| --- |
| license: mit |
| library_name: transformers |
| pipeline_tag: text-generation |
| datasets: |
| - LightningCreeper/MIA |
| base_model: |
| - Qwen/Qwen2.5-VL-7B-Instruct |
| - Qwen/Qwen3-8B |
| --- |
| |
| # Memory Intelligence Agent (MIA) |
|
|
| Memory Intelligence Agent (MIA) is a memory framework designed for deep research agents (DRAs). It transforms agents from "passive record-keepers" into "active strategists" using a sophisticated **Manager-Planner-Executor** architecture. |
|
|
| - **Paper:** [Memory Intelligence Agent](https://huggingface.co/papers/2604.04503) |
| - **Repository:** [https://github.com/ECNU-SII/MIA](https://github.com/ECNU-SII/MIA) |
|
|
| ## Overview |
|
|
| MIA replaces traditional "memory dumps" with a specialized architecture to enable efficient reasoning and autonomous evolution: |
|
|
| - **The Manager**: A non-parametric memory system that stores and optimizes compressed historical search trajectories to eliminate bloat. |
| - **The Planner**: A parametric memory agent that produces search plans and evolves its strategy via Continual Test-Time Learning during inference. |
| - **The Executor**: A precision instrument that searches and analyzes information guided by the search plan. |
|
|
| MIA employs an alternating reinforcement learning paradigm to enhance cooperation between components and establishes a bidirectional conversion loop between parametric and non-parametric memories. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @article{qiao2026mia, |
| title={Memory Intelligence Agent}, |
| author={Jingyang Qiao and Weicheng Meng and Yu Cheng and Zhihang Lin and Zhizhong Zhang and Xin Tan and Jingyu Gong and Kun Shao and Yuan Xie}, |
| journal={arXiv preprint arXiv:2604.04503}, |
| year={2026} |
| } |
| ``` |
