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

A team of researchers from leading institutions, including Shanghai Jiao Tong University And Zhejiang University Developed what they call the first “memory control system” for AI, with a fundamental disability that has impeded models by achieving human persistent memory and learning.

The system called Memostreats memory as a core computational source that can be planned, shared and evolved over time – similar to how traditional operating systems manage CPU and storage sources. The research, Published on July 4 at ArxivShows significant performance improvements in relation to existing approaches, including a boost of 159% in temporary reasoning tasks compared to the memorial systems of OpenAi.

“Large language models (LLMs) have become an essential infrastructure for artificial general intelligence (AGI), but their lack of well-defined memory management systems hinders the development of long-context reasoning, continuous personalization and knowledge consistency,” the researchers write in their paper.

AI systems struggle with persistent memory in different conversations

Current AI systems are confronted with what researchers call the problem “memory silo”-a fundamental architectural limitation that prevents them from maintaining coherent, long-term relationships with users. Every conversation or session in essence starts completely, with models that are unable to preserve preferences, accrued knowledge or behavioral patterns between interactions. This creates a frustrating user experience because an AI assistant can forget the food restrictions of a user in one conversation when asked about restaurant recommendations in the following.

While some solutions like Pick-up-augmented generation (RAG) Try to tackle this by withdrawing external information during conversations, the researchers claim that these “Stateless solution remain without a life cycle control.” The problem runs deeper than simple information in the field of information – it is about creating systems that can really learn and evolve from experience, just like human memory does.

“Existing models are mainly based on static parameters and short -term contextual situations, which means that their ability to follow user preferences or to update knowledge about longer periods,” the team explains. This limitation is particularly clear in Enterprise institutions, whereby AI systems are expected to maintain the context about complex, multi-phase workflows that can span those days or weeks.

New system provides dramatic improvements in the reasoning tasks of AI

Memos introduces a fundamentally different approach through what the researchers call ‘Memcubes“Standardized memory units that can be brought in different types of information and can be compiled, migrated and evolved over time. These vary from explicit text-based knowledge to adjustments at parameter level and activation states within the model, creating a uniform framework for memory management that did not exist before.

Test on the LOCOMO -BENCHARKThat memory -intensive reasoning tasks evaluates, Memos consistently exceeded established basic lines in all categories. The system achieved a general improvement of 38.98% compared to the memory implementation of OpenAI, with particularly strong profit in complex reasoning scenarios that require information about multiple conversation.

“Memos (Memos-0630) consistently first scores in all categories and surpasses strong bases such as Mem0, Langmem, Zep and OpenAi-Memory, with particularly large margins in challenging institutions such as Multi-Hop and temporary reasoning,” the study said.

The system also yielded substantial efficiency improvements, with up to 94% reduction in time-to-five-topping latency in certain configurations through its innovative KV-Cache memory injection mechanism.

These performance profits suggest that the memory bottleneck has been a more important limitation than understood before. By treating memory as a first -class computational source, Memos Seems to unlock reasoning opportunities that were previously limited by architectural limitations.

Technology can reform how companies use artificial intelligence

The implications for Enterprise AI implementation can be transforming, in particular because companies are increasingly dependent on AI systems for complex, continuous relationships with customers and employees. Memos makes it possible to describe the researchers as ‘platform-dependent memory migration’, so that AI memory can be portable on different platforms and devices, so that what they call ‘memory islands’, which are currently breaking down the user context of specific applications in the false count.

Consider the current frustration that many users experience when insights that are explored in one AI platform cannot transfer to another. A marketing team can develop detailed customer personas through conversations with Chatgpt, only to start all over again when switching to another AI tool for campaign planning. Memos tackle this by making a standardized memory format that can move between systems.

The research also outlines potential for ‘paid memory modules’, where domain experts can pack their knowledge in buyable memory units. The researchers propose scenarios in which “a student of medical rotation in clinical rotation may want to study how to manage a rare car -immune disorder. An experienced doctor can diagnostic heuristics, paths and typical case patterns in a structured memory in a structured memory”.

This marketplace model could fundamentally change how specialized knowledge is distributed and monitored in AI systems, creating new economic opportunities for experts and at the same time democratize access to high-quality domain knowledge. For companies this may mean that AI systems quickly use AI systems with deep expertise in specific areas without the traditional costs and timelines related to adapted training.

Three-layer design reflects traditional computer control systems

The Technical architecture of memos Learning for decades of learning from the traditional design of the operating system, adapted for the unique challenges of AI memory management. The system uses a three-layer architecture: an interfacal layer for API calls, an operating layer for memory planning and life cycle management and an infrastructure layer for storage and governance.

The Memscheduler component of the system manages dynamically different types of memory – from temporary activation states to permanent parameter adaptations – selecting optimal storage and collection strategies based on use patterns and task requirements. This means a significant deviation from the current approaches, which usually treat the memory as fully static (embedded in model parameters) or completely short -lived (limited to conversation context).

“The focus shifts from how much knowledge the model learns once or the experience can convert into structured memory and can repeatedly pick up and reconstruct,” the researchers notice, who describe their vision for what they call “mem-training” paradigms. This architectural philosophy suggests a fundamental reconsideration of how AI systems should be designed, which leave from the current paradigm of mass pre-training to more dynamic, experiential learning.

The parallels with the development of the operating system are striking. Just as early computers programmers require memory allocation manually, current AI systems require that developers must carefully orchestrate how information flows between different components. Memos abstracts this complexity, making a new generation of AI applications possible that can be built on top of advanced memory management without requiring deep technical expertise.

Researchers release code as an open source to speed up acceptance

The team has released memos as an open-source project, with Full code available on Github and integration support for large AI platforms, including Huggingface, OpenAi and Ollama. This open-source strategy seems to be designed to accelerate acceptance and encourage community development, instead of pursuing its own approach that could limit the widespread implementation.

“We hope that memos help AI systems from static generators to continuously evolving, memory-driven agents,” noted project leader Zhiyu Li in the Github repository. The system currently supports Linux platforms, with planned Windows and MacOS support, which suggests that the team gives priority to the acceptance of companies and developers over the immediate accessibility of the consumer.

The open-source release strategy reflects a broader trend in AI research where fundamental infrastructure improvements are openly shared compared to the entire ecosystem. This approach has intake of innovation based on areas such as deep learning frameworks historically accelerated and can have similar effects for memory management in AI systems.

Tech Giants racing to resolve AI memory restrictions

The research arrives when large AI companies struggle with the limitations of current memory approaches, and emphasize how fundamentally this challenge has become for industry. OpenAi recently introduced Memory functions for Chatgptwhile Anthropic” Google And other providers have experimented with different forms of continuing context. However, these implementations are generally limited in size and often miss the systematic approach that memos offer.

The timing of this study suggests that memory management has emerged as a critical competitive battlefield in AI development. Companies that can effectively solve the memory problem can get significant benefits in the preservation and satisfaction of users, because their AI systems can build up relationships deeper, more useful relationships over time.

Industry observers have long predicted that the next major breakthrough in AI would not necessarily come from larger models or more training data, but of architectural innovations that better imitate human cognitive capacities. Memory management represents precisely this kind of fundamental progress – one that could unlock new applications and use cases that are not possible with the current stateless systems.

The development forms a part of a broader shift in AI research into more statefeous, persistent systems that can collect and evolve knowledge over time – possibilities that are essential for AGI. For leaders of Enterprise technology that evaluate AI implementations, memos can be significant progress in building AI systems that retain the context and improve over time, instead of treating any interaction as isolated.

The research team indicates that they are planning to explore cross-model memory, self-evolving memory blocks and a wider ecosystem of the “memory market”. But perhaps the most important impact of memos is not the specific technical implementation, but the proof that treating memory as a first-class computational source can unlock dramatic improvements in AI options. In an industry that has largely focused on the size of the scale model and training data, Memos suggests that the next breakthrough of better architecture instead of larger computers.