At the beginning of the conversation, Fan Chenggong first demonstrated MemoryBox, a personal memory management tool. On the left are different memory spaces such as life and work; on the right are nearly a thousand imported historical ChatGPT sessions and synchronized local files. What this demonstration truly aims to prove is not just another AI entry point, but something more fundamental: when a user switches from ChatGPT to Gemini, and then to Claude, what should truly not be lost along with the entry point is personal memory. Behind this judgment is the intuition Fan Chenggong has formed over the past 20-plus years in the storage industry.
As the co-founder of MemVerge, he also created the open-source project MemMachine and is preparing to launch a new product, MemoryBox. In Fan Chenggong's view, the next stage of AI is not just about competing in model capabilities, but also about who controls the ownership of memory. Looking at a deeper level, this battle for memory ownership is actually a deeper infrastructure migration: the storage of AI is being rewritten by the logic of memory.
Core Insights
- Memory is becoming an independent layer.
Starting in 2025, telecom operators are deeply entering the market, leveraging their cost and subsidy advantages to secure over half of the orders from major tech companies, and directly adopting the mature supply chains cultivated by third parties. Traditional third-party service providers will either exit the competition or be downgraded to subcontractors, bearing the risk of secondary settlement.
- Memory is not a bucket.
Episodic memory, semantic memory, procedural memory, and document memory correspond to four completely different governance logics. The true difficulty of memory lies not in storing it, but in compression, deduplication, recall, and conflict management.
- Models can be centralized, but memory cannot.
Large models can share public network data for training, but long-term memory is inherently private data. Therefore, local first, data sovereignty, and memory sovereignty will become long-term architectural choices.
- A unified memory layer will not be completed by model vendors.
Major companies will certainly continue to supplement their respective memories, but a memory layer shared across models, Agents, and applications will maintain its independent space for a long time.
- Personal memory is the entry point; organizational memory is the main battlefield for commercialization.
When memory begins to accumulate team experience, task processes, and organizational know-how, what it competes for is no longer just a feature, but the foundational capability of an enterprise.
The following is the transcript of this interview, slightly modified without changing the original meaning.
01. Memory Rewrites Storage in the AI Era
Aifenyi: You have been in the storage industry for over 20 years and are now working on memory. How do you view the relationship between memory and storage?
Fan Chenggong: To me, these two things are conceptually connected. The three most core elements of a computing system have always been computing, storage, and communication. In the AI era, large models are the new computing layer, and memory is the new storage layer. Of course, communication is still needed in the AI era, but these three core functions remain unchanged. Following this logic, you will find that memory is the new way of processing information in the AI era. What it stores is no longer just files and tables, but users' history, preferences, task processes, relationships, and accumulated experience.
Aifenyi: In the previous storage layer, databases played a core role. In the AI era, will the value of databases be weakened?
Fan Chenggong: That is a very sharp question. My view is that databases will not be eliminated, but they will no longer dominate the entire landscape. Traditional database technologies remain important and will certainly continue to be part of modern systems; it is just that with the introduction of large models, methods for data processing, data preservation, and data querying will all change. So if you ask me whether the proportion of databases will decline, I think new elements will come in, and the proportion of the traditional system in the overall system may decrease, but databases, file systems, and such will continue to hold their ground. Even if there are dedicated memory models in the future, a lot of content will still need to be written and organized through other layers. The future memory system will be more complex; the database will just be one part of it, not the whole.
Aifenyi: How do you judge the current stage of the memory market? Many people feel that model vendors have not yet truly exerted effort in this area.
Fan Chenggong: I don't think they haven't exerted effort. Over the past 12 to 18 months, at least the model vendors in the US have been gradually enhancing their memory capabilities. ChatGPT has short-term and long-term memory systems, and cross-session profile memory has also been improving. Gemini has persistent context, and Claude also has its own memory system. So, if you only use one model, its memory capability is actually getting stronger and stronger. The major companies are not ignoring it; they are building memory within their own models.
Aifenyi: Are major companies building memory, or are they doing it as an ancillary feature of AI super apps?
Fan Chenggong: Yes, major companies are unlikely to actively support other companies' models. Their business model relies on the models themselves, and memory is also part of enhancing model stickiness. But the reality is not a single-model world. Individuals switch between different models, and enterprises use multiple models, multiple Agents, and multiple apps simultaneously. Cross-model memory, multi-Agent shared memory, and unified memory layers are mainly being developed by startups, open-source projects, and new teams. I don't think model giants will dominate the memory space. The memory built by major companies will certainly have its audience, but independent memory systems will definitely have their space too.
Aifenyi: Why haven't enterprises built a unified memory layer on a large scale yet?
Fan Chenggong: Because most enterprises have not yet reached the multi-Agent stage. Problems are often concentrated in a single scenario, such as coding, customer service, or a specific single-point Agent, and they have not yet reached the step of comprehensively considering the architecture. Therefore, most enterprises have not yet integrated multiple apps and multiple models into a shared memory system for coordinated management. I think the direction of a unified memory layer is clear, but the timing is not yet fully mature.
02. Memory Is Not a Bucket: Four Types of Memory, Four Governance Logics
Aifenyi: How do you define memory?
Fan Chenggong: I think the field of memory is quite complex; at least we need to classify it first. The first type is episodic memory, which refers to specific things that have happened. What time, with whom, and what was done at what location; it is further divided into short-term and long-term, and long-term episodic memory will also undergo processes like merging, deduplication, and conflict resolution. The second type is semantic memory, which is the stable summary extracted from many episodic memories. The third type is procedural memory, which is about how to do things. For example, how to break down a task into steps, how I usually express myself, how I present; skills can also be seen as a kind of procedural memory to some extent. Broader still, there is document memory, which refers to accumulated content like files, knowledge, and knowledge graphs. Different types of memory have different ways of writing, updating, querying, and recalling. You cannot just stuff user preferences, task processes, operational steps, and document content into a big warehouse and expect the system to sort them out on its own. A truly good memory system does not just store things; it needs to know what should be distilled into semantic memory, what should be retained as episodic, and what should be consolidated into procedural memory.
Aifenyi: In the long run, will there be a unified memory architecture that can address all four types of memory simultaneously?
Fan Chenggong: I think there will be, and likely more than one. There should be a relatively complete memory technology that can cover episodic, semantic, procedural, and document memories. But there is a premise here: memory will not become a large pool shared by everyone like large models. Memory is inherently private and personalized. So it can have a relatively unified technical architecture, but it will not be the case that everyone's memory enters the same system for unified processing.
Aifenyi: Which types of memory are currently covered by the memory of large models themselves?
Fan Chenggong: Taking ChatGPT as an example, memory within a single session is essentially short-term memory. Long-term memory is closer to semantic memory, which is the user's profile memory. True long-term episodic memory has not been covered yet, and procedural memory is not yet mature.
Aifenyi: In terms of technical implementation, what are the mainstream approaches you see?
Fan Chenggong: The first type is vector databases, which turn content into embeddings and then perform semantic queries in the vector space. This was first adopted by RAG and later became an important part of memory projects. The second type adds reranking on top of vectors. The results of vector retrieval, combined with the model's reranking judgment, yield better relevance performance. The third type organizes memory around time and location, because memory is deeply related to time. The fourth type is the markdown file approach used by many agents today, overlaid with different retrieval methods. The fifth type goes a step further, directly building dedicated memory models and training long-term memory into the models. Other methods will emerge in the future, but the goals are all the same: to make memory a system that can be organized, compressed, deduplicated, and managed for conflicts.
03. Memory Is a Countervailing Force Against Models and NVIDIA
Aifenyi: What is the core product philosophy of MemoryBox right now?
Fan Chenggong: It's very simple: I am the master of my own memory. Memory is different from large models. The data used to train large models mostly comes from the public internet and is public data; but memory is private, personalized, or belongs to the enterprise itself. You cannot mix my memory and your memory in the same system for processing. Therefore, our central viewpoint is to build private memory. Whether for individuals or enterprises, everyone should have their own memory system. What we provide is this software, rather than collecting everyone's memory into a public pool.
Aifenyi: Is this also why MemoryBox is made into an independent product?
Fan Chenggong: Yes. MemoryBox is a product we launched for end-users. It can be installed on your computer, phone, or your own cloud drive. All memories are stored there; we cannot see them, only the owner can. It can connect to the file system, as well as your conversation history with various AIs, organizing this content into different memory spaces. You can select one of these memory spaces as context to interact with one or more models. I can import all my ChatGPT conversations, and my local files and cloud drive files can also be synchronized. In the end, they won't just be piled up in a single list, but will be organized into different memory spaces, such as life, work, or a specific project.
Aifenyi: If only conversation history is synchronized, will some memory be lost?
Fan Chenggong: The original conversation records will all come in, which is the most important part. Based on these records, we can also make summaries ourselves, such as extracting personas and creating profile memory. The long-term memory that ChatGPT creates for users can be exported as much as possible through prompts, but whether it is 100% is hard to say.
Aifenyi: Palantir's CEO has also recently expressed dissatisfaction with model vendors. Has the US market elevated local first and privacy first to a very important position?
Fan Chenggong: There is indeed a strong tension here. On the one hand, model vendors are becoming increasingly powerful, holding high discourse power in both the US and China; on the other hand, more and more individuals and enterprises are realizing that they cannot hand over all their data, knowledge, and memory at once, and everyone will emphasize data sovereignty and memory sovereignty. The AI field needs a countervailing force to balance NVIDIA and foundation models, so I think local first is not the only answer, but it will certainly become one of the long-term trends.
Aifenyi: Besides local first, what other technical paths can further protect memory and privacy?
Fan Chenggong: There are several directions. First, in the future, many enterprises, and even some individual users, will enter a hybrid model mode through model routing. Small models are getting better and better; besides using large models, enterprises will also deploy their own models. For tasks that are highly sensitive to privacy or do not actually require large models, they can be handled first in their own models and memory systems, and only when necessary will requests be routed to external large models. Second, if interaction with external models is still required, some information can be desensitized and replaced with other identities to minimize privacy exposure. In other words, ownership is not just a statement of values; it will in turn determine how the system is deployed, how models are routed, and how memory is kept.
04. Start with Personal Entry Points, Then Move to Organizational Memory
Aifenyi: How far are evaluations like LoCoMo and LongMemEval from real-world scenarios?
Fan Chenggong: I think these are certainly valuable; they are a very good start for technically measuring information retrieval systems. But this field is developing very quickly, and many benchmarks are now somewhat shallow, with everyone able to achieve relatively high scores.
Aifenyi: What metrics does your internal evaluation system focus on?
Fan Chenggong: There are mainly two. The first is the improvement in answer quality. Without memory, AI might only score 30; with memory, it can score 80, which is a very tangible improvement. The second is saving token costs. If the system can remember past content, organize it well, and retrieve it on demand, you don't need to feed the long context back in every time. Third, we also look at how many tokens the memory system itself consumes. Fourth is latency, which is the impact of this system on speed and overall performance. Simply put, we look at both benefits and costs; we cannot just say it is smarter after adding memory without looking at how much heavier, slower, and more expensive the system itself has become.
Aifenyi: Why move from the developer-facing open-source project MemMachine to the end-user product MemoryBox?
Fan Chenggong: Open-source products mainly target developers, especially agent developers. For this group, open source is important because they need to know what your mechanisms are. But a change now is that many people working on memory treat it as a backend-oriented, auxiliary capability to help Agents or apps work. This is certainly correct; memory does have this aspect. But we feel that memory is also a very intimate and core concept for ordinary individual users. Because as humans, memory itself is the source of our uniqueness. To a large extent, I am actually the accumulation of my memories. So in the AI era, how to protect one's memory and how to make memory truly work for oneself rather than being taken away by platforms—this in itself can stand as a product.
Aifenyi: The MemoryBox team previously had strong 2B (to-business) experience. What is the reason for choosing to enter the memory field from the 2C (to-consumer) side?
Fan Chenggong: I think in the AI field, the boundary between 2C and 2B is actually blurring. Although our first version of users were individual users, many of them are professionals themselves, and the scenarios are mainly productivity and work scenarios. Transitioning from here to enterprises is quite natural. Therefore, I don't think 2C and 2B are completely disconnected paths. Market promotion and sales methods are certainly different, but in terms of product architecture and technology, they are connected. Once personal memory is established, moving many things to the organizational layer will be more natural.
Aifenyi: What are the core goals for 2026 and 2027?
Fan Chenggong: Our most important goal this year is to build a product that users truly like and are willing to use continuously. It would be even better if it could become a phenomenal product, but at the very least, we need to first create a viable memory product to give users a new AI experience. By next year, we hope to further solidify the memory foundation on this basis. In addition to personal memory, we also need to develop organizational memory. At that time, we will serve both individual and enterprise users. I think this path has the potential to produce phenomenal products and also has very clear commercialization space, especially when it reaches the organizational memory stage, the commercial path will be clearer.