MIA is a dataset jointly released in April 2026 by East China Normal University, Shanghai Innovation Institute, and Harbin Institute of Technology for training and evaluating intelligent agents with long-term memory and task execution capabilities. Related research papers include... Memory Intelligence AgentThe aim is to enhance the long-term memory utilization and multi-step decision-making capabilities of intelligent agents. This dataset contains approximately 21,000 inference trajectory data, covering the entire process of problem solving, planning, searching, and execution, and is suitable for agent inference and reinforcement learning research.

Data Structure

This dataset contains the following components:

• Training: Data for two-stage reinforcement learning (RL) training of the executor and planner.
• Testing: Evaluate benchmarks across multiple datasets (e.g., LiveVQA, HotpotQA) to measure research and inference performance.
• TTRL: Data specifically selected for continuous learning during testing, enabling the planner to adjust its strategy during inference.
• Image search caching: Supports caching for image-to-image search tasks.