Recent developments in LLM agents have largely focused on improving the capacity of the execution of complex tasks. However, a critical dimension remains under -explored: memory – the capacity of agents to persist, to recall and reason on the information specific to the user in time. Without persistent memory, most LLM-based agents remain stateless, unable to build a context beyond a single invite, limiting their usefulness in real contexts where consistency and personalization are essential.
To solve this problem, Mirix AI introduces Mirix, a modular multi-agent memory system explicitly designed to allow a robust long-term memory for LLM-based agents. Unlike flat systems and purely focused on the text, Mirix incorporates structured types of memory through methods – including visual entry – and is built on a multi -aging architecture coordinated for memory management.
Basic architecture and memory composition
Mirix has six specialized composition memory components, each governed by a corresponding memory manager:
- Central memory: Store the persistent agent and user information, segmented in “persona” (agent, tone and behavior) and “human” (user facts such as name, preferences and relationships).
- Episodic memory: Capture of horrible events and user interactions with structured attributes such as event_type, summary, details, actors and horoditing.
- Semantic memory: Eacls abstract concepts, knowledge graphics and named entities, with entries organized by type, summary, details and source.
- Procedure memory: Contains structured workflows and task sequences using clearly defined steps and descriptions, often formatted as JSON for easy handling.
- Resource memory: Maintains references to external documents, images and audio, recorded by title, summary, type of resource and content or link for contextual continuity.
- Knowledge vault: Secures textual facts and sensitive information such as identification information, contacts and API keys with strict access controls and sensitivity labels.
A Meta Memory Manager Orchestra The activities of these six specialized managers, allowing a routing of intelligent messages, hierarchical storage and specific recovery operations. Additional agents – with roles such as the cat and the interface – are in collaboration within this architecture.
Active recovery and interaction pipeline
Mirix's main innovation is its Active recovery mechanism. On the user's entry, the system first depicts a subject independently, then recovers the relevant memory inputs of the six components and finally label the data recovered for contextual injection in the resulting system invite. This process decreases dependence on obsolete parametric knowledge and provides a much stronger response base.
Several recovery strategies – in particular embedding_match,, bm25_matchAnd string_match—S are available, ensure precise access and in memory context. Architecture allows a new expansion of recovery tools as needed.
Implementation and application of the system
Mirix is deployed as an application of a multiplatform assistant developed with React-Electron (for the user interface) and Uvicorn (for the Backend API). The wizard monitors the screen activity by capturing screenshots every 1.5 second; Only non -redundant screens are kept and memory updates are triggered by lots after collecting 20 unique screenshots (about once per minute). Downloads on the Gemini API are streaming, allowing effective processing of visual data and latency less than 5 seconds to update memory from visual inputs.
Users interact via a cat interface, which is dynamically based on the agent's memory components to generate contextual personalized responses. The semantic and procedural memories are made as extensible arbres or lists, offering transparency and allowing users to audit and inspect what the agent “remembers” them.
Evaluation on multimodal and conversational references
Mirix is validated on two rigorous tasks:
- Screenshot: A visual reference of questions and answers requiring a long-term persistent memory on high-resolution screenshots. Mirix surpasses the basic lines of the generation of the generation with recovery – siglip and gemia – by siglip and gemals – by 35% in the LLM-As-Aa-Judge precisionwhile reducing recovery storage requirements by 99.9% compared to text methods.
- Loco: A textual reference evaluating long conversation memory. Mirix realizes 85.38% average precisionoutperform solid open source systems such as Langmem and MEM0 by more than 8 points, and the approach of the higher limits of full context sequence.
The modular design allows high performance in the fields of multimodal and text inference only.
User cases: portable devices and the memory market
Mirix is designed for extensibility, with a medium for light AI laptops – including intelligent glasses and pins – via its effective modular architecture. Hybrid deployment allows both the handling of memory based on devices and the cloud, while practical applications include the summary of real -time meetings, granular location and context recall, and dynamic modeling of user habits.
A visionary characteristic of Mirix is the Memory market: A decentralized ecosystem allowing secure memory sharing, monetization and collaboisation of collaborative AI between users. The market is designed with fine-grain privacy controls, end-to-end encryption and decentralized storage to guarantee the sovereignty of the data and the self-ownership of users.
Conclusion
Mirix represents a significant step towards the endowment of agents based on LLM with a human -type memory. Its structured multi-agent composition architecture allows abstraction of robust memory, multimodal support and real-time contextual reasoning. With empirical gains through difficult benchmarks and an accessible multiplatform application interface, Mirix establishes a new standard for AI systems with memory.
Faq
1. What makes Mirix different from existing memory systems like MEM0 or ZEP?
Mirix introduces multi-component composition memory (beyond the storage of text passage), multimodal support (including vision) and a multi-agent recovery architecture for long-term memory management, more evolving, precise and rich in context.
2. How does Mirix ensure low latency memory updates from visual entries?
Using downloads in combination with gemini APIs, Mirix is able to update visual memory based on screenshot with less than 5 seconds latency, even during active user sessions.
3. Is Mirix compatible with closed source LLMs like GPT-4?
Yes. Since Mirix works as an external system (and not as a model plugin or recycling), it can increase any LLM, regardless of its basic architecture or license, including GPT-4, Gemini and other proprietary models.
Discover the Paper,, Github And Project. All the merit of this research goes to researchers in this project.
Sponsorship opportunity: Reach the most influential AI developers in the United States and Europe. 1M + monthly players, 500K + community manufacturers, endless possibilities. (Explore sponsorship)
Sajjad Ansari is a last year's first year of the Kharagpur Iit. As a technology enthusiast, he plunges into AI's practical applications by emphasizing the understanding of the impact of AI technologies and their real implications. It aims to articulate complex AI concepts in a clear and accessible way.
