Memory Intelligence Agent
Summary: arXiv:2604.04503v1 Announce Type: new
Abstract
Deep research agents (DRAs) integrate large language model (LLM) reasoning with external tools. Memory systems enable DRAs to leverage historical experiences, which are essential for efficient reasoning and autonomous evolution. Existing methods rely on retrieving similar trajectories from memory to aid reasoning, while suffering from key limitations such as ineffective memory evolution and increasing storage and retrieval costs.
To address these problems, we propose a novel Memory Intelligence Agent (MIA) framework, consisting of a Manager-Planner-Executor architecture. The Memory Manager is a non-parametric memory system that can store compressed historical search trajectories. The Planner is a parametric memory agent that can produce search plans for questions. The Executor is another agent that can search and analyze information guided by the search plan.
Key Features of the MIA Framework
The development of the MIA framework involves several innovative features:
- Cooperative Learning: The framework adopts an alternating reinforcement learning paradigm to enhance cooperation between the Planner and the Executor, ensuring efficient task execution.
- Continuous Evolution: The Planner is enabled to evolve continuously during test-time learning, allowing updates to be performed on-the-fly alongside inference without interrupting the reasoning process.
- Efficient Memory Evolution: A bidirectional conversion loop between parametric and non-parametric memories is established to achieve efficient memory evolution, facilitating better access to historical data.
- Reflection and Judgment Mechanisms: The incorporation of reflection and unsupervised judgment mechanisms boosts reasoning and self-evolution capabilities in an open-world context.
Experimental Validation
Extensive experiments conducted across eleven benchmarks demonstrate the superiority of the MIA framework compared to existing methods. These experiments validate the effectiveness of the proposed architecture in enhancing memory utilization and improving reasoning outcomes. The results indicate that MIA not only addresses the limitations of previous approaches but also sets a new standard for memory integration in AI systems.
Conclusion
The Memory Intelligence Agent (MIA) framework represents a significant advancement in the field of artificial intelligence, particularly in the integration of reasoning and memory systems. By addressing the critical issues of memory evolution and retrieval costs, MIA stands poised to enhance the capabilities of deep research agents in various applications. As AI continues to evolve, frameworks like MIA will play a crucial role in enabling more intelligent and autonomous systems.