In the first half of 2026, memory evolved from a technical concept into a distinct industry track. In late May, Tencent Cloud launched Agent Memory, boosting long-term memory accuracy from 48% to 76%. In July, Alibaba Cloud intensively integrated memory capabilities into its database product line. Meanwhile, Volcengine open-sourced OpenViking, reducing token consumption costs by 91% through memory-driven optimization, with GitHub stars surpassing 21,000.
Financing for numerous vendors accelerated simultaneously. Memory Tensor completed a hundred-million-yuan-level Pre-A round. Red Bear AI completed a several-hundred-million-yuan A+ round, reaching a post-investment valuation of nearly 3 billion yuan. Capital is voting for the memory track with real money. The market is bustling, yet consensus on the definition of memory remains elusive, still in an early chaotic stage. Investors struggle to define boundaries, and enterprise clients find it hard to evaluate value. Thus, this nascent industry requires a fundamental research study to clarify its origins.
The "memory" mentioned within the scope of this report specifically refers to Agent Memory—the experiential asset accumulated, dynamically updated, and continuously guiding decisions during long-term interactions of AI agents, no longer conflated with concepts like knowledge bases or context windows. Through this report, iFenxi attempts to answer three questions: What exactly is memory? Why can memory form an independent market rather than just an accessory function of models? How large is the memory market, and who are the competitors?
Key Conclusions Memory is an independent and rapidly growing infrastructure track in the era of AI agents. iFenxi estimates that the size of China's memory market will grow from 1.44 billion yuan in 2025 to 64.25 billion yuan in 2030, with a compound growth rate exceeding 110%. Among them, the memory operating system will reach a scale of 21.85 billion yuan by 2030, accounting for about 34% of the memory market, making it the fastest-growing segment. In the future, the memory market will continue to evolve in four directions.
First, evolving from experience accumulation to decision accumulation, memory upgrades from a recording tool to a decision asset.
Second, Memory as a Service (MaaS), where memory capabilities shift from internal agent modules to independently callable services.
Third, the rise of edge and local memory, driven by data sovereignty demands pushing memory migration to the edge and local environments.
Fourth, the memory market moves towards standardization, with unified interfaces, formats, and evaluation systems driving cross-platform memory migration.
In terms of the competitive landscape, five types of players converge on the memory track, including cloud giants, model vendors, independent memory platforms, enterprise memory application vendors, and consumer memory products. Based on research on over 20 vendors, iFenxi establishes a competitive quadrant, with the horizontal axis measuring technical capabilities and the vertical axis measuring market influence, placing relevant vendors into four quadrants.
Figure 1: Competitive Quadrant of China's Memory Market
In the Leaders quadrant, Tencent Cloud Agent Memory and Zhibian Technology represent two distinct development paths. Relying on its cloud ecosystem and enterprise-grade compliance system, Tencent Cloud's memory products are mature with strong commercialization capabilities, representing the cloud giant route. As an independent memory platform, Zhibian Technology squeezed into the Leaders quadrant through self-evolution capabilities and citation chain traceability governance, proving that technical depth is sufficient for independent vendors to break through the ecological barriers of cloud giants. In the Innovators quadrant, Kimi Memory Space leverages its long-context genes, using a memory model to determine what to remember, offering a transparent and controllable experience.
Memory Tensor pioneers the MemOS memory operating system form, landing in state-owned enterprises and securing Pre-A round support from Huawei Hubble and Honor Strategic Investment. In the Practitioners quadrant, Red Bear AI is a typical representative. With over 500 clients, having completed a several-hundred-million-yuan A+ round with a post-investment valuation nearing 3 billion yuan, it demonstrates outstanding commercialization capabilities. In the Explorers quadrant, Baidu AI Cloud provides memory capabilities in the form of vector database modules, embedded within the internal scenarios of the Qianfan platform. This is still an early attempt by a cloud vendor, with product independence and market landing still in their infancy.
Based on the competitive quadrant, the competitive logic of the memory market can be outlined. The Leaders quadrant is still mainly dominated by cloud giants, which occupy the top positions relying on infrastructure, customers, and commercialization capabilities; independent memory platforms can also squeeze into the top Leaders relying on technical depth; model vendors are mainly clustered in the Innovators and Explorers quadrants, with clear directions but products not yet fully formed.
01 Market Definition
In today's AI industry, "memory" is one of the most abused terms. Some believe memory is a knowledge base, some think it's the ultra-long context window of a model, and others consider it a temporary cache for historical dialogues. The same term points to completely different technical implementations, making it hard for investors to judge boundaries and enterprise clients to evaluate value. iFenxi believes that the root of market confusion is that memory has never been rigorously defined. This chapter first corrects three common misconceptions about memory, then provides a clear definition, and finally establishes a usable classification framework.
1.1 Misconceptions in Market Perception
There are three misconceptions that downgrade the understanding of memory in general market perception.
The first misconception is equating memory with knowledge bases, RAG, or ultra-long contexts, which is the most common misinterpretation. Knowledge bases store general facts, RAG solves retrieval augmentation, and ultra-long contexts extend the window for a single dialogue; all three serve to let the model call more information in the current single response. Memory goes beyond this; it must also record what the agent has done in the past, the results, which approaches were effective, and let these experiences influence future decisions. Knowledge can be ingested offline, but experience can only be accumulated over time. Treating memory as knowledge storage erases the most critical time dimension in memory.
The second misconception is treating memory as part of the foundational model's capabilities. Some argue that as long as the foundational model extends the context window long enough or has enough parameters, the memory problem is automatically solved. This treats memory as a mere accessory function of the model. iFenxi finds this view untenable. Model memory solves the immediate retrieval of information in a single inference, cleared once the session ends, whereas memory must persist across sessions, tasks, and versions over the long term. The former is a computing power issue; the latter is a system architecture issue. Cramming memory into model capabilities obscures a key fact: memory is an infrastructure capability that requires independent design, independent operation and maintenance, and independent evolution.
The third misconception is treating enterprise memory and consumer memory as two completely different tracks. The market habitually describes enterprise memory as knowledge management, process accumulation, and organizational collaboration, while consumer memory is described as the preferences and habits of personal assistants. iFenxi believes these two share the same underlying origin. The project experience accumulated by an organization and the usage habits accumulated by an individual are essentially the same thing—experiential assets formed through long-term interactions of agents. The only differences lie in the ownership, permissions, and governance intensity of the experiential assets; the principles of memory itself are no different. Severing the two will cause vendors to miss the opportunity to compete in both markets simultaneously.
1.2 Definition of Memory
The essence of memory is the experiential asset continuously accumulated by an agent during its growth. iFenxi defines memory as the experiential asset formed, stored, and dynamically updated by an agent during long-term interactions, supporting the agent in reusing past information across sessions and models, and making more effective decisions and actions. To understand this definition, three things must be clarified. First, the boundary between knowledge and memory. Knowledge is static facts that can be called upon stably over the long term; memory is the experience unique to a specific agent or organization, growing more extensive and accurate with use. Memory records what this agent or organization has experienced and learned in specific contexts. Mistaking a knowledge system for a memory system is the source of Misconception One. Second, the connotation of memory goes far beyond model memory. Model memory is the context for a single inference, allowing the model to think more comprehensively in the current round; memory is the accumulation of experience across sessions, making the agent stronger over a longer time scale. Model memory serves thinking; memory serves growth. The former is just a small workspace of the latter during the instant of inference. Third, the composition of memory is composite. It includes both individual and organizational experience; both factual information and behavioral patterns, decision logic, and long-term reflections. Memory is not simple data retention; it simultaneously records the facts that occurred and the judgment habits, action patterns, and review conclusions formed by the agent as a result.
1.3 Four Core Capabilities of Memory
To determine whether a memory system is viable, it must simultaneously achieve understanding, evolution, usability, and trustworthiness. Multimodal capability solves understanding. Memory should uniformly understand text, images, audio, video, and business system data.
Agents do not face single-format inputs; an industrial drawing, a piece of equipment time-series log, or a voice command could all become experiences to be accumulated. If a memory system cannot uniformly encode these heterogeneous pieces of information into a single experience, it can only store fragments in isolated format silos. For example, manufacturing enterprises can solidify multimodal drawings and equipment time-series logs into reusable industrial memory assets, stepping beyond document retrieval to support equipment fault prediction, accumulate O&M experience, and help new employees get up to speed quickly. Self-evolution capability corresponds to evolution. Memory should be able to self-update, compress, reflect, forget, and transfer. Memory is not an ever-increasing log; as experience accumulates, it needs to eliminate outdated information, compress redundant details, distill lessons from multiple failures into reflections, and transfer experience learned in one task to new tasks.
Product capability solves usability. Memory should support continuous read/write, full lifecycle management, cross-platform deployment, and handle versioning and conflicts. For a memory system to be adopted by real businesses, merely solving technical principles is far from enough; it needs to support high-frequency continuous read/write, manage the full lifecycle of memory from creation to archiving, support cross-platform migration, and handle versioning and conflicts in multi-person collaboration. Governance capability solves trustworthiness. Memory should possess permission management, isolation, traceability, and trust capabilities. When memory bears the experiential assets of an organization, who can read, who can modify, which experiences belong to which project, and where a certain conclusion comes from all become unavoidable issues.
1.4 Memory Types
Having clarified the capabilities memory needs to possess, let's look at the types of memory itself. Classifying memory is to provide the market with a common coordinate system.
First, memory can be divided by content into three types: semantic memory, episodic memory, and procedural memory. This classification originates from the CoALA framework in the field of cognitive science. Semantic memory answers "what I know," representing objective knowledge and facts; episodic memory answers "what I have done and the results," representing experiences; procedural memory answers "how I do it and why," representing decision logic and behavioral patterns. It should be noted that these three types of memory are parallel; an agent may simultaneously possess a large amount of semantic memory, a small amount of episodic memory, and gradually forming procedural memory. The difference among the three lies only in content attributes, not maturity.
Second, memory can also be divided by the subject of ownership into three levels: individual memory, project memory, and organizational memory. Individual memory belongs to a single agent or user, recording independently formed experiences; project memory belongs to a specific project or task, recording collective experience accumulated during project execution; organizational memory belongs to the entire organization, recording long-term accumulated, cross-project reusable experiences and practices at the organizational level. Currently, memory vendors in the market have launched corresponding products mapping this three-level structure.
02 Independence of the Memory Market
Having defined what memory is, the next question is why memory can become an independent market and why it shouldn't just be an accessory function of foundational models. The reasons are twofold: First, the experience borne by memory does not belong to any model and naturally coordinates with multiple models; Second, building memory has independent data capability thresholds, which is entirely different from the data governance of models. Combined, these two reasons form the basis for the establishment of the memory market.
2.1 Memory Does Not Belong to the Model
Models are responsible for the thinking of agents; memory is responsible for the growth of agents. Models determine what an agent can do; memory determines how well it can do it. Models solve reasoning capabilities; for the same question, the stronger the model, the better the answer in the current round. However, the value of an agent lies not only in a single answer but in understanding an organization or a user better through long-term interactions. Imagine two agents using the same model: one has accumulated the project experience of the enterprise over the past two years, and the other has not. Faced with the same new problem, the performance gap between the two will be massive. The cause of this gap is memory. Models are responsible for doing a good job in the current round of thinking; memory is responsible for accumulating all past experiences so that the next round of thinking starts from a higher baseline. The project experience an organization spends two years accumulating belongs to the organization itself, not to any model vendor. The long-term accumulated usage preferences of a user belong to the user, not to any model vendor. Models can be changed, upgraded, or switched from Vendor A today to Vendor B tomorrow, but these experiential assets must persist across models and products over the long term. If memory is attached to a specific model, changing the model once means the enterprise has to reset all past experiences to zero, which no enterprise can afford. Cross-model, cross-product, long-term persistence—these are precisely the characteristics an infrastructure should have. Infrastructure is relied upon by multiple parties and does not change with the replacement of upper-layer applications; a database won't be replaced just because an upper-layer software is changed. Memory accumulates long-term experience at the organizational and user levels, similarly requiring cross-model, cross-product long-term persistence. Therefore, the positioning of memory can only be an independent AI infrastructure. iFenxi judges that an enterprise's most valuable experiential assets cannot be bet on a model vendor that might be replaced at any time. This is the root of why memory can become an independent market. This judgment is the first cornerstone for the establishment of the memory market.
2.2 Unique Data Threshold of Memory
The second reason memory can form an independent market is its data threshold. Memory must process dirty data in real businesses, which is entirely different from the clean corpora used in model training. This data threshold cannot be solved during the model training process. Specifically, memory vendors need to solve three core data problems. First, data preprocessing and parsing. Memory needs to accumulate heterogeneous data in real businesses, such as nested tables and cross-page tables in PDFs, or complex annotations in mechanical drawings. Foundational models suffer a sharp drop in accuracy when parsing these files. Memory vendors mostly need to develop proprietary parsing small models to specifically handle the data. Second, multimodal data management. Memory faces massive heterogeneous data, with text, images, audio, video, and business system data mixed together. How to store, organize, compress, and govern it is a complete engineering problem. The larger the data scale, the harder the threshold for this capability becomes. Moreover, since this data must persist long-term and be callable at any time, the requirements for storage structure and governance mechanisms are entirely different engineering systems compared to the ultra-long context during model inference. Third, memory computation. This is the most unique aspect of memory. Retrieval, reranking, updating, reflection, and forgetting—these computational actions are all based on the unit of experience, not the unit of a single piece of data. To give a simple example, a multi-hop retrieval aims to answer the causal chain among a series of conclusions; an isolated fact cannot answer such questions. The memory system must remember this causal chain so that when asked related questions next time, it can retrieve the cause and effect together, rather than just returning an isolated conclusion. This kind of experience-based computation cannot be replaced no matter how strong the model's computing power is. Computing power solves how fast to calculate; memory computation solves whether the causality is remembered. The two are different things. Stacking these three capabilities together constitutes the moat for memory vendors. Only those who can cross the data threshold can gain a foothold in this market.
03 Market Size
As an independent market, the size and growth space of memory are unavoidable issues. Why it can grow so large and how large it can eventually become determine the weight of this track.
3.1 Driving Factors of the Memory Market
The fundamental reason driving the rapid growth of the memory market is that the AI industry is transitioning from a knowledge-driven stage to an experience-driven stage. The core value in the era of large models lies in knowledge production; models learn world knowledge through massive training data and complete knowledge invocation through reasoning capabilities. As model capabilities continue to improve, the cost of knowledge acquisition is dropping rapidly. More and more agents will possess similar knowledge levels and reasoning capabilities, and the gap in model capabilities will gradually narrow.
Against this backdrop, experience begins to become the key factor determining the competitiveness of agents. Knowledge determines the lower limit of capability; experience determines the upper limit. Therefore, experience is becoming the new means of production in the era of agents, and memory becomes the core infrastructure for managing and utilizing this means of production. The transformation of experience into a means of production is first reflected in the evolution of agent capabilities. Currently, most agents still rely on real-time model reasoning to complete tasks. The value of memory lies in transforming each task execution process into long-term experiential assets. Task objectives, execution paths, decision logic, result feedback, and optimization strategies can all be continuously accumulated and reused in subsequent tasks.
As experience continuously accumulates, agent execution efficiency steadily improves, error rates continuously drop, and autonomous optimization capabilities gradually form. Memory thus becomes an important foundation for agents to achieve autonomous evolution. Notably, when agents can directly invoke historical experience to make decisions, it can also significantly reduce the consumption of reasoning tokens. The transformation of experience into a means of production is equally reflected in enterprise AI construction and the competition of consumer-grade AI products. An enterprise's core competitive advantage often comes from experience and know-how. Business experience, industry experience, and decision-making experience have long been held by a few core employees. When employees leave, the experience is lost with them. Experience is difficult to replicate, inherit, and utilize at scale, which has always been a critical issue in enterprise operations.
As digital employees gradually take on more tasks, enterprises are beginning to try to transform organizational experience into inheritable digital assets. The assetization of organizational experience is becoming an important direction for enterprise intelligent upgrading. The transformation of experience into a means of production is also reflected in the changing competitive logic of consumer-grade AI products. As model capabilities tend towards homogenization, users' evaluation criteria for AI products are changing. Users no longer just focus on whether the model is smart, but whether the product truly understands them. The most competitive AI products in the future may not have the strongest models, but they will definitely have the most complete user memory.
3.2 Panoramic Map of the Memory Market
As experience gradually becomes the new means of production in the era of agents, a complete system is forming around the storage, management, and application of experiential assets. According to iFenxi's definition, the memory market mainly consists of three layers: memory hardware, memory operating systems, and memory applications. Among them, memory hardware is responsible for experience storage, memory applications are responsible for experience consumption, and memory operating systems are responsible for experience management and scheduling.
Figure 2: Panoramic Map of the Memory Market
From the perspective of market size, all three layers will benefit from the growth dividends brought by the popularization of agents. However, from the perspective of value distribution, the memory operating system is expected to become the core segment with the fastest future growth and the highest value proportion. Currently, the market's understanding of memory still remains at the storage level.
As agents begin to take on increasingly complex tasks, the true challenge of memory is not storage capacity, but computational capabilities such as management, orchestration, and scheduling. Therefore, memory systems require not only storage capabilities but also computational capabilities. Future memory operating systems need to possess capabilities such as memory extraction, compression, association, orchestration, and reflection. iFenxi believes that the memory operating system will be located at the core of the value chain, connecting downwards to hardware infrastructure on one hand, and serving various applications upwards on the other.
3.3 Market Size Forecast
iFenxi's calculations show that the size of China's memory market will grow from 1.44 billion yuan in 2025 to 64.25 billion yuan in 2030, with a compound growth rate exceeding 110% during the period.
Figure 3: Forecast of China's Memory Market Size from 2025 to 2030
In terms of market distribution, the memory operating system market size will reach 21.85 billion yuan in 2030, accounting for 34% of the overall memory market. Meanwhile, its market size compound growth rate is also the fastest. The five-year compound growth rate of the memory operating system exceeds 152%, while the compound growth rate of the overall memory market in the same period is 114%.
Figure 4: Market Size of Memory Operating Systems from 2025 to 2030
As an independent infrastructure market, iFenxi estimates that by 2030, the proportion of memory in agent infrastructure will approach 26%, becoming the second-largest independent market second only to models.
Figure 5: Proportion of Memory in Agent Infrastructure from 2025 to 2030
04 Competitive Landscape
After researching over 20 vendors, iFenxi categorizes the current participants in China's memory market into five classes. Interestingly, while everyone flies the flag of "memory," they are doing completely different businesses.
4.1 Classification Map of Five Types of Vendors
Cloud giants start from databases, aiming to turn memory into an AI upgrade of storage infrastructure, conveniently migrating their familiar customers along. Model vendors treat memory as an accessory capability to make their models smarter and consume more tokens; whether memory is good depends on whether it can feed the model. Independent memory platforms emphasize not taking sides with any model, elevating memory to an independent data sovereignty asset on par with foundational models, tearing open a market gap through neutrality.
Enterprise memory application vendors work backward from business outcomes, treating memory as a pre-engine for reasoning; whoever helps marketing convert more wins. Consumer memory products are the most radical, simply letting users pay for memory itself, turning backend capabilities into frontend products. These five types of players have different resource endowments and bet on different directions.
Figure 6: Classification Map of Players in the Memory Market
4.2 Analysis of Five Types of Vendors
4.2.1 Cloud Giants
Cloud giants are the most active players in the memory market, especially database teams.
Tencent Cloud Agent Memory is led by the TencentDB database team. Alibaba Cloud is also seeking breakthroughs in memory products on its database product line. Volcengine's open-source project OpenViking at the beginning of the year also originated from the VikingDB vector database team. The core starting point of cloud giants is consistent: memory is a natural extension of database capabilities. Luo Yun, Deputy General Manager of Tencent Cloud Database, judges that future databases need to manage agent organizations. Permission granting between main agents and sub-agents, active-standby switching, and multi-version data management—these were once marginal database requirements and are now core propositions.
Ge Shuangbo, Senior Product Expert of Alibaba Cloud Database, makes a more vivid judgment: models determine how smart AI is, while databases determine how far AI can run. The upper limit of AI is determined by models, and the lower limit is determined by databases. iFenxi believes that not only cloud giants but current database vendors are expected to define themselves as AI memory companies in the future. From the perspective of database vendors, memory is essentially a data access system in the AI era.
From a competitive dimension, the advantages of cloud giants in the memory track also extend from databases. Their underlying vector databases and massive data storage experience are accumulations that other vendors find hard to catch up with in the short term. Meanwhile, cloud giants' databases possess mature enterprise-grade security and compliance systems, reducing compliance concerns during enterprise procurement. However, cloud giants also face two core constraints when building memory.
First, their perspective naturally leans towards storage and retrieval, focusing on what to remember, how to store, and how to retrieve, while paying insufficient attention to the full lifecycle management of memory, including when to remember, forget, merge, overwrite, and how to iteratively evolve.
Second, insufficient resource investment, with a risk of becoming a marginal endeavor. The memory teams of cloud giants are usually sub-projects of the database department, with low staffing and strategic priority, which does not match the status of memory in future AI Infra and is highly likely to become a white elephant. Cloud giants currently have the highest degree of productization in the memory market and the easiest category form for customers to understand and procure, but the core problem they solve is the AI upgrade of storage infrastructure, not promoting agents to better accumulate experience and self-evolve.
4.2.2 Model Vendors
The starting point for model vendors to build memory is completely different from cloud giants; their goal is to make model capabilities stronger and token consumption higher. More importantly, model vendors all regard continuous learning as the next leap capability point, and memory is the essential foundation for achieving continuous learning. For example, when planning GLM-5.2, Zhipu AI listed memory as a key technology to break through.
In an internal letter at the beginning of 2026, Tang Jie, founder of Zhipu AI, explicitly listed memory as a core proposition for the next-generation models. The Transformer architecture, widely used for nearly 10 years, is showing deficiencies, including the computational overhead of ultra-long contexts, memory mechanisms, and update mechanisms. The core logic of the model vendor path is to define memory as an accessory enhancement capability of the model. Its memory architecture design naturally leans towards its own model ecosystem, and the value of memory is reflected through the improvement of model or agent effects. Therefore, in the cognition of model vendors, memory should make their own models more usable, and memory itself does not exist independently of the model.
From a competitive dimension, the advantages of model vendors come from the spillover momentum of the main battlefield of models, with the data flywheel being particularly prominent. During the process of users calling models, a large amount of data that can be used to improve memory capabilities is generated, and this data will be used for the next generation of memory architecture tuning. However, security and compliance are the most obvious soft spots for model vendors.
Their governance systems require user data to be closed-loop and flow back to serve their own models. This practice of keeping all private data tightly in their own hands naturally steps on the gray area of compliance. iFenxi judges that model vendors have an unsolvable knot when building memory: the memory system is bound to its own model ecosystem from the very beginning, leaving almost no room to use it in other models or applications.
4.2.3 Independent Memory Platforms
Independent memory platforms start from a different mindset. They do not view memory as an accessory function of a specific database or model, but define it as an independent AI Infra on par with foundational models. The core propositions of this path can be summarized into three points.
First, the neutrality of memory itself is its greatest value. The boundaries of models are expanding infinitely, unrestrainedly absorbing the knowledge of all humanity, personal experiences, and enterprise know-how, causing huge disruption to the business world. The neutrality of memory makes it the strongest symbol of data sovereignty in the AI era, preventing individuals and organizations from being distilled into waste by models. Fan Chenggong, co-founder of MemVerge, stated that both individuals and enterprises should have their own memory systems.
Second, memory is the pioneer of self-evolution in the AI era. Zhan Chaoqun, CEO of Zhibian Technology, stated in an exchange that memory is the infrastructure for agent evolution. Memory receives new data every day, self-evolving through continuous reflection, verification, forgetting, and conflict management. Compared to the data flywheel of models, the data flywheel of independent memory platforms is more perfect because its evolution process occurs on the side of terminal individuals and organizations, completed spontaneously without human intervention.
Third, independent memory platforms have strong compatibility and can be combined with any model and any application. This architectural freedom is something other types of players cannot provide.
From a competitive dimension, the advantages of independent platforms are reflected in different aspects. First, independent memory platforms are all training small models focused on memory. Referring to the human brain, small-parameter models are more adaptable to the memory system. On one hand, the response speed must be fast when reasoning calls memory, and large-parameter models will cause delays. On the other hand, from the perspective of data sovereignty, the future trend of memory is edge or localized models. Small-parameter models have low computing power requirements and are more aligned with the needs.
Second, in terms of product capabilities, independent platforms cover the full lifecycle of memory. Taking Zhibian Technology as an example, its three-layer architecture of workspace, project, and actor realizes the collaboration of organizational and individual memory. Within an enterprise, the data of the legal department, R&D department, and operations team are accumulated in different projects but collaboratively evolve under the same workspace. This organizational memory stacking system is currently rarely seen in similar designs in other memory products.
4.2.4 Enterprise Memory Applications
Enterprise memory application vendors have taken a more direct path, defining memory backward from business outcomes. Their core starting point is that better reasoning effects come from memory optimization, which can completely subvert traditional enterprise application software. By calling memory first and then the model, execution effects are greatly improved, and token consumption is significantly reduced.
Wen Deliang from Red Bear AI stated in an interview that this industry will ultimately not be about who has the best technology, but who has the best commercialization. The core advantage of this path lies in the most direct commercial validation. Enterprise clients do not care about the technical architecture of memory; they only care about whether the reply effect of the memory-driven customer service robot has improved and whether the marketing conversion rate has increased, taking business outcomes as the sole test standard. Currently, scenarios such as customer service, marketing, and education are taking the lead in landing enterprise memory applications.
iFenxi judges that the limitation of this path lies in the generalizability of scenarios. Forcing memory capability construction through vertical scenarios may achieve the extreme in a single scenario, but lacks cross-scenario abstraction and reusability. Whether the memory-driven reasoning logic verified in the customer service scenario can be migrated to scenarios like financial risk control, industrial O&M, and medical diagnosis remains unproven.
4.2.5 Consumer Memory Products
Consumer memory products have taken a more radical path, not letting memory hide as backend infrastructure, but making it a frontend product that users are willing to pay for directly. Their core starting point is an intuitive judgment: a person is largely composed of their accumulated memories, and paying for memory is the most natural demand of consumers. Consumer memory products mainly come in two forms: software and hardware.
The core concept of the software form is to uniformly flow back the dialogues and agent trajectories in various AI applications to a memory center. When a user initiates a request, the context is rewritten based on memory and distributed to the appropriate AI application, ultimately delivering the optimal result. The hardware form of consumer memory products is always-on audio or video wearable devices. By recording users' all-weather audio and video data, they become the spiritual companions who understand users best.
Obviously, the advantage of consumer memory applications lies in user stickiness. When users accumulate all their cross-platform, cross-model AI interaction histories into a memory software or hardware, the migration cost will rise exponentially over time. iFenxi judges that consumer memory hardware products have greater growth space in the long-term perspective because their memory comes from the physical world rather than the digital world, which is the most scarce data asset at present. However, hardware products rely not only on the development of memory technology itself but are also constrained by the hardware itself, such as the battery life of always-on video devices and the edge multimodal data processing capabilities.
4.3 Competitive Quadrant
After completing the type classification of market participants, it is necessary to further evaluate the competitive posture of each vendor. This chapter uses a two-dimensional competitive quadrant of technical capability and market influence to evaluate the main players one by one. The construction of this evaluation framework is based on iFenxi's continuous tracking research on China's memory market.
The research team investigated over 20 vendors, systematically collecting data on their technical architecture, product maturity, scenario landing, and commercialization, and establishing evaluation dimensions and weights accordingly. The horizontal axis of the competitive quadrant is technical capability, weighted and synthesized by four sub-dimensions: multimodal (20%), self-evolution (30%), product (25%), and governance (25%). The vertical axis is market influence, equally weighted and synthesized by two sub-dimensions: scenario (50%) and commercialization (50%).
Figure 7: Competitive Quadrant of China's Memory Market
Based on the scores of the two dimensions, vendors fall into four quadrants, corresponding to different competitive postures. Leaders rank at the forefront in both dimensions, leading in technical capability and having achieved scaled commercial landing, constituting the current benchmark reference for the market. Innovators have outstanding technical capabilities but limited market influence, mostly forming advantages in specific directions such as long context, self-evolution, or memory retrieval. Their commercialization and scenario coverage are still in the early stages, with the potential to migrate to Leaders.
Practitioners have strong market influence but relatively weak technical capabilities. They usually rely on the customer base of existing cloud or large model businesses to extend memory capabilities. Their underlying technical accumulation is shallow or relies on external solutions. They have a share advantage in the short term but face pressure for technical reinforcement in the medium to long term. Explorers are in the early stages of both dimensions, mainly consisting of startup teams and open-source projects. Their productization and commercialization are not yet complete, representing potential variables in the future market landscape. iFenxi believes that the endgame of the memory market must be deeply integrated with reasoning. The reasoning here is not only logical reasoning but also includes dimensions such as emotion. For example, memory plus logical reasoning represents better work outcomes, while memory plus emotional reasoning represents a better life state.
Therefore, the good or bad of memory technology and products cannot be judged merely by comparing memory recall effects or token consumption, but should be measured around the success rate of task execution. When the benchmark of the memory market shifts to a comparison of task effects, it will be the true beginning of memory competition.
4.4 Evaluation of Key Vendors
4.4.1 Tencent Cloud, Representative of Cloud Giants
Tencent Cloud is the representative of the cloud giant path; its strengths and shortcomings both come from the same origin: the database.
In terms of product advantages, the Agent Memory product is independently developed from the bottom up by the Tencent Cloud database team. Built on Tencent Cloud Vector Database, it is an independent memory management foundation with complete APIs and consoles, integrated with CodeBuddy, allowing enterprises to use it out of the box.
Technically, it constructs a four-layer memory architecture: L0 stores raw sessions, L1 breaks them down into atomic information and builds vector indexes, L2 aggregates them into scenario-based files, and L3 precipitates them into personalized profiles. Luo Yun, Deputy General Manager of Tencent Cloud Database, explained it this way: L3 is mandatory disclosure directly implanted into the context, L2 adopts progressive disclosure for on-demand reading, and L0 and L1 are retrieved through vectors and keywords.
This layered design enabled it to pull long-term memory accuracy from 48% to 76% and user fact recall rate from less than 30% to over 79% in the PersonaMem evaluation. In addition, relying on the ecosystem of Tencent's multimodal large models, Tencent Cloud's memory products have strong multimodal basic parsing capabilities.
In terms of governance and compliance, Tencent Cloud has a mature enterprise-grade multi-tenant system, and compliance capabilities are readily available, reducing concerns during enterprise procurement. In terms of scenarios, consumer-side WorkBuddy plus enterprise-side agents can be rolled out in multiple scenarios. The disadvantages of Tencent Cloud's memory products are concentrated on low priority and insufficient resources. The memory product team is a sub-project of the database department, with low staffing and strategic priority, suspected of being a white elephant. For Tencent Cloud, memory is just an incremental product, not a core strategy, and the continuity and intensity of resource investment are questionable.
4.4.2 Kimi, Representative of Model Vendors
Kimi is the representative of the model vendor path. Its advantages come from its technical heritage in long texts and consumer-side product experience, while its shortcomings are reflected in multimodality and product positioning. Let's talk about the advantages first. Kimi's confidence in building memory lies in its long-term accumulation in long contexts. Founder Yang Zhilin stated that context length is the memory of large models and one of the key factors determining large model applications. Kimi is not only the world's first intelligent assistant to support 200,000 Chinese characters input but has also opened internal testing for 2 million characters lossless context.
At the product level, the Kimi Memory Space offers a good, transparent, and controllable experience. Users can clearly see what is stored in the memory, why it is stored, and how it can be modified. This transparency is rare in memory products. In terms of governance, Kimi has made differentiation: users manage their memory completely autonomously, and data is not used for model training, alleviating the compliance soft spot most criticized for model vendors. In terms of self-evolution mechanisms, Kimi uses a separate memory model to judge what to remember and already has a set of update logic. Now let's talk about the disadvantages.
At the multimodal level, Kimi's memory is currently pure text and cannot process heterogeneous data such as images, audio, and video, limiting the scenarios it can serve. In terms of product positioning, memory is not an independent product but just a function of Kimi, mainly focusing on consumer-side personal assistant scenarios. It has no independent commercialization and relies on the membership system and API for monetization.
4.4.3 Zhibian Technology, Representative of Independent Memory Platforms
Zhibian Technology is a representative vendor of independent memory platforms, positioned as a neutral enterprise-grade productivity memory infrastructure. Its capabilities can be understood from five dimensions.
First is product philosophy. Zhibian Technology believes that memory has compound interest. In the framework of Productivity Intelligence = Thinking Power × Memory Power, foundational models provide general thinking power, while memory provides sustainably accumulated experience, decisions, and states. The two are a multiplication relationship, not a substitution. Zhibian Technology advocates the "compound interest" of memory, allowing the results of each task to precipitate into reusable experience, thereby shortening the learning path for the next task. This also implies the neutrality premise of the independent memory layer: models can be iterated, switched, and combined, but enterprise memory should be reused across models, agents, and cycles, belonging to the customers themselves. This is the fundamental difference between independent memory platforms and the built-in memory of giants.
Second is organizational memory stacking. Giants and open-source solutions mostly do not distinguish between organizational and personal memory. Zhibian Technology is one of the few independent vendors that has turned organizational memory governance into a product. Zhibian's organizational memory architecture revolves around the three-layer structure of Workspace (project container), Project (memory asset), and Actor (subject of ownership), allowing the same piece of knowledge to precipitate different memories according to roles. For example, for the same product document, after-sales focuses on troubleshooting, R&D focuses on version iteration, and management focuses on GTM. Different roles take what they need, forming memory collaboration at the company level.
Third is measuring memory by "task success rate" rather than "how much is remembered." Mainstream evaluations like LoCoMo and LongMemEval test whether it can remember, but memory ultimately serves reasoning, and it should test whether it can complete tasks relying on memory. The new evaluation standard, MemoryArena, advances the evaluation to interactive scenarios where sub-tasks are interdependent.
In the controlled comparison of the same model and the same evaluation script in MemoryArena, Zhibian Technology's equal-weight average task success rate across five domains is 20.5%, higher than the strongest control group's 13.5%, with the most obvious advantage in the physical reasoning domain (60% vs 45%). A more telling detail appears in the progressive multi-hop retrieval domain. The sub-Q&A accuracy of Mem0, which is mainly based on similarity retrieval, is higher (8.9% vs 6.7%), but the final task success rate is lower (15% vs 20%). This shows that "remembering more" does not equal "getting it done." Fourth is that traceable and governable memory conforms to enterprise governance.
The advantage of Zhibian Technology's governance layer comes from the team's genes: the core team comes from Alibaba Cloud Database and has long served enterprise-side clients. Open-source projects generally do not touch governance, and giant memory leans more towards consumer-side privacy. Zhibian Technology's differentiation comes from enterprise-side accumulation.
Specific governance capabilities are reflected in several aspects: First, citation chain traceability. Retrieval comes with source Projects and fragment-level corner marks, and the decision trajectory is fully traceable. Because scenarios like finance and compliance require not only conclusions but also the ability to restore how the conclusion was derived. Second, a permission model dynamically calculated based on identity, task, and context, covering multiple dimensions from team, space, project, role to entity level. Third, lineage-based cascade deletion. A deleted fact has often spawned multiple copies such as indexes, summaries, rules, and caches. Zhibian tracks by lineage and cascades to clear them, deleting the source data while deleting its impact, which is particularly critical for compliance audits. Fifth is multimodal data parsing, which determines the upper limit of recall accuracy.
Multimodal parsing is the prerequisite for recall accuracy; if parsing is wrong, subsequent memory and reasoning will be wrong. Zhibian Technology's judgment is that the hardest part of document understanding is not OCR, but restoring 2D, hierarchical layouts into memory without losing structure, relationships, original values, and sources. It takes the route of self-developed parsing models. The barrier is not the size of model parameters, but the long-term accumulated format adaptation and data engineering. For example, processing complex layouts such as nested tables, cross-page tables, and merged cells in PDFs, as well as modal coverage of text, images, videos, tables, PDFs, code, and databases, which are the main forms of enterprise knowledge assets.
05 Future Trends
The integration of memory and reasoning is the endgame. In the coming years, memory will evolve in four directions: decision precipitation, Memory as a Service, edge data sovereignty, and market standardization.
5.1 From Experience Precipitation to Decision Precipitation
To better integrate with reasoning, the next step for the memory market is not only to precipitate experience but also decision logic. The traditional enterprise information system follows the "data-knowledge-decision" chain, first organizing data into knowledge, and then using knowledge to support decisions. In the era of agents, this chain evolves into "knowledge-context-memory-execution." Memory replaces the intermediate knowledge precipitation link and directly serves execution. The bigger change is behind. Enterprises used to precipitate knowledge, now they precipitate experience, and in the future, they will precipitate decisions. Decision logic, behavioral patterns, and organizational experience will replace knowledge as the most important assets of an enterprise. Because agents not only need to know what the answer is but also need to understand why to do it this way and under what circumstances. For example, when a senior employee leaves, what they take away is often not knowledge, but the set of judgment basis for making decisions: why choose this plan, why reject that client. Precipitating decisions is leaving this set of judgment basis to the organization, so it does not leave with individuals. Experience precipitation is happening right now; decision precipitation is the increment that will arrive on a large scale in the next stage. When memory can not only remember what was done but also remember why it was done this way, memory upgrades from a recording tool to a decision asset.
5.2 Memory as a Service
Memory is transforming from an internal module of an agent into an independent cloud service, a trend that can be called Memory as a Service. Commercially, the billing model is also changing accordingly. Past memory costs were bundled into the token consumption of model reasoning; in the future, it will shift to an independent billing method of memory storage plus retrieval times. Memory changes from a cost item to a commodity that can be priced and charged independently, which is itself a signal of its independence as a market. Technically, a unified memory layer across agents is becoming the direction. The same user shares the same memory across different agents. A user develops preferences in one application, and another application can continue to use them without having to teach it from scratch. Once the memory layer is independent, it corroborates the core judgment of this report: memory is an independent market, not an accessory to any model.
5.3 Edge and Local Memory, Data Sovereignty and Privacy
Regulations are tightening, and memory is moving towards the edge and local environments. The EU's GDPR and China's Personal Information Protection Law are both reinforcing a principle: personal data must be manageable, controllable, and deletable. What is stored in memory is the most core experience of a person or an enterprise. Once this data goes to the cloud, there is a risk of being processed and utilized by third parties. The stricter the regulation, the more enterprises tend to deploy memory locally, keeping data off devices and within enterprise boundaries. This trend has already started to land on the edge. Some memory vendors have begun to integrate memory capabilities into edge devices such as mobile phones and home appliances, storing memory locally and only sending out necessary results. Behind edge and local memory is data sovereignty. A person or an enterprise has complete control over its own experience and does not hand it over to any platform. The further memory moves towards the edge, the higher the neutral value of independent memory platforms.
5.4 The Memory Market Moves Towards Standardization
For the memory market to truly grow up, the final step is standardization. Evaluation systems, memory protocols, memory interfaces, and industry benchmarks are gradually taking shape. Before unified standards, each vendor's memory has its own format and interface. Once an enterprise uses a certain vendor, it is locked in, and the cost of switching vendors is extremely high. The community is promoting unified memory protocols, allowing memory modules of different frameworks to interoperate. An application can call memory from different vendors, and one vendor's memory can be reused by multiple applications.
Once standards are established, memory will no longer be the private technology of a single vendor but a layer of public infrastructure. By then, memory is expected to become the new infrastructure layer in the AI era after foundational models. This also means that whoever takes the lead in standard-setting will seize the initiative in the second half of the memory market.
Scan the code to get the "2026 iFenxi · China Memory Market Research Report"