China's Exclusive Multimodal Long Memory AI Company Secures Tens of Millions Financing - Betting on Active Intelligence

Base die manufacturers will not take over the Memory layer, and active intelligence will first emerge in Memory enterprises.

By | Wang Xinyi

Edited by | Zhang Yuxin

One-sentence Introduction

Thalamus Intelligence, the only domestic company focused on multi-modal long memory, has launched a native multi-modal memory foundation model, betting that AI will evolve from general-purpose to personalized, and eventually move towards proactive intelligence.

Proactive intelligence refers to the capability that AI can actively interact with users at the right time and in an appropriate manner on the premise of fully understanding users. To achieve proactive intelligence, Memory is an unavoidable threshold that must be crossed.

Financing Status

Recently, Thalamus Intelligence has completed a seed round of financing of tens of millions of yuan, with investors including top-tier Shenzhen-based funds and industrial capital. The proceeds from this round will be mainly used for technology R&D and talent team expansion.

Team Introduction

Thalamus Intelligence was founded in November 2025. Zhang Yuan, the founder and CEO, graduated from Peking University with dual academic backgrounds in electronics and economics. She once served as COO at an autonomous driving company, and also has working experience in the venture capital industry and embodied intelligence field.

The average age of the team is around 26 years old, and its core members mainly come from enterprises and universities including Alibaba Damo Academy, Tencent, SenseTime, the Chinese University of Hong Kong, the Chinese University of Hong Kong, Shenzhen, Peking University, Xi'an Jiaotong University, KIT, and Fudan University.

Products and Business

The industry consensus in the Memory field first took shape in academia at the end of 2025. This is a very new field. When meeting investors, Zhang Yuan is often asked questions such as: Will the foundational model build memory capabilities on its own? Will the Memory layer exist independently after all? How should responsibilities be divided between foundational model developers and memory-focused enterprises?

Zhang Yuan has a firm answer to all these questions: The memory layer will definitely exist independently as a long-term infrastructure.

She explained that different manufacturers have different foundational model architectures and training corpora, with different strengths in specific tasks. Users constantly switch between different models, which leads to the long-standing problem of memory silos. In the process of AI evolving from general-purpose to personalized, there is a missing intermediate layer that can continuously learn and remember users. Therefore, a user-centric third-party memory layer will inevitably exist independently of the foundational model, which is exactly the opportunity Zhang Yuan has identified.

However, for her, the first priority is to face the unsolved problems in the industry. At the current stage, there are certain pain points in the Memory field:

The first is the cost problem: most solutions directly feed a huge amount of full context into the window, resulting in extremely high inference costs. The second is the context fragmentation problem: it is difficult to effectively connect dialogues across sessions and tasks. The third is the modality missing problem: most existing memory solutions are limited to plain text and cannot process multi-modal information such as vision and audio.

Moreover, due to the inherent sequential disorder caused by the Transformer architecture, current large models cannot understand time. In the collaboration of Agents, memory silos often undermine task execution effects, as user information is locked in separate Agents and cannot be shared smoothly.

Zhang Yuan believes that if AI wants to truly integrate into the physical world, it must take human multi-modal perception as the reference from the very beginning of design. The memory architecture should be natively multi-modal from the start, instead of converting information into text for storage afterwards.

Accordingly, Thalamus Intelligence launched a memory foundation module named MemAura in June, providing customers with ADK (Agent Development Kit) and API (Application Programming Interface) services.

MemAura supports long memory accumulation and fast memory retrieval, adopts an efficient context mechanism, and can effectively reduce inference costs. In terms of performance results, MemAura delivers outstanding performance: The total Token consumption on the input side is reduced by 40%-49%, the memory retrieval latency is controlled within 400 milliseconds, the end-to-end first response can be generated within 1 second, and the overall accuracy can reach over 80%.

In the memory processing workflow, MemAura operates in a very different way from traditional memory systems.

Traditional memory systems usually extract, compress and summarize user inputs, then store them in the original information library, store data through a specific architecture or knowledge graph, and finally retrieve information relying on the Agent mode.

However, MemAura adopts a bionic memory processing mechanism, whose specific workflow is divided into three steps: first, identify events, imitate human memory mechanism to retain key information and filter out redundant noise; then, MemAura encodes the filtered information; the third step is to partition information, and isolate private memory from non-private memory.

In addition, they also adopt a tiered hot-cold storage strategy: for high-frequency accessed evidence, hot storage is used to realize fast retrieval; for low-frequency data, cold storage is applied to save resources. This strategy not only ensures that core data is accessible at any time, but also effectively controls the overall cost.

Zhang Yuan told us that in the multi-modal memory field, a trillion-yuan track, Thalamus Intelligence is not facing head-on competition at present. Although this track can accommodate many players, only a few startups are paying attention to it now, and the highest valuation among related companies is less than 20 billion yuan, so the competition is far from fierce.

Thalamus Intelligence's current main customers include high-volume companion hardware products such as companion robots, as well as vertical Agent scenarios such as AI customer service and digital employees. Customers use the memory API to help the foundational model call original information when answering questions and executing tasks, so as to improve interaction effects.

Different scenarios have different priorities for memory capabilities. For example, the requirements for companion products for children aged 3-5 and companion products for users aged 18-25 vary greatly. Zhang Yuan said, "We don't want to do excessive customized development, so we encapsulate Memory capabilities into different ADKs according to the weight of the scenario map, and customers can choose the corresponding ADK for access."

In terms of business model, MemAura charges through a combination of API call volume and scenario authorization, and customers can select different levels of memory capabilities according to their actual needs.

"At present, most of our customers are still focused on plain text or text plus voice scenarios. With the maturity of multi-modal related technologies, the modalities covered by customers will become richer, and more customers from embodied scenarios will come to verify our solutions," Zhang Yuan said.

Core Barriers

Since the company was founded, their technical route has completed two iterations, and they are gradually exploring the third iteration.

The first generation is the STKG Spatio-Temporal Knowledge Graph, whose specific implementation is: they place time and space in the physical layer, put multi-modal perception information in the perception layer, and build a cognitive layer to form a cognitive network for forward reasoning.

Its Benchmark performance is very impressive, achieving SOTA results in both LoCoMo (Reasoning and Timing) and LongMemEval (Interaction and Preference) tests.

However, for customers, after reaching a certain critical value, the improvement of Benchmark scores is not that meaningful. Customers care more about latency and Token cost. Therefore, they have made some trade-offs in engineering, carried out architecture-level iteration, turned to imitate the encoding and cognitive map of the human hippocampus and the integration mode of the cerebral cortex to build the memory system, and developed the bionic memory architecture MemAura.

In terms of data performance, MemAura has a leading edge of one order of magnitude. Currently, the latency of traditional online customer service scenarios can reach about 10 seconds, and the latency in companion scenarios is about 2 seconds, while MemAura can control the latency within 400 milliseconds.

At present, Thalamus Intelligence is still continuously exploring Memory related technologies, and they are using E2P (Embedding-to-Prefix) related technologies and fine-tuning the Transformer architecture to achieve the next iteration. At present, in the laboratory environment, they have achieved the effect of greatly saving Token consumption and approximate continuous learning in the field of preference extraction.

In May 2026, Thalamus Intelligence, together with NVIDIA, the Hong Kong University of Science and Technology, and the Chinese University of Hong Kong, released the world's first Benchmark for multi-modal long memory named MEMLENS. It is now open source, and has ranked top three in the Daily Paper list of the Hugging Face AI field.

For the first time, they placed 27 vision-language large models and 7 Memory Agents under the same set of multi-modal data, set 4 standard context lengths (32K, 100K, 128K, 256K), and carried out a complete controlled experiment, drawing the following conclusions:

First, vision-language large models usually do not perform any compression, and directly feed all content into the context window. As the length increases, the model performance drops sharply. Although the performance of Memory Agents is relatively stable, due to the design defects of the memory architecture, a large amount of visual and multi-modal data is lost at the writing stage.

Second, as the context length grows, both Memory Agents and vision large language models are more prone to hallucinations, and tend to confidently output false information when facing questions without sufficient evidence.

Third, many post-training methods for Memory weaken the model's ability to refuse to answer after training, which in turn increases the risk of hallucinations.

It is worth noting that this Benchmark proves that foundational model manufacturers and memory companies should have their respective divisions of labor, focus on their own areas of expertise, and occupy irreplaceable niches in the industry: foundational models should focus on visual perception, text-image alignment and underlying reasoning; memory enterprises should focus on cross-modal information retrieval, evidence organization and status update.

Founder's Thoughts

The end point of personalization is proactive intelligence.

Therefore, we judge that the end state of Agent should not be task-driven, but state-driven. In the future, Agents can help users make Planning based on their historical memory, and self-drive to complete tasks.

In the near future, we will see some proactive intelligent products or capabilities. If proactive intelligence emerges, it will most likely appear in companies focused on Memory research.

What Memory should do in the future is to precipitate user and task preferences, distinguish long-term status from short-term fluctuations, so as to better support proactive intelligence.

At present, most Memory-focused enterprises are focusing on context windows and RAG (Retrieval-Augmented Generation). However, I do not think the context window is equivalent to real memory. It is essentially a temporary working memory maintained by relying on front-end interaction and short-term context. As for RAG, the current mainstream in-memory implementation solutions mostly adopt the GraphRAG (knowledge-graph-based RAG) architecture, but it is essentially an external vector database that completes similarity recall through vector retrieval.

Many current enterprises, such as open source projects in Silicon Valley including Mem0 and MemGPT, still adopt the Vector Native or GraphRAG methods to implement memory functions.

Such solutions have obvious limitations. Taking pure vector retrieval as an example, it lacks modeling of information relevance, and is essentially more like a knowledge base that can only handle simple one-on-one question-and-answer scenarios. Real memory is not static question and answer, but needs to track the dynamically changing timeline and user intentions, and keep updating and adjusting.

Long Context and Memory Agent (memory) are not an either-or choice, but can be balanced through engineering trade-offs.

DeepSeek V4 has a 1M context window. After this model was launched, many people naturally thought of a question: will foundational models integrate memory capabilities, or will there be an independent Memory layer in future foundational models?

Our Benchmark gives the answer: foundational models should focus on extending the context length, improving KV cache (a data caching mechanism) related work, and strengthening underlying reasoning capabilities such as visual perception and text-image alignment. Memory enterprises should focus on finding cross-modal information to provide better services for users. The two sides have clear divisions of labor.

The core of Memory is not storage, but decision-making.

If it is only simple storage, traditional databases and storage systems have long been able to achieve it, and AI memory is completely unnecessary. But the real value of Memory lies in that it can make a series of judgments: whether to write information into memory, which part of the information should be written into memory, what kind of memory to recall, and when to actively initiate recall when answering a certain type of question. Based on this, Memory carries two core values:

The first value is to release the upper limit of AI potential. There is a key intermediate layer between the upper limit of the foundational model and the actual reachable upper limit of AI, which includes the complexity of the AI system and the personalized data closed loop (namely Memory). The personalized data closed loop is the core infrastructure that releases the upper limit of AI potential.

The second value is to support proactive intelligence. The foothold of AI always lies in human beings: to serve people, delight people, and empower people with higher efficiency, which is its unchanging mission. Whether it is Agent, physical or non-physical AI, all ultimately need proactive intelligence for support. The essence of proactive intelligence is decision-making: on the basis of fully understanding the personalized data closed loop of users, decide when to initiate interaction, how to initiate interaction, and what content to present.

In the early stage of household robot implementation, the demand for embodied Memory will surge.

In the process of AI moving into the physical world, its interaction entrance is multi-modal, so multi-modal long memory is an inevitable trend. No interaction with human beings can bypass memory. Especially when the interaction time is longer, the task chain is longer, and the involved modalities are richer, memory will play a more critical role.

In addition, in long-distance workflows, whether it is long-cycle Agent tasks or long-cycle embodied intelligence tasks, the richer the data modalities and the longer the memory span, the higher the data value. In other words, the closer the data is to the real state of the physical world, the higher the barrier it builds.

AI needs an anchor point mapped to the physical world, and this anchor point must be time.

Because time is unique, Memory cannot exist independently of time. No matter how our paradigm changes, we attach great importance to the role of time.

Image Source | Provided by the enterprise

Feel free to connect for discussions~