Chinese researchers unveil MemOS, the first ‘memory operating system’ that gives AI human-like recall

Including the team of researchers of leading agencies Shanghai Jiao Tong University and University of Zhejiang The man has called the first “memory operating system, which prevents the main restriction that prevents the main restrictions preventing the main restrictions from steadfast memory and obstruction of AI systems.

System, called MemosTreatment of memory as a source of calculation and improves over time – a lot of traditional operating systems as the cpu and how to manage storage sources. Research, Was published on July 4 in the archiveCompared to Openai’s memory systems, it demonstrates a 159% increase in temporary thinking tasks, including increasing 159% of increasing performance.

“Large Language Models (LLS) have become an important infrastructure for artificial general intelligence (AGI), but the lack of well-prescribed memory management systems, prevent long-term justification, sustainable personalization and knowledge sequence,” researchers their paper.

AI systems are struggling with continuous memory between conversations

Current AI systems face the things they call researchers “Memory silo“Problem – a fundamental architectural restriction that prevents us from maintaining long-term relationships with users.

As some solutions Returned generation (dwarf) During the talks, the investigators are left to apply for this, the problem is that it works deeper than the employment of simple data – it is related to the establishment of systems that can learn from experience as a person’s memory.

“The existing models are mainly relied on static parameters and short-term contextual states, keep track of user choices or update knowledge during extended periods,” the team explains. This restriction appears in the intended enterprise parameters, which is expected to maintain a context between the complex, multi-stage workflows that can pass through the day or weeks or weeks.

The new system offers dramatic progress in AI reasoning positions

Memos Provides a majority different approach to what researchers called “Memcubu“Standard memory units that can cover various data and transferable and developed over time.

Test Lokomo BenchmarkEvaluating the intensive reasoning tasks to memory, memories have been set up on all categories. The system, compared to the implementation of Openai’s memory, with strong benefits in complex meditation scenarios, which requires many talks between numerous conversations.

“Memos-0630) is a consistent basis in all categories in a row, powerful bases like memorial, Langmem, ZEP and Openai-memory as challenging settings, such as a very-hop and temporary justification,” research. The system increased significant efficiency in certain configuration in certain configurations in certain configurations through innovative KV-cache memory injection mechanism.

This performance gains show that the memory problem is a more important limit than previously understood. By treating the memory as a first-class calculation source, Memos It seems to open thinking abilities restricted by architectural restrictions.

Technology can change the placement of artificial intelligence in enterprises

An entity can be transformative as an effective effects for the EI placement, especially enterprises, customers and employees in complex, sustainable relationships AI systems. Memos ensures how researchers describe “Cross-platform memory migration“To allow AI memories to be portable on various platforms and devices, it reduces what they call”Memory Islands“This is currently in the context of user within special applications.

Many user experience in investigated on an AI platform, review the user experience when the concepts investigated on an AI platform. A marketing group can prepare detailed customer staff through chatgept conversations, just to start from scratch while crossing a different AI tool for campaign planning. Memories address this by creating a standard memory format that can move between these systems.

The study also belongs to the potential “Paid memory modulesDomain experts can order the memory units taken in the memory units, “An experienced doctor can learn how to manage the status of a unique autoimmune.

This Marketplace model can vuly distribute special knowledge in the AI ​​systems, which creates new economic opportunities for professionals, creating new economic opportunities for professionals in democratizing domain knowledge. For enterprises, it can quickly place the AI ​​systems with a deep expertise in special areas in traditional costs and time related to special trainings and special trainings.

Three-layered design mirrors traditional computer operating systems

This Technical architecture of memos AI reflects ten years from the design of the traditional operating system adapted for unique challenges of memory management. The system applies a three-layer architecture: an interface for API calls, an operating layer for memory planning and lifelong management and infrastructure layer for storage and management.

System MEMSCHEDULER Component manages different types of memory dynamically – to permanent parameter changes from temporary activation situations – the selection of optimal storage and search strategies based on task requirements. This usually represents a significant detection of the current approaches that accepts the memory as a completely static (model parameters) or entirely ephemeral (conversion context).

“Focus reveals and reconstruction and reconstruction of the model to structured memory and reconstruction and reconstruction of” because researchers’Mem-training“Paradigms. This architectural philosophy proposes to be a basic revision of how to build AI systems, more dynamic, experienced learning from the current paradigility of mass training.

Parallels are surprising to the development of the operating system. Like the first computers required to manage the memory allocation of programmers, the current AI systems require to close the extent to how developments are flooded among different components. Memos This complex, which prevents a new generation of a new generation AI application that can be built on complex memory management without demanding a deep technical expertise.

Researchers leave the code as an open source to accelerate adoption

The team was released Memos As an open source project, The full code available in GitHub Integrated support to large AI platforms, including Huggingface, Openai and Ray. This open source strategy is designed to promote community development than to take a property approach to acceptance and restrict community development.

“We hope that memories help to develop stable generators, developing memory agents,” Project devise Zhiyu Li Gitub depot. Currently, the system supports Linux platforms, with Windows and MacOS support, the team is immediately accepted by the entity and the developer immediately in the possibility of consumer.

Open source release strategy reflects a wider trend in a research that explicitly shared the improvement of the basic infrastructure to benefit from all ecosystems. This approach has accelerated an innovation historically in areas such as deep learning frames and can affect memory management in AI systems.

Technical giants races to solve AI memory restrictions

Research, major AI companies, fighting the restrictions of current memory approaches, are struggling with the restrictions of current memory approaches that stress the industry as the industry. Openai recently introduced Memory features for Chatgptwhile Anthropical, GoogleAnd other providers have experienced with different forms of persistent context. However, these applications are usually limited to coverage and often no systematic approach Memos provides.

The time of this study shows that memory management is a critical competitive battlefield in the development of the AI. Companies that can effectively solve the memory problem can have important advantages in user detention and satisfaction, because AI systems will be able to establish deeper and more useful relationships over time.

Industrial observers have long predicted that the next bigger progress in the AI ​​will be better than larger models or more training information, but better than architectural innovations. Memory management represents this type of fundamental development – it is a kind of unlocking new applications and use unavailable cases in existing citizenship systems.

Development represents a larger change in research, more than the state, resistant systems, which develops and developing timely knowledge – is important for artificial general intelligence. For enterprise technology leaders who evaluate AI applications Memos An important progress can be made in the establishment of AI systems, which can protect the context and improve the context, rather than approaching each interaction as isolated.

The research team shows that they plan to develop and develop self-evolving memory blocks, self-evolving memory blocks, self-developing memory market. However, perhaps the most important effect of memories will not be a special technical application, on the contrary, such as the first class source source, the priest of the memory can unlock dramatic progress in AI opportunities. In an industry, memories, which focus mainly on the size and training information, show that the next progress can come from better architecture.