In the first half of 2026, memory evolved from a technical concept into a dedicated industry track. In late May, Tencent Cloud launched Agent Memory, boosting long-term memory accuracy from 48% to 76%. In July, Alibaba Cloud intensively deployed Agent memory capabilities across its database product line. Meanwhile, the Viking team at Volcengine open-sourced the OpenViking project, reducing Token consumption costs by 91% and surpassing 21,000 GitHub stars.
Financing in the memory industry is accelerating in parallel. In July, Memory Tensor completed a hundred-million-yuan Pre-A funding round, primarily backed by angel investors, including Huawei Hubble Investment and Honor Strategic Investment. During the same period, RedBear AI closed a multi-hundred-million-yuan A+ funding round, reaching a post-investment valuation of nearly 3 billion yuan. Capital is now voting with real money for the memory track.
According to estimates by iFenxi, the memory market size will reach 7.85 billion yuan in 2027 and grow to 64.25 billion yuan by 2030, with a compound annual growth rate (CAGR) exceeding 110%. Such a substantial market scale has prompted various players to rush into the arena. So, which vendors are actually participating in this memory competition?
01 All Vendors Converge on "Memory": Five Player Groups Drive the Emerging Track
After surveying over 20 vendors, iFenxi found that current participants in China's memory market can be categorized into five types. Interestingly, while everyone flies the "memory" banner, they are actually pursuing completely different business models.
Cloud giants start from databases, aiming to turn memory into an AI upgrade of storage infrastructure while seamlessly migrating their existing customers. Foundation model vendors treat memory as an auxiliary capability to make their models smarter and consume more Tokens; the usability of memory depends on whether it can fully feed the model. Independent memory platforms emphasize neutrality toward any model, elevating memory to an independent data sovereignty asset on par with foundation models, leveraging neutrality to carve out a market niche.
Enterprise memory application vendors work backward from business outcomes, treating memory as a pre-engine for reasoning; whoever helps marketing achieve more conversions wins. Consumer memory products are the most aggressive, simply asking users to pay for memory itself and turning back-end capabilities into front-end commodities.
These five types of players possess different resource endowments and bet on different directions, fundamentally driven by their inherent positions. Some believe memory will be the first to ignite the developer ecosystem and become Agent infrastructure; others argue memory inherently possesses B2B enterprise attributes and must focus on commercialization in the enterprise market; still others contend that memory-driven consumer hardware will become a phenomenal product in 2027, with consumers paying directly for it.
Figure 1: Classification Map of Players in the Memory Market
02 Cloud Giants: Memory Products Grow from Database Teams, Memory Defined as the Next-Generation Database
Cloud giants are the most active players in the memory market, particularly the database teams. Tencent Cloud's Agent Memory is led by the TencentDB database team. Alibaba Cloud is also seeking breakthroughs in memory products within its database product line, while Volcengine's open-sourced project OpenViking earlier this year also originated from the VikingDB vector database team.
The core starting point for 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 will need to manage Agent organizations. Permission granting between primary and sub-Agents, primary-standby switching, and multi-version data management—these were once marginal requirements for databases but have now become core propositions.
Ge Shuangbo, Senior Product Expert at Alibaba Cloud Database, offers a more vivid judgment: models determine how smart AI is, while databases determine how far AI can go; the upper limit of AI is determined by models, and the lower limit by databases. iFenxi believes that not only cloud giants but current database vendors will likely 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 perspective, the advantages of cloud giants in the memory track also extend from databases. Their underlying vector databases and massive data storage experience represent accumulations that other vendors can hardly 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, iFenxi believes that cloud giants face two core constraints in building memory.
First, their perspective naturally leans toward 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 risks making it a marginal endeavor. Cloud giants' memory teams are typically sub-projects within the database department, with low staffing and strategic priority. This misaligns with the future status of memory in AI Infra and is highly likely to render it a marginal endeavor.
Cloud giants currently represent the most productized and easily understood and procured category in the memory market. However, the core problem they solve is the AI upgrade of storage infrastructure, rather than driving Agents to better accumulate experience and self-evolve.
03 Foundation Model Vendors: Making Memory Serve the Model Ecosystem, Memory as an Enhanced Capability Rather Than an Independent Product
The starting point for foundation model vendors to build memory is entirely different from that of cloud giants; their goal is to make model capabilities stronger and consume more Tokens. More importantly, foundation model vendors all regard continuous learning as the next leap in capability, and memory is the essential foundation for achieving continuous learning. For instance, when planning GLM-5.2, Zhipu AI identified memory as a key technology for breakthrough. In an internal letter in early 2026, Tang Jie, founder of Zhipu AI, explicitly listed memory as a core proposition for next-generation models, stating, "The Transformer architecture, widely used for nearly 10 years, is showing shortcomings, including the computational overhead of ultra-long contexts, memory mechanisms, and update mechanisms." The core logic of this path for foundation model vendors is to define memory as an auxiliary enhanced capability of the model. Its memory architecture design naturally leans toward its own model ecosystem, and the value of memory is reflected through the improvement of model or Agent performance.
Therefore, in the understanding of foundation model vendors, memory must make their own models more usable, and memory itself does not exist independently of the model. From a competitive perspective, the advantage of foundation model vendors comes from the momentum spillover of the main model battlefield, with the data flywheel being particularly prominent. During the process of users invoking models, a massive amount of data is generated that can be used to enhance memory capabilities, and this data will be used for tuning the next-generation memory architecture. However, security and compliance are the most obvious soft spots for foundation model vendors. Their governance systems require user data to flow back in a closed loop and serve their own models. This practice of keeping all private data tightly in their own hands naturally steps into the gray area of compliance.
iFenxi judges that foundation model vendors face an unsolvable knot in building memory: the memory system is tied to their own model ecosystem from the very beginning, leaving almost no room to apply it to other vendors' models or applications.
04 Independent Memory Platforms: Pioneering Self-Evolution
Independent memory platforms take a different approach. They do not view memory as an auxiliary function of a specific database or model, but define it as an independent AI Infra on par with foundation models. The core propositions of this path can be summarized in three points.
First, the neutrality of memory itself is its greatest value. The boundaries of models are expanding infinitely, irresistibly absorbing the knowledge of all humanity, personal experiences, and enterprise know-how, causing massive 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 by models into "waste." Fan Chenggong, co-founder of MemVerge, stated, "Whether individuals or enterprises, everyone should have their own memory system."
Second, memory is the pioneer of self-evolution in the AI era. Zhan Chaoqun, CEO of Zhibian Technology, stated in an exchange, "Memory is the infrastructure for the evolution of intelligent agents." Memory receives new data every day and self-evolves 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 occurs on the side of end-users and organizations, completing 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 perspective, the advantages of independent platforms are manifested in different dimensions. First, independent memory platforms are all training small models focused specifically on memory. Referencing the human brain, small-parameter models are more suited to the memory system. On the one hand, response speeds must be fast when reasoning invokes memory, and large-parameter models would cause latency.
On the other hand, from the perspective of data sovereignty, the future trend of memory is edge-side or localized models. Small-parameter models have low computing power requirements and are more aligned with these 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 achieves the synergy between organizational and individual memory.
Within an enterprise, data from the legal department, R&D department, and operations team is accumulated in different projects but evolves collaboratively under the same workspace. This organizational memory stacking system currently has almost no similar design in other memory products.
05 Enterprise Memory Application Vendors: Driving Memory Capabilities from Business Outcomes
Enterprise memory application vendors have taken a more direct path, defining memory in reverse from business outcomes. Their core starting point is that better reasoning results come from memory optimization, which can completely disrupt traditional enterprise application software. By invoking memory before invoking the model, execution performance is significantly improved, and Token consumption is also markedly reduced. Wen Deliang of RedBear AI stated in an interview, "Ultimately, this industry will not be about who has the best technology, but about who commercializes the best."
The core advantage of this path lies precisely in the most direct commercial validation. Enterprise customers do not care about the technical architecture of memory; they only care about whether the reply performance of memory-driven customer service robots has improved and whether marketing conversion rates have increased, taking business outcomes as the sole testing 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. Driving memory capability construction with vertical scenarios may achieve perfection in a single scenario, but it lacks cross-scenario abstraction and reusability. Whether the memory-driven reasoning logic validated in customer service scenarios can be transferred to financial risk control, industrial operation and maintenance, medical diagnosis, and other scenarios remains unproven.
06 Consumer Memory Products: Waiting for a Blockbuster to Emerge
Consumer memory products have taken a more aggressive path, not letting memory hide as back-end infrastructure, but making it a front-end product that users are willing to pay for directly. Their core starting point is an intuitive judgment: a person is largely accumulated by their memories, and paying for memory is the most natural demand for consumers. Consumer memory products mainly come in two forms: software and hardware.
The core concept of the software form is to uniformly route back dialogues and Agent trajectories from various AI applications into 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 consists of always-on audio or video wearable devices. By recording users' round-the-clock audio and video data, they become spiritual companions who understand users best.
Obviously, the advantage of consumer memory applications lies in user stickiness. When users accumulate all their cross-platform and cross-model AI interaction histories into a single memory software or hardware, the migration cost will rise exponentially over time. iFenxi judges that consumer memory hardware products have greater room for growth from a long-term perspective because their memory comes from the physical world rather than the digital world, making it the most scarce data asset today. However, hardware products rely not only on the development of memory technology itself but are also constrained by the hardware itself. For example, the battery life of always-on video devices and edge-side multimodal data processing capabilities.
07 AI Memory: Where Will It Ultimately Ignite?
The five types of players have each bet on different explosive points, but there is currently no clear trend on where memory will ultimately ignite. Different ignition points determine the path choices of different types of vendors. iFenxi believes that the endgame of the memory market must be deeply integrated with reasoning. The reasoning here is not just logical reasoning but also includes dimensions such as emotion. For example, "memory + logical reasoning" represents better work outcomes, while "memory + emotional reasoning" represents a better state of life.
Therefore, the quality of memory technology and products cannot be judged solely by comparing memory recall performance and Token consumption, but should be measured around the success rate of task execution. Only when the benchmark of the memory market shifts to a competition of task outcomes will the memory competition truly begin.
iFenxi is about to release the "2026 iFenxi China Memory Market Research Report." Based on systematic research on over 20 memory-related vendors, it will comprehensively present the performance of each vendor in the competitive quadrant. Stay tuned for continuous updates.