Paper 2604.04503
Memory Intelligence Agent
- Published
- Apr 2026
- Research lab
- Independent
- Citations
- 4
- GitHub
- 792 stars
01 In brief
Summary
The paper introduces the Memory Intelligence Agent (MIA), a framework for deep research agents that combines a Manager-Planner-Executor architecture with dual memory systems.
The Memory Manager stores compressed historical trajectories as non-parametric memory, while the Planner, trained via alternating reinforcement learning, serves as parametric memory.
The Executor follows plans and interacts with tools.
MIA enables test-time learning, updating the Planner's parameters during inference, and includes reflection and unsupervised judgment mechanisms for self-evolution.
Experiments across eleven benchmarks show MIA improves GPT-5.4 by up to 9% on LiveVQA and 6% on HotpotQA, and with a Qwen2.5-VL-7B Executor achieves an average 31% improvement, outperforming Qwen2.5-VL-32B by 18%.
MIA also surpasses prior memory baselines by an average of 5% and demonstrates progressive self-evolution in unsupervised settings.
02 From the paper
Abstract
Deep research agents (DRAs) integrate 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 of 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. Memory Manager is a non-parametric memory system that can store compressed historical search trajectories. Planner is a parametric memory agent that can produce search plans for questions. Executor is another agent that can search and analyze information guided by the search plan. To build the MIA framework, we first adopt an alternating reinforcement learning paradigm to enhance cooperation between the Planner and the Executor. Furthermore, we enable the Planner to continuously evolve during test-time learning, with updates performed on-the-fly alongside inference without interrupting the reasoning process. Additionally, we establish a bidirectional conversion loop between parametric and non-parametric memories to achieve efficient memory evolution. Finally, we incorporate a reflection and an unsupervised judgment mechanisms to boost reasoning and self-evolution in the open world. Extensive experiments across eleven benchmarks demonstrate the superiority of MIA.