The center of the AI trade in 2026 has been memory. Each week, as I go through X looking for news, insights, and opinions on AI, memory comes up more than any other topic. Whether you believe AI is still in the early innings or that it is a bubble about to burst, memory fits both arguments. That is what keeps it at the center of the trade.
This week, I came across a Korean-language interview with KAIST Professor Kim Jung-ho, often called the “father of HBM.” After going through the transcript and reflecting on this weekend’s video about Micron’s earnings report, especially its commentary on the future of memory demand, I came away even more convinced that most people still underappreciate the importance of memory for the future of AI. As Dr. Kim put it, “AI equals memory.”
That line immediately reminded me of a book I read years ago: Joshua Foer’s Moonwalking with Einstein. If you have not read it, put it on your summer reading list, especially if you are interested in understanding why memory is becoming so important for AI. The book shows what your own brain is capable of when memory is used architecturally, not as random storage, but as a structured system for organizing and retrieving information. It stayed with me because it changed how I thought about memory.
The book is a true story. Foer first arrived at the U.S. Memory Championship as a journalist covering the event, not as a competitor. The rest I will leave for you to read, but the key point is that he spent the next year training in the ancient techniques of the memory palace with the hope of participating himself. That is what makes Moonwalking with Einstein so useful for thinking about AI: the lesson is not that he suddenly acquired a better brain. The lesson is that he gave his memory an architecture.
That is where the connection to AI becomes so powerful. Long before AI, the Greeks understood that intelligence depended on memory. Plato gave us the philosophy of memory: the idea that learning, knowledge, and recollection are deeply connected. Simonides of Ceos gave us the architecture of memory: the method of loci, or memory palace, where orators learned to memorize long speeches by placing each idea inside a familiar physical space and mentally walking through it in order.
Foer brings that ancient technique back to life in Moonwalking with Einstein. One of the reasons the book stays with readers is that he does not merely explain the memory palace; he makes you experience it. He walks through a list of ordinary items, pickled garlic, cottage cheese, smoked salmon, wine, socks, hula hoops, a snorkel, a dry ice machine, and turns them into bizarre images placed along a familiar route. By the end, the reader remembers the list not because the brain suddenly became more powerful, but because the information was given an architecture. I still remember a colleague, Dave Patterson, hearing me talk about that part and remembering it word by word as well, like your favorite song growing up that never leaves you.
That is the right metaphor for AI. We have spent the last three years building a more powerful brain. But intelligence does not scale on brainpower alone. It scales when the brain has a memory palace: a structured system for storing, retrieving, and acting on information. Memory is not just storage. Memory is architecture. It is the difference between having a warehouse full of information and having a working system that knows where everything is, how to retrieve it, and when to use it.
That idea is the right starting point for understanding the next phase of AI. The debate investors keep having is whether we can still be in the early innings of the AI buildout when the stock charts already look like they have had a full cycle. I understand the argument. Many of the AI winners have already moved dramatically. There are signs of speculation throughout the US, Korea and Taiwan. The memory companies and the power and infrastructure names have already created enormous market capitalization. From the charts alone, it is easy to say this must be late. But that is the wrong measuring stick. The better question is not how far the stocks have moved. The better question is how far the intelligence stack still has to evolve.
The reality is that the brain keeps getting smarter. Every new model generation is not just a slightly better chatbot. The models have become more capable at coding, reasoning, tool use, computer use, long-horizon tasks, and workflow execution. Anthropic described Claude Opus 4.5 as setting a new standard across coding, agents, and computer use, and that is the point investors need to sit with. The market is watching stock charts, but the technology curve is watching capability. As the models become more useful, now with loops and tags, the system around the model has to expand. A smarter brain needs a larger memory palace. It needs more context, more working memory, more persistent storage, more retrieval, more bandwidth, and more power-efficient access to everything it has learned or needs to know. We are now Moonwalking with Agents.
This is why I think the recent jump in agentic capability is the memory moment. ChatGPT made the world realize the model could talk. Claude, Opus, and the new agentic systems are making people realize the model can work. That changes the memory requirement completely. A chatbot answers a question. An agent runs a process. It reads files, searches databases, writes code, opens tools, remembers what it tried, compares alternatives, revises plans, and continues over time. That requires more than a prompt window. It requires a real memory architecture. One user may eventually run ten agents. One company may run thousands. A large enterprise may run millions of small agentic workflows in parallel. Each one of those workflows needs context, state, retrieval, and persistence.
The analogy Dr. Kim used is simple. The GPU is like a brilliant analyst. But memory is the analyst’s desk, filing cabinet, library, whiteboard, and courier system. If the analyst is brilliant but has no desk, he cannot organize the work. If he has no filing cabinet, he cannot keep history close by. If he has no library, he cannot access the broader knowledge base. If the courier system is slow, he spends the day waiting for the right files to arrive. The analyst may be brilliant, but the system is broken. That is where AI is heading. The model may be increasingly intelligent, but if the memory subsystem cannot feed it fast enough, hold enough context, and retrieve the right information at the right time, the intelligence gets trapped.
This is why the “late innings” argument based only on stock charts misses the deeper point. The charts highlight how investors had to catch up to the rise of the agents. Everyone underestimated the rise of the agents. Everyone including Andrej Karpathy.
In October 2025, just one month before Opus 4.5, Andrej Karpathy told Dwarkesh Patel that agents were better understood as a decade-long project, not a one-year event. His diagnosis was telling: “You can’t just tell them something and they’ll remember it.” That is the memory problem in one sentence. Agents do not become real workers until they can carry context, learn continuously, retrieve prior state, and use memory over time.
Now with agents, the memory requirements are just starting to be measured for the next several phases of the AI buildout. Chatbots were the first use case. Agents are the second. Edge devices and embodied AI will be the third and most memory intense. Each step requires more memory, not less. The smarter the brain becomes, the larger the memory palace has to become around it.
That is why Professor Kim Jung-ho’s interview was so important. His line was essentially that “AI equals memory.” He did not mean GPUs, math, models, or software stop mattering. He meant the bottleneck is shifting. In the training era, compute created intelligence. In the inference era, memory determines how useful that intelligence becomes. He made the point even more directly when he said that the capability of AI is increasingly determined by the capability of memory. That is the investment insight. The market is still trained to think of memory as a cyclical semiconductor input. Dr. Kim is arguing that memory is becoming the architecture of intelligence.
The important part of his interview is that he does not stop at HBM. HBM is the high-speed working memory that keeps the accelerator fed, but agentic AI and physical AI require something larger. As context windows expand, multimodal data grows, and agents begin to carry history across time, AI needs far more than fast DRAM sitting next to a GPU. That is where he introduces HBF, his framing for high-bandwidth flash, and eventually HBS, or high-bandwidth SRAM, as part of a future 3D memory-centric architecture. His vision is not a GPU with memory attached. It is a memory-centric AI computer where compute moves closer to the data because moving the data has become the bottleneck.
The reason this still feels early is that the largest memory demand waves are only beginning. Dr. Kim specifically points to agentic AI and physical AI as the next step changes. A chatbot answers a question. An agent runs a process. A robot, autonomous car, or humanoid has to interpret the physical world in real time. These systems are not just reading text. They are reading reality. They need video, audio, spatial maps, sensor inputs, motion planning, object recognition, safety constraints, and real-time control. A chatbot can pause before answering. A car or humanoid cannot pause before deciding whether to brake, turn, reach, grasp, or stop. Physical AI does not reduce the memory problem. It explodes it.
Micron’s earnings call confirmed the same thesis from the corporate side. Micron said AI system performance is increasingly dependent on the memory subsystem’s performance and capacity, and it explicitly framed memory as a strategic asset in the AI era. That is a very different statement from the old memory cycle. In the old world, memory was inventory, pricing, and utilization. In this world, memory becomes a strategic bottleneck for customers trying to secure the future of AI infrastructure.
The most important part of Micron’s commentary was that the memory demand story is not limited to data centers. Micron tied future demand to agentic AI, automobiles, robotics, and humanoids. It noted that L2+ and above vehicles carry more than five times the memory and storage content of an average vehicle, and that humanoid robots carry ten times the memory of an average L2+ vehicle. That is the early-innings argument in one sentence. We are still mostly debating GPUs and data centers, while the next memory demand waves are coming from agents, AI PCs, AI phones, full self-driving, robotics, and humanoids.
Micron also discussed long-term strategic customer agreements, including agreements representing roughly $100 billion of cumulative revenue from 2026 to 2030 at minimum contract pricing. That matters because customers are no longer treating memory like a normal commodity input. They are trying to secure supply. When customers sign long-term agreements with pricing floors and commitments, they are telling you that memory has moved from a cyclical component to a strategic resource.
That is the real investment conclusion. The market keeps asking whether the AI buildout is late because the charts are extended. The charts represent how investors underestimated when the IQ would reach the point where the agentic world became a reality. Now with investors aware, I expect the second derivative of the charts will slow as part of a mid-cycle slowdown but the memory architecture says we are still early in demand. We have built the first versions of the brain. Now we need to build the memory palace around it. The scarce resource is controlled by very few companies, and the strategic importance is rising for corporations and countries at the same time. AI started as a compute race. It is becoming a memory race. In the training era, compute created intelligence. In the agentic era, memory gives intelligence context. In the physical AI era, memory gives intelligence situational awareness. That is why memory is becoming the operating layer of AI or as Dr. Kim said, AI equals memory.
