxin and Li, Yufei and Gao, Yisen and Xie, Zhongwei and Luo, Hongyu and Pan, Jeff Z. and Song, Yangqiu},
  year = {2026},
  eprint = {2610.29249},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  doi = {10.48550/arXiv.2610.29249},
  url = {https://arxiv.org/abs/2610.29249}
}