Summary

MIA (Memory In Intelligence Agent) addresses critical memory and reasoning limitations in deep research agents, transforming them into active strategists. It targets researchers and developers building advanced AI agents that require robust memory management and dynamic strategy evolution, offering significant performance and efficiency gains.

How It Works

MIA employs a Manager-Planner-Executor architecture. The Manager optimizes memory storage, the Planner uses Continual Test-Time Learning (TTRL) to dynamically adapt research strategies, and the Executor precisely follows research blueprints. This framework integrates parametric and non-parametric memories via Reinforcement Learning (RL) for autonomous agent evolution in open-world scenarios.

Quick Start & Requirements

Requires Python 3.10+ (Python 3.10.0). Installation involves Conda environment setup (verl) and running install.sh. Dependencies include Flash-attention (CUDA 12). Supports online/offline text and image search, often needing API keys. Deployment for training/inference is complex, requiring multi-node configurations with vLLM and substantial GPU resources. Links: Paper (2604.04503), Models (Hugging Face).

Highlighted Details

• Significantly boosts SOTA LLM performance on benchmarks like LiveVQA and HotpotQA.
• Enables smaller models (e.g., Qwen-2.5-VL-7B) to outperform larger closed-source models in non-tool-calling settings.
• Establishes a new benchmark in memory-augmented architectures, outperforming contemporary SOTA agent memory frameworks.
• Offers OpenClaw skills with integrated memory and trust-worthy judgment.

Maintenance & Community

Led by ECNU and SII researchers. No explicit community channels (Discord, Slack) or social media links are provided in the README.

Licensing & Compatibility

Released under the MIT License, permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

The setup and deployment are highly complex, requiring multi-node infrastructure and significant GPU resources, posing a barrier to entry. Planned "High-Efficiency" and "Trust-worthy" versions suggest current iterations may lack these advanced features. The extensive service deployment instructions indicate a system designed for research environments rather than immediate plug-and-play use.