Chinese researchers unveiled notes, the first “memory operating system” gives a summons of artificial intelligence -like intelligence

A team of researchers from leading institutions, including Shanghai Jiao Tong and Zhejiang University He developed what they call the first “memory operating system” of artificial intelligence, dealing with the basic restriction that hindered artificial intelligence systems from achieving a human -like memory and learning.

The system is called NotesThe memory is treated as a basic account resource that can schedule, share and develop over time – such as how to manage traditional operating systems for the CPU resources and storage resources. Search, It was published on July 4 on ArxivIt shows significant improvements in performance on current methods, including an increase of 159 % in temporal thinking tasks compared to Openai memory systems.

“LLMS models have become an essential infrastructure for artificial general intelligence (AGI), however their lack of well -specific memory management systems hinders the development of long thinking, continuous allocation, and the consistency of knowledge,” researchers write in Determine them.

Artificial intelligence systems are struggling with the ongoing memory via conversations

Current artificial intelligence systems face what researchers call.Memory silo“The problem-basic architectural restrictions that prevent them from maintaining coherent and long-term relationships with users. Each conversation or session begins mainly from the zero point, with the inability of the models to retain preferences, accumulated knowledge or behavioral patterns through interactions.

While some solutions Pre -recovery generation (rag) Trying to process this matter by withdrawing external information during the talks, the researchers argue that these are “sexual solutions without controlling the life cycle.” The problem works deeper than the restoration of simple information – it is related to creating systems that can really learn and develop from experience, just as it resembles human memory.

“Current models depend mainly on fixed parameters and short -term contextual states, which limits their ability to track user preferences or update knowledge over the extended periods,” the team explains. This restriction becomes particularly evident in the settings of institutions, as AI systems are expected to maintain the context through a complex and multiple stages work that may extend for days or weeks.

A new system provides exciting improvements in the tasks of thinking about artificial intelligence

Notes It provides a different approach mainly through what researchers call “Memcubes-The unified memory units that can envelop different types of information, their composition, deportation and development over time. These range from explicit knowledge based on the text to the amendments at the teacher level and stimulation cases within the model, which creates a unified framework for memory management that was not previously present.

Test on Locomo measurementThat evaluate the intensive thinking tasks of memory, the notes have constantly outperformed the basic lines established in all categories. The system achieved a total improvement of 38.98 % compared to the implementation of Openai’s memory, with particularly strong gains in complex thinking scenarios that require communication information through a multiple conversation.

“Notes (0630) are ranked first in all categories, outperforming strong foundation lines such as MEM0, Langmem, Zep and Openai-Memory, with special margins in particular in difficult places such as multiple thinking and temporal logic,” according to the research. The system also provided significant improved efficiency, with a decrease of up to 94 % in time access time to the first in certain configurations through the innovative KV-CACHE memory mechanism.

These performance gains indicate that the memory bottle neck was more important restrictions than previously understood. By dealing with memory as a first -class account resource, Notes It appears to open the thinking possibilities that were previously restricted to architectural restrictions.

This technology can reshape how to spread artificial intelligence companies

The effects of the AI’s deployment of AI can be transformative, especially since companies are increasingly dependent on artificial intelligence systems of complex and continuous relationships with customers and employees. Notes What can be described as researchers as “Memory deportation via platforms“Allow the memories of artificial intelligence to be carried through different platforms and devices, and to break what they call”Memory islands“This prohibits the context of the user currently within specific applications.

Think about the current frustration that many users face when the visions that were explored on the AI ​​platform cannot be explored. The marketing team may develop detailed customer personalities through conversations with ChatGPT, only to start the zero point when turning into a different tool of artificial intelligence for planning for campaigns. This notes are dealt with by creating a unified memory format that can move between systems.

The research also determines the possibility “Paid memory units“Where the two field experts can pack their knowledge of the purchase memory units. Researchers perceive the scenarios where“ the medical seeker in clinical rotation may want to study how to manage the rare autoimmune condition. An experienced doctor can cover up diagnostic reasoning, interrogation paths, and typical condition patterns in an organization’s memory “can be installed and used by other artificial intelligence systems.

This market model can mainly change how to distribute specialized knowledge and invest in artificial intelligence systems, creating new economic opportunities for experts while adding a democratic character to access high -quality knowledge. For institutions, this may mean the spread of artificial intelligence systems quickly with deep experience in specific areas without traditional costs and timetable tables associated with allocated training.

It designs three layers of traditional computer operating systems

the Technical architecture for notes Contracts of learning reflects the traditional operating system design, adapted to the unique challenges of managing artificial intelligence. The system uses a three -layer structure: an API interface layer, an operating layer for memory schedule, life cycle management, and storage and governance infrastructure layer.

order MEMSCHEDULER The dynamic component manages different types of memory – from temporary activation situations to permanent parameters ’adjustments – choosing optimal storage strategies and retrieval based on use patterns and important requirements. This is a great exit from the current methods, which the memory usually treats either completely fixed (included in the parameters of the model) or fast -demise (limited to the context of the conversation).

“The focus turns from the amount of knowledge that the model learns once to whether the experiment can turn into an organization’s memory, recover and rebuild it repeatedly,” the researchers note, describing their vision of what they call.Memory training“Models. This architectural philosophy indicates the basic rethinking of how to design artificial intelligence systems, and to stay away from the current model of tremendous training towards the most dynamic learning.

Similar to the development of the operating system striking. Early computers have also required programmers to manually manage memory customization, the current artificial intelligence systems require developers to organize how information flows between different components. Notes Abstracts this complexity, which is likely to enable a new generation of artificial intelligence applications that can be built over advanced memory management without the need for deep artistic experience.

The researchers launch an open source symbol to accelerate the adoption

Establish the team Notes As an open source project, with A complete symbol is available on Jabbap Supporting integration of major artificial intelligence platforms including Hugingface, Openai and OLLAMA. This open source strategy is designed to accelerate the adoption and encourage community development, rather than a royal approach that may limit the implementation on a large scale.

“We hope Memos will help the artificial intelligence systems from fixed generators to continuous developers, which depend on memory,” Zhiyu Li commented in the GitHub warehouse. The system currently supports Linux platforms, with Windows and MacOS support, indicating that the team gives priority to institutions and developers on the possibility of immediate access to the consumer.

The open source version strategy reflects a wider direction in artificial intelligence research, as constituent infrastructure improvements are shared in public in favor of the entire ecosystem. This approach has historically accelerated innovation in areas such as deep learning frameworks and can have similar effects of memory management in artificial intelligence systems.

Technology giants race to solve artificial intelligence restrictions

The research reaches when the major artificial intelligence companies contradict the restrictions of current memory approaches, highlighting the importance of this challenge to the industry. Openai was recently presented Memory features for chattingwhile manand GoogleAnd other service providers have experienced different forms of constant context. However, these applications were generally limited in the range and often lacking the systematic approach Notes He provides.

The timing of this research indicates that the memory management has appeared as the basis of a decisive competitive battle in developing artificial intelligence. Companies that can solve the memory problem effectively acquire great advantages in keeping the user and their consent, as their artificial intelligence systems will be able to build deeper and more useful relationships over time.

Industry observers have long expected that the following major penetration in artificial intelligence will not necessarily come from larger models or more training data, but from architectural innovations that better simulate human cognitive capabilities. Memory management represents this type of basic progress exactly – which can open new applications and use incomplete situations with current sexual systems.

Development is part of a broader transformation in artificial intelligence research towards more occurrence and continuous systems that can accumulate and develop knowledge over time – the capabilities that are necessary for artificial general intelligence. For institution technology leaders who evaluate artificial intelligence applications, Notes It can be a great progress in building artificial intelligence systems that maintain context and improve over time, rather than dealing with each reaction as isolated.

The research team indicates that they are planning to explore memory sharing via the model, self -evolving memory blocks, and the development of a broader ecosystem “memory market” in future work. But perhaps the most important effect of notes will not be the specific technical implementation, but rather the evidence that memory processing as a first -class mathematical resource can open exciting improvements in the capabilities of artificial intelligence. In the industry, it has largely focused on the size of the model model and training data, Memos indicates that the following penetration may come from a better structure instead of the largest computers.

Daily visions about business use cases with VB daily

If you want to persuade your boss at work, you have covered VB Daily. We give you a scoop from the inside about what companies do with obstetric artificial intelligence, from organizational transformations to practical publishing operations, so that you can share visions of the maximum return on investment.

Read with us privacy policy

Thanks for subscribing. Check more VB news bulletins here.

An error occurred.