---
title: "Is OpenClaw's Memo Failing? Four Founders Diagnose the Problem"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-04-04T12:09:59+00:00"
canonical: "https://ffcap.cn/en/research/src-20260404-01html"
source: "https://uniqueresearch.substack.com/p/src-20260404-01html"
language: "en"
---

# Is OpenClaw's Memo Failing? Four Founders Diagnose the Problem

_Original · Unique Research · 2026-04-04 · Shanghai_

_Editor's note: This historical edition preserves both the reported narrative and the detailed panel conversation, including repeated quotations, the room-management interruption, and the source's occasional shift into third-person narration. “Memory.md is dead” is the article's rhetoric, not a verified software deprecation or version-specific defect finding. Funding, domain spending, audience reach, product capabilities, comparative accuracy, valuations, and demonstrations remain claims reported by the article or attributed to the speakers; no independent performance, financing, privacy, or partnership audit is implied. The speaker's “three or four years” description of startup tech giants and the Google/Pixel analogy are retained as source statements, not verified corporate chronology. “Today,” “yesterday,” and “a few days ago” refer to the conversation, whose exact event date is not established here, rather than automatically to the April 4 publication date. The moderator's “Cloud” wording is retained without silently replacing it with DeskClaw. English transliterations of Chinese names remain provisional. Predictions about education, employment, and AI replacing websites or professions are the speakers' views, not established outcomes or Unique Research advice. The extracted text is translated below; the 19 source images still require a separate completeness review before this draft can be approved._

Extraordinary Awards

Your AI Is Losing Its Memory

Memory.md Is Dead. The Memory Wars Are Only Just Beginning.

"

No one can record all their know-how, abilities, and experiences in a single memo.

The speaker was Song Jian, founder of NoDesk AI. He had been especially busy over the Lunar New Year because the team's DeskClaw was going head-to-head with the tech giants. With coverage from 20 official media outlets in Zhejiang and tens of millions of views across the internet, he had acquired an unexpected online notoriety: “Hangzhou Lobster Guy.”

But memory was giving him a bigger headache than the nickname.

A Single Memo Cannot Support AI's Memory Ambitions

User numbers at Song Jian's company have kept rising since launch. The product initially relied on OpenClaw, written by Peter, but has now moved to an in-house system.

“There was a major problem with his original Memory.md.”

Song Jian did not mince words onstage. “Put simply, no one can record all their know-how, abilities, and experiences in a single memo, and Memory.md is essentially just a memo.”

Laughter rippled through the audience. But beneath it lay anxiety.

For an enterprise, deciding when to use short-term memory and when to use medium- or long-term memory involves considerable complexity. A single memo document cannot handle it.

“OpenClaw has, at the very least, fully unleashed the capabilities of large models. It is certainly a good paradigm.” But then he shifted: “The new problem is that information has exploded and capabilities have grown stronger, but memory has become a disaster zone.”

So Song Jian and his team proposed a concept called “Evo-Memory,” aiming to build a closed loop from memory retrieval and synthesis through to evolution.

In Advertising, Hallucinations Cost Real Yuan

Song Jian offered an example.

“What should record your personal life, what should record everyday routines, and what counts as learning? In e-commerce, it becomes even more of a disaster. What needs doing every day? What needs doing at this particular moment? What should improve after it is done? What must stay unchanged every day after it is done?”

He paused, his tone turning serious.

“Even with something like ROI, placing ads is not a matter of adding or removing a little red dot. Ultimately, it is real yuan that bear the cost. Whether ad spending rises or falls, whether ROI points move forward or backward—these errors and hallucinations are still very evident in today's OpenClaw Memory.md.”

This means vertical providers must put substantial thought into defining the operating domain and use cases. Once those are defined, they must put substantial thought into categorizing them within Memory.

“That means memory management becomes even more important when we provide To B services. Otherwise, enterprises simply cannot put it into practice.”

A Silicon Valley Player's Answer: Multimodal Visual Memory

Shen Junxiao (Shawn) offered a different answer.

He is the founder and CEO of Memories.ai. The founding team previously worked on AI assistants at Meta. They raised US$15 million in Silicon Valley to work on something that sounds like science fiction: multimodal memory and visual memory.

“We believed then that there would definitely be Millions of Agents in the future, replacing all these websites and apps.”

But there is a problem: when you interact with these Agents in natural language, they are not sufficiently Personalized. You cannot give your personal information to every Agent.

“So there has to be an intermediate layer: the AI assistant layer. It connects you to all these future Agents. If those AI assistants are to be sufficiently personalized, they need Memory—and it has to be your Memory.”

Shawn believes future AI assistants will be visual and multimodal. Wearables and robots are both part of embodied intelligence.

“If we put that whole category under embodied intelligence, multimodal memory becomes very important.”

He spent a seven-figure sum in US dollars on the memories.ai domain. Asked whether it was worth it, he simply replied: “Worth it.”

Memory Needs Action to Create Value

But Shawn also acknowledged that memory is difficult to turn into a standalone product.

“Products place great emphasis on Time to value—the time it takes to realize value. Memory's Time to value is too long. You need a lot of time to know whether memory system A or B is better; it is hard to perceive the difference within ten minutes.”

And it is not for humans to use. It is for an Agent.

“So you are building memory Infrastructure for an Agent.”

To demonstrate how fast, effective, and detailed this Infra is, they built a First party app.

“Just as Google made its own Pixel phones to promote Android back then.”

A few days earlier, his Co-founder had met Amazon's leadership. They wanted to use the team's technology and asked what it could do.

“We told them to ask whatever they liked. They asked: ‘How many restaurants did you see on the way to our office? List everything about all those restaurants.’ The system immediately listed it all.”

Once a First party app is connected to the backend, you can reach a Time to value moment very quickly.

“It is difficult with memory alone. But I do not think that just because it does not exist now, it cannot exist in the future. Memory could be a standalone product; for now, it needs to be combined with an Agent.”

The End of Education?

Zhang Qiming approached memory from another angle.

He has made two products: Aibrary, which helps people learn AI, and BotLearn, which helps AI learn. Aibrary can recommend books, converse with you through books, and turn large collections of books into podcasts.

“Internally, we joke that Aibrary exists so you do not have to read books—or can read fewer books and read them faster.”

But Wu Wei posed a pointed question: “Now that AI intelligence has reached this level, why do humans still need an education?”

Zhang Qiming did not dodge it.

“That is a very difficult question. In the long run, if AGI really arrives, we think it may be meaningless for people to learn many things.”

But along the way, there must be a stage of human–machine collaboration. People would take more responsibility for metacognition and posing questions, then push an Agent to solve them. The Agent would implement many repeatable tasks whose capabilities can accumulate.

“We think that when people send their children to school today, they are using 19th-century teaching methods and 20th-century knowledge to educate 21st-century children.”

He offered two suggestions:

“First, introduce children to AI products as early as possible. Let them play with all kinds of AI tools from an early age. Second, instead of pushing your children to excel, push yourself. As artificial intelligence develops, children may struggle to find work in the future. So work harder on yourselves and leave them a little more wealth.”

The Founders' One Liners

Song Jian said: “Instead of being afraid, take action. Once you act, put your heart into it. As for the tech giants, other tech giants will take them on.”

Zhang Qiming said: “Instead of pushing your children to excel, push yourself.”

Shawn said: “Focus. Concentrate on what you are doing at the current stage; do the next stage when you get there. Finding exceptional people and teams matters most.”

Chen Hong, founder of memU, said: “I hope everyone can turn their abilities into Skills.”

A Final Word

Memory looks like a technical problem. In reality, it is a matter of survival.

When your AI assistant cannot remember what you said last week, when your Agent “loses its memory” at a critical moment, when memory hallucinations burn real money in your ad campaigns—you realize that Memory.md really is just a memo.

And a memo cannot support the AI era.

VII. More Details from the Conversation

Panel topic: The Memory Revolution: Long-Term Memory and Contextual Understanding for AI Agents

Moderator: Wu Wei, Founder and CEO of Unique Research

Panelists: Shen Junxiao, Founder and CEO of Memories.ai | Zhang Qiming, Co-founder of BotLearn/Aibrary | Song Jian, Founder of NoDesk AI | Chen Hong, Founder of memU

Wu Wei: Welcome, everyone! We do not actually meet very often in China. Sometimes we run into one another at events in Silicon Valley, so it is a rare pleasure to see you all here. For this opening Panel, I think we should first have everyone briefly introduce themselves, their company, and their relationship with Memory. I believe Memory is very important to every one of you.

Shen Junxiao: Hello, everyone. My name is Shawn, and I am the Founder and CEO of Memories.ai. Memories.ai is a technology company based in Silicon Valley. Our founding team all previously worked on AI assistants at Meta. When we were working on AI assistants at Meta, we felt that memory was the single most important thing for an AI assistant, without exception.

Wu Wei: Why is memory so important?

Shen Junxiao: Because ChatGPT had not appeared at that point, but at Meta we already had the concept of Agents. We believed then that there would definitely be Millions of Agents in the future, replacing all these websites and apps. But when you interact with those Agents in natural language, they are not sufficiently Personalized. You cannot give your personal information to every Agent. So there must be an intermediate layer: the AI assistant layer. It connects you to all those future Agents. If those AI assistants need to be sufficiently personalized, they need Memory—and it must be your Memory.

We also believe future AI assistants will be visual and multimodal. Your wearables, for example, and today's robots are actually part of the AI assistant category. If we classify that entire category as embodied intelligence, multimodal memory becomes very important. What we work on is precisely multimodal memory and visual memory. So we left Meta and raised US$15 million in Silicon Valley to build visual memory for future AI assistants and embodied intelligence.

Wu Wei: Understood. You previously worked at Meta and are based in Silicon Valley. How do you assess Meta's strategy in this wave of AI? It feels as though it sometimes moves very quickly, but at other times seems to be firing in all directions.

Shen Junxiao: It is certainly quite aggressive. But overall, I feel AI is developing incredibly quickly in Silicon Valley. Unless a company has good Infra—infrastructure—that coordinates both top-down and bottom-up efforts, it really is hard to win in an AI era where speed is decisive. So among Silicon Valley's big tech companies—Microsoft, Amazon, Google, and Apple—which has the greatest opportunity? Definitely Google.

Wu Wei: Definitely Google? You are not more optimistic about Anthropic or OpenAI?

Shen Junxiao: In our eyes, Anthropic and OpenAI count as “startup tech giants,” not traditional tech giants. Startups established only three or four years ago do not have many legacy problems; their Infra is built for AI. Their valuations are already in the hundreds of billions of US dollars, so I think they are already big enough in scale. If you ask who has the best opportunity, personally I think it is Google, followed by OpenAI and Anthropic. At the moment, Anthropic actually seems to be catching up from behind and moving ahead.

Wu Wei: Yes, I saw that OpenAI released something new today, too. No matter—Mr. Zhang, please introduce yourself.

Zhang Qiming: Hello, everyone. My name is Zhang Qiming. Our company is called Ouraca, short for OurAcademy, a shorthand for an “AI university of the future.” Our first product is called Aibrary. Because we want to build an AI university for future education, we started by building a library. Books are actually historical memory. We hope that if we can explain books well, we can probably explain many other things people learn well, too. Along the way, we have also done a lot of research—for example, into the kinds of talent future society will need. Artificial intelligence is developing especially fast, while traditional education already has many problems. After OpenClaw took off, we launched another product, BotLearn. We hope to accumulate human skills so an Agent can learn and evolve, enabling human–machine collaboration. So we mainly work on applications and evolution for models or Agents.

Wu Wei: So Aibrary helps people learn AI, while BotLearn helps AI and Agents learn?

Zhang Qiming: Aibrary is a library. It can recommend books, converse with you through books, and turn large collections of books into podcasts so you can read them quickly. Internally, we joke that Aibrary exists so you do not have to read books—or can read fewer books and read them faster. BotLearn, meanwhile, aims to help an Agent learn more, allowing you to collaborate with the machine.

Wu Wei: Now that AI intelligence has reached this level, why do humans still need an education?

Zhang Qiming: That is a very difficult question. In the long run, if AGI really arrives, we think it may be meaningless for people to learn many things. But along the way, there must be a stage of human–machine collaboration. People may take more responsibility for metacognition and posing questions, then push an Agent to solve them. The Agent would implement many repeatable tasks whose capabilities can accumulate. We think that when people send their children to school today, they are using 19th-century teaching methods and 20th-century knowledge to educate 21st-century children. Everyone can clearly see that programmers can now use Copilot to write a lot of code with a single sentence, and creating a PPT no longer requires a copywriter and designer—you can generate it directly with one sentence. So education will need to change substantially to meet future needs.

Wu Wei: OK. I am moderating onstage, and I can hear a bit of noise up there. Vera, please help me close the two doors. If anyone wants to have a discussion in the venue, please move outside; it is more suitable for conversation out there. We are all here to learn, so please keep quiet, all right? Good. Next, Mr. Song. I feel you have been especially busy over the past month. We have attended your events; please introduce yourself.

Song Jian: Hello, everyone. I am Song Jian from NoDesk AI. As Mr. Wu said, this Lunar New Year really has been quite busy, because as a young Chinese startup, we are going head-to-head with the tech giants at this very moment. We have a product called DeskClaw. DeskClaw should still be very popular in China right now. Many users say that, in their eyes, Qclaw, WorkBuddy, and DeskClaw are now competing directly. Of course, I am not trying to DISS Fu Sheng's products or Alibaba's. As a young company born and raised in Hangzhou—the secretary told me today that I must give this a shout-out onstage—we come from China Cloud Valley, within Zijingang Science and Technology City in Xihu District. Recently, because 20 official Zhejiang media outlets covered us, with perhaps tens of millions of views and hundreds of thousands of comments across the internet, I acquired my first-ever nickname as an object of online mockery: “Hangzhou Lobster Guy.” I never imagined having such a melodramatic name.

But I still think that, whether it is the tech giants or us, we are all essentially riding OpenClaw's momentum. We hope to combine general-purpose and vertical approaches, because everyone has their own domain. For us, the aim is to put Agent applications into practice in e-commerce and make them vertical. At this stage, I still feel OpenClaw has brought many things that are different.

Returning to today's topic, memory has been painful for us, too. Our user numbers have kept rising since launch. Initially, we relied on OpenClaw written by our friend Peter, but now we have switched to an in-house system. There was a major problem with his original Memory.md. Put simply, no one can record all their know-how, abilities, and experiences in a single memo, and Memory.md is essentially just a memo. It is even less adequate for an enterprise. Deciding when to use short-term memory and when to use medium- or long-term memory involves considerable complexity. A single memo document cannot handle it.

So ultimately, you still need a Memory System to help categorize and manage these Markdown files. For example, this Markdown file manages my Coding preference, another manages work, and another manages my personal life. Different Markdown files represent different Memory, and the Memory System decides how to manage them. But the problem now is that if you hand all that management over to an Agent, it may create a great many Categories for you, causing the file system to explode. So you still need a Solution for more sensible management. A Memory System offers that kind of good memory-management Solution.

Chen Hong: Yes, yes, exactly.

Wu Wei: Mr. Song, you just touched on this a little: your Cloud effectively draws on open source and then builds an in-house agent system. How do you do that in a way that ensures users have a better experience of memory management?

Song Jian: I think the most important thing I have learned at today's event is that I definitely need to work with Chen Hong's company afterward. As fellow startups, a spirit of collaboration matters. Second, from our perspective, we are more like elementary school students when it comes to iterating on memory. We were only founded in early ’25 and had our first anniversary on March 18. During this turbulent 1 year, we have had to put both e-commerce Agent use cases and Agent applications into practice.

After OpenClaw appeared, Memory.md blew up again. Song Jian and his team proposed a concept called “Evo-Memory,” aiming to build a closed loop from memory retrieval and synthesis through to evolution. This was mentioned in an article our other Co-founder, Wang Xiaodong, published on our official account yesterday.

OpenClaw has, at the very least, fully unleashed the capabilities of large models. It is certainly a good paradigm. Previously, there were flaws in how the model connected to its body, hands, and feet. The new problem is that information has exploded and capabilities have grown stronger, but memory has become a disaster zone. As mentioned earlier, what should record your personal life, what should record everyday routines, and what counts as learning? In e-commerce, it becomes even more of a disaster. What needs doing every day? What needs doing at this particular moment? What should improve after it is done? What must stay unchanged every day after it is done? Even with something like ROI—return on investment—placing ads is not a matter of adding or removing a little red dot. Ultimately, it is real yuan that bear the cost. Whether ad spending rises or falls, whether ROI points move forward or backward—these errors and hallucinations are still very evident in today's OpenClaw Memory.md.

This means that, as vertical providers, we must put substantial thought into defining the operating domain and use cases. Once those are defined, we must put substantial thought into categorizing them within Memory. That means memory management becomes even more important when we provide To B services. Otherwise, enterprises simply cannot put it into practice.

Wu Wei: Yes. For enterprises, aside from security and permissions, data matters. Then comes memory, because memory is also part of data. OK, Mr. Zhang, what is your view? After all, you have to handle a great many books and a lot of learning material.

Zhang Qiming: Let me continue along the same lines. Memory is very important. Take a person: everything they have experienced determines their present state. The same is true when we interact with a model or an Agent. From the perspective of education or books, how do we handle a model's parametric memory, including our understanding of the user themselves, well?

As everyone mentioned, it is difficult for users to feed in all their know-how and experiences. In our work, we found that practically every one of our users has LinkedIn. LinkedIn records where they went to school, which companies they worked at, and what roles they held. So we guide users to input their LinkedIn, making it much easier to personalize around their background. On that basis, we combine it with their interactions with us: what questions they asked, what books they read, how long they spent reading each book, and which parts they repeatedly listened to or asked about. We can then give highly personalized feedback. For example, if everyone comes from the internet industry and is reading Zero to One—0 to 1—or books on economics and finance, we can bring your personal experience directly into the explanation and discuss material that fits you. I think that is a very good approach.

Simulating a person is one thing. Along the way, we have also explored simulating a book's author to answer questions. Before the concept of a Skill existed, we proposed internally that every book is a snapshot of its author's worldview and methodology at a particular stage. For example, everyone working in the internet industry needs to drive growth, so we extracted all the core logic and cases from a large collection of growth books. If you are working on a product feature, you can input it and simulate the author of a book, using that book's real cases and methods for analysis. So books are actually a distillation of the memories of many people from history.

We have built many interesting Skills: you input a question, and a whole group of “books” appears to answer you. When their views differ, you can even have the “books” debate. Add your personal experience from LinkedIn—or even give it all your documents—and it can provide very comprehensive, detailed advice, with much higher accuracy than using a model directly.

Wu Wei: I think that is a very good approach. Most of my memory is actually in my WeChat chat history. If AI could analyze all of my WeChat chat history, I reckon it could simulate me completely. Also, while you were talking about books, I was thinking about a question. I have written a book myself, though mine is really just a casually written how-to book. In the future, might book authors also use a lot of AI, essentially rehashing old material and drawing on all kinds of news sources? Could that mean books will not be as valuable in the future as they were in the past?

Zhang Qiming: That will definitely happen. So from this perspective, we think that if we can bring many people with ideas, approaches, and methodologies into a community, they can put their thinking there and others can ask them questions. They themselves can also learn more, in a process of real-time iteration and evolution.

Wu Wei: Very interesting. Finally, let us talk a little more about OpenClaw and business development. Returning to Memories.ai, I think this is great—how much did you spend on that domain? A seven-figure sum in US dollars? Was it worth it?

Shen Junxiao: It was worth it.

Wu Wei: You certainly identified the key point. You must have a particularly strong voice on memory. What you do is multimodal memory management, which may become the core infrastructure for all future AI smart hardware with a camera. How did you decide to position yourselves in this area so early? What challenges were there, or has the moment simply arrived?

Shen Junxiao: We have actually worked on memory for a long time. We have always promoted memory, but now we promote memory together with the Action that follows. Why? Because we think memory is difficult to turn into a standalone product, and products place great emphasis on Time to value—the time it takes to realize value. Memory's Time to value is too long. You need a lot of time to know whether memory system A or B is better; it is hard to perceive the difference within ten minutes. And it is not for humans to use. It is for an Agent. So you are building memory Infrastructure for an Agent.

To demonstrate how fast, effective, and detailed our Infra is, we need a good Demo, a First party app—just as Google made its own Pixel phones to promote Android back then. Without this App to demonstrate it, it is difficult to persuade future OEM and ODM companies to use our technology. A few days ago, my Co-founder met Amazon's leadership. They wanted to use our technology and asked what it could do. We told them to ask whatever they liked. They asked: “How many restaurants did you see on the way to our office? List everything about all those restaurants.” The system immediately listed it all.

Once a First party app is connected to the backend, you can reach a Time to value moment very quickly. It is difficult with memory alone. But I do not think that just because it does not exist now, it cannot exist in the future. Memory could be a standalone product; for now, it needs to be combined with an Agent.

Wu Wei: It really feels as though future large models could swallow everything, combining all kinds of Agent functions. Back to Chen Hong: do you think there is a greater opportunity in multimodal memory Infra, or in the direction you are taking? And my next question is: what is your company's North Star metric, and why?

Chen Hong: Well, I think each has its own strengths. What we actually care more about is technological development. It is not a business metric in the traditional sense, but maintaining our technological leadership and keeping pace with everyone.

Let me also offer a suggestion. There are many tools today for improving efficiency within companies or for individuals. If you use tools like DeskClaw, the little lobster, you can actually turn all your abilities into Skills. In one sentence: I hope everyone can turn their abilities into Skills.

Wu Wei: Good, thank you, Mr. Chen. A round of applause! Turning abilities into Skills sounds great. But once they become Skills, could large models ultimately swallow them all and use them for training? Mr. Song, your one-sentence conclusion?

Song Jian: My one sentence is: instead of being afraid, take action. Once you act, put your heart into it. As for the tech giants, other tech giants will take them on.

Zhang Qiming: I work in education, so let me offer two suggestions. First, introduce children to AI products as early as possible. Let them play with all kinds of AI tools from an early age. Second, instead of pushing your children to excel, push yourself. As artificial intelligence develops, children may struggle to find work in the future. So work harder on yourselves and leave them a little more wealth.

Wu Wei: So everyone made the right choice coming here today! We are all learning. Shawn, your one-sentence conclusion?

Shen Junxiao: I do not know whether most people here are founders or investors. Shall we have a show of hands? Founders, please raise your hands. Investors? It seems most people here are Founders. I have not completely figured it out myself, so I will not hand out advice indiscriminately. But I have learned two things so far. First is Focus. Focus and sequencing: concentrate on what you are doing at the current stage; do the next stage when you get there. Focus matters a lot. Second, finding exceptional people and teams matters most. Those are my two things.

---

Original publication: https://uniqueresearch.substack.com/p/src-20260404-01html
On-site reading page: https://ffcap.cn/en/research/src-20260404-01html