Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

MIRIX: Multi-Agent Memory System for LLM-Based Agents

Created by
  • Haebom

Author

Yu Wang, Xi Chen

Outline

MIRIX is a modular multi-agent memory system developed to overcome the limited memory capabilities of existing AI agents. It processes not only text but also visual and multi-modal information, providing useful memory capabilities in real environments. It consists of six memory types: Core, Episodic, Semantic, Procedural, Resource Memory, and Knowledge Vault, and a multi-agent framework that dynamically manages them, enabling efficient storage, inference, and retrieval of diverse and long-term user data. It outperforms existing systems in the ScreenshotVQA and LOCOMO benchmarks, and in particular, ScreenshotVQA achieved 35% higher accuracy and 99.9% storage capacity reduction compared to the RAG criterion. It also provides package applications for users, providing real-time screen monitoring, personalized memory construction, intuitive visualization, and secure local storage.

Takeaways, Limitations

Takeaways:
Presenting a new memory system that overcomes the memory limitations of existing AI agents
Increasing applicability to real environments through multimodal information processing
Achieve state-of-the-art performance on ScreenshotVQA and LOCOMO benchmarks
Efficient memory management and reduced storage capacity
Providing user-friendly packaged applications
Limitations:
The specific Limitations is not explicitly mentioned in the paper. Future research is expected to require additional performance improvements and verification of applicability to various environments.
There is a need to verify the long-term stability and scalability of developed package applications.
Additional evaluation of generalization performance for different types of data is needed.
👍