Memory Intelligence Agent (MIA)

Paper | GitHub

MIA (Memory In Intelligence Agent) is a memory framework designed for deep research agents (DRAs). It transforms agents from "passive record-keepers" into "active strategists" using a Manager-Planner-Executor architecture. This repository contains the datasets and data artifacts used to train and evaluate the MIA framework.

Dataset Description

The dataset includes the following components:

• Train: Data used for the two-stage Reinforcement Learning (RL) training of the Executor and Planner.
• Test: Evaluation benchmarks across multiple datasets (e.g., LiveVQA, HotpotQA) to measure research and reasoning performance.
• TTRL: Specifically curated data for Continual Test-Time Learning, allowing the Planner to adapt its strategies during inference.
• Image Search Cache: Supporting cache for image-to-image search tasks.

Framework Architecture

MIA replaces standard memory systems with a sophisticated three-part architecture:

• The Manager: A non-parametric memory system that stores compressed historical search trajectories.
• The Planner: A parametric memory agent that produces search plans for questions.
• The Executor: An agent that searches and analyzes information guided by the search plan.

For detailed instructions on environment setup and how to use these files for training or inference, please visit the official GitHub repository.