DolphinBench

DolphinBench measures whether an agent can use long-term memory to take the right actions on a user's behalf. It contains three synthetic users, approximately 500,000 tokens of history per user, and 600 tool-using tests across their work and personal lives.

The benchmark grades the actions an agent takes through simulated apps, not just what it says it remembers.

Load the dataset

from datasets import load_dataset

messages = load_dataset("mem0ai/dolphinbench")
facts = load_dataset("mem0ai/dolphinbench", "facts")
tests = load_dataset("mem0ai/dolphinbench", "tests")

Contents

Configuration	Rows	Contents
messages	13,539	Dated conversation history for each synthetic user.
facts	1,058	Information established by the histories and the messages that establish it.
tests	600	Task requests, required facts, expected tools, app state, and grading checks.

The grading, app_state, and shared_app_state columns contain JSON. The complete benchmark runner, simulated apps, and release validation are maintained in the DolphinBench GitHub repository.

Links

• Website
• Paper
• GitHub

Intended use

DolphinBench is evaluation data for comparing agent harnesses, models, and memory systems. Do not train on its histories, facts, tests, or grading criteria if you intend to report DolphinBench results.

All people, organizations, conversations, and app states in the dataset are synthetic.

Citation

@article{rathi2026dolphinbench,
  title={DolphinBench: Mapping the Pareto Frontier of Agent Memory},
  author={Rathi, Soumil and Yadav, Deshraj and Singh, Taranjeet},
  journal={arXiv preprint arXiv:2609.24971},
  year={2026}
}