RedBear AI Completes Hundreds of Millions of Yuan Series A+ Financing, Extends from To B Services to To C Applications Based on AI "Memory Science" | 36Kr Exclusive

36Kr learned today that RedBear AI has officially closed a Series A+ funding round of hundreds of millions of RMB, pushing its post-money valuation to nearly 3 billion RMB. This round was co-led by Zhejiang Jiuwei Private Equity Fund, Jiaxing Zhangyuan Venture Capital, and existing shareholder Graceful Summit Capital. This marks the sixth financing round RedBear AI has completed in just 15 months.

According to sources, the proceeds from this funding will be primarily used to further advance human-brain-inspired foundational research on AI memory science, accelerate the deep integration of the OpenBear general large model and the MemoryBear memory science system, and expand market coverage across four core application scenarios: intelligent customer service, intelligent marketing, ChatBI, and smart education.

Following the early-stage race focused on scaling large model parameter counts and computing power, the industry is gradually entering a deep commercialization phase where "applications define success". Currently, technologies centered on the "Memory Layer" dimension of large models have become a key strategic priority.

Financing events in this field have been increasingly frequent and notably accelerating: Engram, a 13-person team with Andrej Karpathy—who led Claude pre-training at Anthropic—among its investors and a $600 million valuation, has publicly debuted; Clipto, positioned to connect personal data to the agent ecosystem's memory layer, has received multiple follow-on investments from HSG, GL Ventures, and other institutions; domestic startups including "Memory Tensor", which completed a nearly 100 million RMB angel round to develop the open-source memory operating system MemOS, and MemoraX AI, which secured over 100 million RMB in total across seed and seed+ rounds in spring 2026, have all built up sufficient capital reserves in this technical direction.

Since its founding in April 2024, RedBear AI has established the technical roadmap of "Memory-Driven AGI". Closely integrating technological innovation with the actual industrial demands of enterprises, while breaking through the memory bottlenecks that hinder the real-world deployment of traditional foundational large models, is the core proposition RedBear AI aims to address in this wave of AI advancement.

I. Why does AGI need a dynamic "Hippocampus"?

Over the past two years, the industry has widely prioritized "model training over practical application deployment". Massive computing resources and capital have been poured into pre-training foundational large models, yet when these models are rolled out for enterprise use cases, they frequently encounter multiple practical bottlenecks.

Restricted by their underlying architecture, traditional foundational large models commonly suffer from discontinuous conversations or context forgetting in long dialogue scenarios due to token limitations; meanwhile, their high knowledge forgetting rate undermines the consistency and stability of service delivery. Additionally, token consumption of large models typically grows linearly, directly driving up the long-term computing costs for enterprises running AI systems, and the lack of causal logic verification makes models prone to "hallucinations", compromising business reliability.

The RedBear AI team believes that simply enhancing the capabilities of foundational large models is insufficient to solve real-world enterprise problems. Targeting these common industry pain points, RedBear AI first proposed the concept of "AI Memory Science" in 2024.

To achieve genuine cognitive intelligence and support large-scale enterprise deployment, large models must evolve their information processing approach from "static information accumulation" to "dynamic organism-style management". In 2026, RedBear AI achieved a critical milestone in its technical ecosystem: it implemented a second round of deep integration between the OpenBear large model and its self-developed MemoryBear memory science system, on top of which it built a native Agent architecture.

This technical logic that deeply combines the "cerebral cortex (responsible for reasoning)" and the "hippocampus (responsible for memory)" is pushing enterprise-grade AI applications to evolve from the passive "question-and-answer" tool state to a new era of "memory-driven reasoning".

II. RedBear AI's Technology Map: The Triangular Matrix of OpenBear, MemoryBear, and Native Agent

The core capability built by RedBear AI is a closed-loop ecosystem where the OpenBear general large model, MemoryBear memory science system, and native Agent architecture are interdependent.

1. OpenBear General Large Model: Underlying Capability Base Designed for Memory

OpenBear is a new-generation AGI model independently developed and pre-trained by RedBear AI. It adopts the MoE (Mixture of Experts) sparse architecture at its core, with a parameter scale reaching the hundred-trillion level.

From a design philosophy perspective, OpenBear is among the first pre-trained large models in the industry built around the "engineered for memory" principle. At the initial stage of model architecture design, RedBear AI embedded dedicated memory interfaces and attention mechanism optimization modules inside the model, enabling it to conduct high-frequency, low-latency information interaction with the external MemoryBear memory science system, and realizing seamless integration of model reasoning capabilities and memory capabilities. This large model system emphasizes cross-domain generalization, long-term memory, and self-reflection features, and can understand a wide range of professional complex real-world tasks without dedicated customized optimization.

OpenBear Product Interface

2. MemoryBear Memory Science System: Hierarchical Dynamic Memory and Brain-inspired Mechanisms

If the large model itself is responsible for logical reasoning, MemoryBear serves as its memory hub. Drawing on the cognitive mechanisms of the human brain, MemoryBear integrates technical theories including the ACT-R dual-memory architecture, the Ebbinghaus Forgetting Curve and activation model, intelligent semantic pruning algorithms, and a 3D self-reflection engine.

Structurally, MemoryBear has established a hierarchical dynamic system spanning from instant to permanent memory, including a working memory layer, a short-term memory layer, and a long-term memory layer. To support the efficient operation of this hierarchical architecture, MemoryBear has overcome multiple core underlying technologies:

Temporal Memory Technology: Enables AI to identify the event timestamps embedded in information and construct a complete timeline, endowing AI with a sense of historical context.

Dynamic Semantic Network Technology: Automatically extracts key entities and core viewpoints from massive long texts or unstructured data, building a dynamic knowledge graph.

Intelligent Semantic Pruning Technology: Based on the Ebbinghaus Forgetting Curve principle, it automatically identifies and periodically deletes redundant, outdated, or incorrect information, maintaining the accuracy of the memory library and reducing token consumption during large model operations.

3D Self-Reflection Engine: When the system enters a low-load period, it automatically re-verifies existing memories and corrects potential errors, thereby mitigating the hallucination flaw inherent to large models.

Minimized Memory Sharing Technology: Establishes a unified memory hub in multi-Agent collaboration scenarios, where each agent only retains critical memories required to complete its own tasks, and other agents access these memories on demand, reducing overall computing power consumption.

According to technical metrics disclosed by RedBear AI, with the support of MemoryBear, the knowledge forgetting rate of its foundational large model is controlled below 1%; token consumption is reduced by 25 times; industry ambiguity preprocessing rate is kept below 1%; and the model hallucination rate is lowered to approximately 0.2%.

MemoryBear Product Interface

3. Native Agent Architecture: Evolution from Passive Tools to Active Collaboration

At the intersection of the OpenBear large model and the MemoryBear memory system, RedBear AI has encapsulated its native Agent architecture. The core feature of this architecture is "memory-native": the Agent system is built with inherent memory capabilities from its inception, allowing it to retain detailed records of every interaction with specific users and the final execution results of tasks, which serve as input for subsequent optimization. Furthermore, this architecture supports autonomous planning and multi-Agent collaboration, enabling agents to implement dynamic task execution and automatic resumption under a secure and controllable permission management system.

Additionally, built on the technical foundation of the MemoryBear memory system, RedBear AI has further extended memory capabilities to the developer tool domain, launching CodeBear, a memory-enabled programming large model platform. Unlike traditional AI code assistance tools, the core differentiation of CodeBear lies in "memory-driven development". In real-world development environments, CodeBear can independently understand requirements, plan technical solutions, write code, and run validations. Leveraging the MemoryBear hierarchical memory system, it continuously accumulates project experience, coding standards, and business logic from every interaction, forming reusable personalized development knowledge deposits. This effectively addresses the pain points of broken context and high redundant communication costs in traditional programming tools, significantly improving the development efficiency and consistency of complex projects.

III. Breakdown of Four Application Scenarios: A Record of Large Model Application Deployment

Built on the aforementioned "Model + Memory + Agent" technical ecosystem, RedBear AI has currently launched four core application products in the market, aiming to verify the commercial value of AI memory science across different niche tracks.

1. Intelligent Customer Service: Personalized Retention in Long-cycle Interactions

The primary bottleneck of traditional intelligent customer service systems is the "one case, one closure" model, which lacks the ability to retain customers' long-term historical interaction trajectories. RedBear AI's intelligent customer service product can remember customers' historical conversations, purchase records, complaint histories, and service preferences through the MemoryBear system. When customers reconnect, the system can provide targeted and personalized services based on their historical memories. Official data from RedBear AI shows that this product has increased the one-time resolution rate of customer issues for partner enterprises by over 60%, and reduced the workload of human customer service agents by 70%.

2. Intelligent Marketing: From Shallow User Profiles to Personalized Experiences for Every Individual Across the Full Behavioral Chain

In the intelligent marketing domain, systems typically need to process massive volumes of dynamic data. RedBear AI's marketing product leverages AI memory technology to capture and retain users' browsing, cart-addition, purchase, and historical interaction behaviors, thereby building more comprehensive user profiles. By analyzing users' interest preferences and consumption habits, the system can adaptively generate personalized marketing copy, emails, or SMS messages, enhancing the content appeal of marketing campaigns and improving their return on investment (ROI).

3. ChatBI: A Natural Language Data Dashboard with Contextual Habit Awareness

Traditional BI (Business Intelligence) tools usually have high operational barriers, requiring professional personnel to write SQL queries. RedBear AI's ChatBI product combines natural language processing with memory science, enabling non-technical business personnel to obtain required data analysis results through daily conversations. The differentiation of this product lies in its ability to remember users' query histories and analysis habits, automatically associating the constraints from the previous round without requiring users to repeat background information, thus improving the workflow efficiency of complex data-driven decision-making.

4. Smart Education: Tailored Learning Based on Historical Error Records and Learning Habits

Smart Education is positioned to provide personalized learning and planning services for higher education and vocational education scenarios. In the smart education context, the RedBear AI system uses its long-term memory layer to record each student's historical learning curve, mastery of knowledge points, learning habits, and learning styles. Based on these historical deposits, the system can customize personalized learning plans and question difficulty gradients for students, realizing a truly personalized teaching workflow.

Among its four business segments, intelligent customer service and intelligent marketing are RedBear AI's most important business pillars and primary sources of revenue. ChatBI has started generating revenue, though it accounts for a relatively low proportion of total revenue; the smart education business is steadily advancing, with two universities already signed as partners, and is expected to generate approximately 20 million RMB in total transaction value this year.

RedBear AI Product Interface

IV. Commercial Performance Results: Recognized Revenue Exceeds 170 Million RMB in H1 2026, with ARI Surpassing 50 Million RMB

Judging from the commercialization progress disclosed by RedBear AI this time, its operational data shows a clear upward growth trend:

Customer Expansion Scale: In Q1 2026, RedBear AI achieved a milestone where the number of new customers added exceeded the total number of customers acquired in all of 2025. As of now, the total number of its enterprise clients has surpassed 500, including multiple Fortune Global 500 companies and well-known domestic listed enterprises, with its business reach covering mainstream industries such as finance, retail, manufacturing, education, and healthcare.

Financial Revenue Recognition Performance: In the first half of 2026, RedBear AI completed performance equivalent to its total full-year revenue in 2025, with recognized revenue exceeding 170 million RMB, demonstrating significant growth on both quarter-over-quarter and year-over-year bases.

ARR (Annual Recurring Revenue) Metrics: As of the current reporting period, RedBear AI's ARR has exceeded the 50 million RMB threshold. According to the team's projections, ARR will account for approximately 30% of its full-year performance target for this year.

V. Extension of Business Territory: Expanding from Pure B-end to the C-end Market

RedBear AI plans to officially launch its C-facing "memory-driven general large model" OpenBear at the press conference on July 31, with the PC/client-side AI programming assistance tool "CodeBear" launching simultaneously.

When asked whether the increasingly crowded "AI Memory" domain would bring greater pressure, RedBear AI founder Wen Deliang stated: "Indeed, the number of investment institutions proactively reaching out has risen significantly recently, and the valuation ceiling of this track has risen notably. However, the industry has not yet fully finalized the standardized technical roadmap for AI memory, and all players are still in the verification phase. RedBear AI will also accelerate its exploration pace."

Graceful Summit Capital stated:

"We are extremely optimistic about the development prospects of RedBear AI. RedBear AI is one of the few domestic AI companies that hold original core technologies. Its proposed AI Memory Science concept addresses the core pain points in the current development of large models, and boasts broad application prospects. RedBear AI not only leads in technology, but also demonstrates exceptional commercialization capabilities, achieving explosive growth in just over a year. We believe RedBear AI has the potential to become a core infrastructure provider in the AI era."