As artificial intelligence evolves from handling single tasks to interacting in complex scenarios, “memory capabilities” have become a key indicator of its intelligence. Traditional AI systems often struggle with the disorganized storage and retrieval of information, making it difficult for them to organize and recall memories in an orderly manner like humans. This significantly limits their application in complex scenarios.
Recently, the MIRIX system, co-developed by PhD student Wang Yu from the University of California, San Diego, and Professor Chen Xi from New York University, has made a groundbreaking change. Not only does it break down the barriers of AI memory technology with an open-source approach, but its innovative design, featuring a multimodal and multi-agent architecture, endows intelligent assistants with memory organization and retrieval abilities that are close to human-like. The simultaneous release of a Mac application allows this cutting-edge technology to quickly be integrated into everyday scenarios.
Six Memory Modules Build AI’s “Cognitive Warehouse”
The most striking breakthrough of MIRIX is its construction of a classification storage system that is highly similar to human memory patterns. Unlike traditional AI systems, which tend to store information in a chaotic manner, this system divides memories into six clearly defined modules, forming a hierarchical “cognitive warehouse.”
Core Memory serves as the “personality cornerstone” of the system, occupying the most crucial storage resources. It records essential user information such as names and preferences, as well as the AI’s own personality traits. When storage capacity exceeds 90%, a refinement mechanism is automatically triggered to ensure that critical information is not diluted.
Situational Memory acts like the AI’s “diary,” where all user interactions and conversations are precisely timestamped, providing a clear timeline for reviewing past interactions.
At the abstract cognition level, Semantic Memory stores general knowledge and factual information, such as “Earth is the third planet from the Sun,” providing the AI with basic knowledge reserves. Procedural Memory functions like an “operation manual,” recording task flows and step-by-step guides to enable the AI to execute standardized actions like “deploying applications” or “formatting conversions.”
For real-time interaction scenarios, Resource Memory temporarily saves materials such as documents, screenshots, and voice notes the user is processing, ensuring seamless handling of multimodal information.
The most noteworthy feature is the Knowledge Vault Module, which uses a tiered encryption mechanism to store sensitive information like API keys and passwords. With strict access control, this ensures data security and addresses privacy concerns in AI memory systems.
Multi-Agent Collaboration Enables “Precise Management” of Memory
If the six memory modules are the “hardware foundation” of MIRIX, then the multi-agent workflow acts as its “operating system.” The system innovatively introduces one Meta Memory Manager and six Sub Memory Managers, forming a collaborative network with clear roles. When new information enters the system, the Meta Manager first acts as the “central dispatcher,” deeply analyzing and determining which memory module the information belongs to, then routes it to the corresponding sub-module.
Each Sub Memory Manager focuses on updating and maintaining memory within its domain. For instance, the Situational Memory sub-module automatically removes duplicate conversation records, and the Semantic Memory sub-module merges similar factual information. Continuous redundancy cleanup ensures the efficient operation of the memory database.
This architectural design has led to a significant performance leap. In the ScreenshotVQA (Screenshot Visual Question Answering) benchmark test, which requires deep understanding of the correlation between images and text, MIRIX’s accuracy surpassed traditional RAG (Retrieval-Augmented Generation) methods by 35%, while compressing storage overhead by 99.9%. Compared to long-text processing methods, accuracy was increased by 410%, while storage costs were reduced by 93.3%. In the LOCOMO long-dialogue task, the system outperformed all existing technologies with an 85.4% performance, particularly excelling in complex conversations involving multi-turn contextual connections, demonstrating the precise retrieval ability of long-term memory. These results confirm the superiority of the multi-agent collaboration model in handling complex memory tasks.
Desktop Application Brings AI Memory into Daily Life
The ultimate value of technology lies in its practical application. The MIRIX team has also launched a Mac application, allowing ordinary users to directly experience this cutting-edge technology. Users simply need to install the app and configure the Gemini API Key, and the system will automatically start recording every detail of their digital life—from email exchanges and document editing to web browsing. All digital footprints will be neatly organized into the memory system.
More importantly, users can interact directly with the intelligent agent, asking questions like, “Where is the report I edited last Wednesday afternoon?” or “How can I retrieve the lost software registration code?” The AI will quickly extract the relevant content from the memory database and provide accurate answers.
In terms of privacy protection, the application uses a local SQLite database to store all memory data, meaning user information is never uploaded to the cloud, eliminating the risk of data leakage. This “local memory + intelligent interaction” model is reshaping the relationship between users and intelligent devices: when AI can accurately remember user habits, preferences, and even work progress, intelligent assistants will evolve from passive responders to proactive “digital partners.” Whether it’s professionals needing to quickly revisit project histories or regular users organizing fragmented information, MIRIX shows great potential in systematizing the chaotic digital life.
As the world’s first open-source multimodal, multi-agent AI memory system, the launch of MIRIX not only promotes the standardization of AI memory technology but also provides the industry with a reusable technical paradigm through its modular design and collaborative architecture. With the participation and iteration of the open-source community, future versions may introduce memory models that better fit different scenarios, allowing artificial intelligence to truly achieve “unforgettable memory and versatile reasoning.”
