给大模型做记忆 郝建业半年融三轮

This article introduces AI memory as an emerging independent track in the large language model (LLM) industry, where leading startups have attracted intensive capital investment thanks to their top-tier teams. Key takeaways are as follows:

1. AI memory is a core foundational capability for LLMs. Unlike traditional AIs that process tasks from scratch each time, memory-enabled AIs can remember user preferences and historical interactions to continuously refine outputs. It is central to improving user experience and retention, and has evolved from a technical concept into an independent industry track.

2. Currently, the field features two main technical approaches. Most early-stage players have adopted the "external memory" route that stores information in external vector databases. MemoraX AI, the startup that recently secured funding, pursues the native model approach, where memory capability is developed through model learning. This approach has ranked first globally across multiple authoritative benchmarks.

3. AI memory has already been deployed in various real-world scenarios including AI coding, interactive entertainment, healthcare, smart homes and wearable devices. It improves task efficiency and user experience, and is widely seen as a core competitive direction in the next phase of LLM development.

As the LLM industry enters a new development stage, the rise of the AI memory track brings new opportunities for brand product R&D and user operation. Key insights are as follows:

1. Shifting consumer expectations: LLMs have evolved from general-purpose intelligence to personalized intelligence. Users increasingly expect AIs to remember their habits, preferences and historical behaviors to deliver tailored services from the very first interaction. Memory has become a core competitive advantage for AI products and a key new user demand.

2. New product development direction: Brands developing AI-enabled products can build differentiated experiences by integrating AI memory capabilities. For example, chronic disease management tools can remember users' long-term health conditions; smart home systems can store preferences of all household members; and interactive entertainment can maintain consistent character arcs over long-term use. All of these effectively boost user retention.

3. Ready-to-use partnership opportunities: MemoraX AI has already launched standardized memory modules and formed a strategic partnership with Huawei Cloud. Brands can access these mature technical solutions to implement memory capabilities quickly and reduce in-house R&D costs.

AI memory has emerged as a new high-growth track in the next phase of LLM development, creating new business opportunities for sellers. Key takeaways are as follows:

1. Large room for market growth: Third-party estimates project that the global Agentic AI orchestration and memory system market will grow from RMB 43.9 billion in 2027 to RMB 199.2 billion in 2030, representing a compound annual growth rate (CAGR) of 35.3%. The track remains in the early stage of development with enormous untapped growth potential.

2. Clear demand across multiple sectors: AI coding requires long-term project memory to cut down redundant communication; interactive entertainment needs memory to maintain consistent characters and storylines; industries including healthcare, smart homes and wearables require memory to enable personalized interactions; and end consumers also have clear demand for personalized AI. There are abundant opportunities across verticals.

3. Lower entry barrier via mature ready-made technologies: Leading memory technology vendors have launched standardized memory modules and formed partnerships with Huawei Cloud. Sellers can leverage existing technologies to focus on developing AI products for niche verticals and enter the track quickly.

The rise of the AI memory track brings new commercial opportunities for factories and opens up new directions for their digital transformation. Key insights are as follows:

1. New demand for product design and manufacturing: More and more factories are launching smart hardware products including smart home devices, wearables and intelligent terminals. Consumers are no longer satisfied with basic AI functions; they expect AI to remember their usage preferences and long-term needs to deliver continuous personalized interaction. Factories can integrate AI memory capabilities into product design to build differentiated products, improve competitiveness and increase profit margins.

2. New commercial opportunities: Factories can partner with AI memory technology vendors to quickly add memory capabilities to their existing smart products, expand product functionality, tap into new market growth, and reach consumers who prioritize personalized experiences.

3. New insights for digital upgrading: In the process of digital and e-commerce transformation, factories accumulate large volumes of production, supply chain and user data. AI memory capabilities can help factories effectively store, retrieve and utilize these data, improve the efficiency of production decision-making and user operation, and support the implementation of digital transformation.

AI memory has emerged as an independent industry track, bringing new growth opportunities for AI-related service providers. Key takeaways are as follows:

1. Clear industry development trend: Base layer LLM technology has gradually matured. The core of next-stage industry competition will be meeting user demand for personalization, and memory is the core infrastructure to enable personalization. The market is growing rapidly with a CAGR exceeding 35%, and capital has already started intensive布局. The track remains in the early high-growth stage with large untapped blue ocean space.

2. Clear core pain points for clients: In the current LLM implementation process, common pain points include high processing costs for long-cycle interaction context, inaccurate retrieval of historical information, low satisfaction of personalized demand and poor experience for long-horizon tasks. Traditional external memory solutions based on vector databases cannot fundamentally solve these problems.

3. Mature solutions are already available: The native model memory approach has been validated by leading players. MemoraX AI has achieved top results across multiple authoritative benchmarks and launched standardized memory products for multiple scenarios. Service providers can integrate these mature technologies to offer memory-enabled solutions for clients across different industries and expand their business scope.

The rise of the AI memory track brings new demands and development opportunities for the ecological construction of LLM platforms. Key insights are as follows:

1. New market demand for platforms: In the process of LLM implementation, clients increasingly need mature long-term memory capabilities to support long-cycle interactions and personalized services. Existing base LLMs lack mature memory capabilities on their own, and require independent memory infrastructure as a supplement. Platforms need to improve this part of their capability ecosystem.

2. Key areas for platform布局: Platforms can proactively布局 the emerging AI memory track by introducing technologically leading memory vendors to enrich their ecosystem and provide clients with more complete LLM solutions. For example, Huawei Cloud has already formed a strategic partnership with MemoraX AI, making it one of the first domestic cloud platforms to布局 the AI memory track.

3. Insights for investment recruitment and risk mitigation: As the track is still in the early stage, platforms can prioritize teams that combine top-tier academic expertise with industrial implementation experience in investment recruitment. Such teams are more likely to deliver technological breakthroughs and reduce cooperation risk. Platforms can also launch targeted support programs for the AI memory track to attract leading teams and seize first-mover advantage in ecosystem building.

This article discloses the latest industrial and technological developments in the LLM field, providing new directional references for AI-related research. Key takeaways are as follows:

1. New industrial trends: The base LLM ecosystem has gradually matured. The core competitiveness of next-generation LLMs is shifting to long-term memory capability. AI memory has evolved from a technical concept to an independent industrial track, attracting intensive capital investment. The global market size is projected to approach RMB 2 trillion by 2030. Currently, the field has formed two technical routes: external memory and native model memory. The native model approach is favored by giants including Google and Meta as well as leading startups, and has become the main research direction today.

2. New unaddressed problems in the field: Existing memory capability evaluation benchmarks suffer from logical contradictions caused by synthetic data generation and fail to reflect real-world scenarios, meaning the "measuring stick" for memory capability is inaccurate. The industry has started to build new open-source evaluation benchmarks that are closer to real-world scenarios to address this pain point, and multiple new benchmarks covering different scenarios have already been launched.

3. Referenceable business model: Leading startups have adopted a model of core underlying technology research combined with modular implementation for vertical scenarios, providing memory solutions for both B2B and B2C products. They have already reached cooperation with clients across multiple industries, validating the commercialization path and providing real-world samples for industrial research.