Why the Next AI Won’t Forget You

A16z published a great explanation of why today’s AI keeps resetting every time you talk to it. Here is what I believe it means for anyone building, buying, or betting on the next wave.

I ordered a Mac Mini a few months ago to run an open-source AI agent called OpenClaw. Sandboxed setup. No connection to personal data. No work systems. First job I gave it was trying to act as my eyes and ears of the news and how it relates to my book and ideas that I write about. If it earns that and does it well, I add the next thing.

I mention this because the first time I sat down to work with it, I noticed something I had been writing about for two years. The AI does not actually remember anything – not in the way you want it to. Every time I opened a new session, I was starting over. Ok – Not fully. There are workarounds built in and I am still working on it. But the intelligence itself was not compounding. It was looking things up.

The piece that Andreessen Horowitz published yesterday put a name on what I was experiencing and made the case that this is the single biggest limitation holding AI back right now. [1] I read it twice. I think they are mostly right. I also think the implications are bigger than they said, and they matter for anyone making bets on where AI goes from here.

The movie Memento, but for your AI

The a16z piece uses Memento as its metaphor. If you have not seen the movie, the main character has a brain injury that prevents him from forming new memories. Every few minutes, his world resets. He copes by tattooing notes on his body and taking Polaroids to remind himself what just happened.

That is how AI works today. The model itself never retains anything from your conversations. What looks like memory is really a collection of workarounds. Chat history is the note. Retrieval systems are the Polaroid. Every interaction, the AI reads the notes, figures out where it is, and starts over.

This works, up to a point. It is also the reason your AI feels like a very smart stranger who has amnesia, instead of a colleague who has worked with you for years.

Why this matters – at least for me

I have a new book coming out this summer called The Invisible Interface. [2] The argument, in one sentence: AI is evolving from a bunch of separate tools into what I call a Personal Operating Layer, or POL. A layer that sits between you and everything digital, remembers how you work, and eventually acts on your behalf.

The whole thesis depends on memory. Not memory as a feature. Memory as the thing that makes the POL possible in the first place.

If the AI cannot remember you, it cannot reason about you. If it cannot reason about you, it cannot act for you. The helper stays a helper and that is useful. But it never becomes the true actor you really want it to be. Without memory that compounds, there is no POL. There is just a very capable contractor who shows up without context every single time.

So when the a16z piece argues that current AI memory is hitting a ceiling, it matters. They are not describing a narrow technical problem. They are describing the gap between where AI is today and where the next generation must go. It’s not a nice to have – it is a must have.

What they are saying

a16z lays out three ways AI can remember.

The first way is what almost every AI uses today. Big context windows and retrieval systems. Essentially giant filing cabinets the AI searches through on every question. It works. But even a huge filing cabinet is still a filing cabinet. The AI must look everything up, every time. It never truly learns.

The second way is a middle ground. You take a general AI and attach specialized memory packs to it. Think of these like snap-on modules that give the AI expertise in a specific area without retraining the whole thing. This is probably where most enterprise AI lives in the next three to five years.

The third way is the frontier. Actually, update the AI’s own model based on what it experiences after it ships. Not just look things up faster. Actually, learn from use.

That third one is what a16z is betting on. It is also what would change everything. Much more than most may anticipate.

What I am seeing while building with OpenClaw

I want to use a concrete example to make this real.

I have been running OpenClaw on a Mac Mini. The job I am giving it is specific. Watch the news. Cross-reference what is happening against the ideas in my book. Flag what supports the thesis, what contradicts it, what is moving faster or slower than I predicted.

That use case is only valuable if the system remembers. It must hold my book’s arguments in its head. It must know which frameworks matter to me. It must remember what I flagged yesterday so it can connect it to what shows up today. Without memory that compounds, it is just a news feed with extra steps.

OpenClaw is built in a way that makes the memory part genuinely mine. The files live on my machine. I can read them. I can move them. I can delete them. That matters.

But here is what becomes obvious the moment you start using a system like this for something real. The memory being portable is not the same as the intelligence being portable. I have been tweaking the setup along the way, trying to make it better. The work of teaching the system what matters to me, how I think, which signals to watch for, that accumulates somewhere harder to see than a set of files.

This is the thing the a16z piece points at but does not fully connect. Memory is not just about what is stored. It is about what the system has learned to do with what is stored. And the deeper the system goes, the harder that is to extract.

Where I agree with a16z

The core argument is right. Looking things up is not the same as learning. A system that can retrieve any fact has not actually gotten smarter. It has just gotten better at searching.

Real learning means compression. Taking a huge amount of experience and squeezing it down into something that captures the pattern, not the specifics. Michael Polanyi called this tacit knowledge. His line, more than sixty years ago, was that “we know more than we can tell.” [3] A radiologist who has read scans for thirty years can spot something off in two seconds. She cannot always explain exactly what made her flag it. Her expertise lives in her visual cortex, not her vocabulary.

That kind of knowledge cannot be typed into a prompt. It can only be learned through experience, compressed into something the system internalizes.

If we want AI to operate at that level, with your business, your patients, your portfolio, we have to let it learn from what happens. Not just take notes on what happened.

Where I think it goes further than they said

A POL that can genuinely learn from use becomes a different kind of product. Not a better chatbot. An operating layer that compounds with your experience.

Right now, every interaction with AI starts close to zero. With real learning built in, year three is qualitatively different from year one. The POL that has been watching you work for three years understands your judgment, your preferences, your patterns, in a way a fresh installation never will.

This is why I think the memory moat is real, and why it is going to be the defining competitive question in AI over the next five years. Not the best model. Not the biggest model. The one that compounds.

The part nobody is pricing correctly

Here is where a venture firm is not going to say what I am about to say.

If AI learns from you, and that learning lives inside the model itself, we have a problem none of the current privacy rules were designed for.

GDPR says you have the right to be forgotten. HIPAA sets rules for how health records can be used and deleted. The EU Data Act requires that you can move your data from one provider to another. All of these assume deletion is a solved problem. You drop a row in a database. It is gone.

When AI learning lives inside the model itself, deletion is not solved. A 2025 research survey on machine unlearning is blunt about it. [4] A separate paper titled Unlearning Isn’t Deletion showed that even when a model appears to have forgotten something, a few minutes of retraining can often bring it back. [5]

Translation: the regulators think deletion works. The research says it mostly does not. Anyone deploying next-generation AI in healthcare, finance, or any regulated industry needs to understand this gap before they sign the contract, not after.

What I am watching

A few things, in plain terms.

Who ships learning-capable AI first, in a real product, not a research paper. That tells me when the POL starts feeling different from the AI we have today.

Whether deletion catches up with learning. If it does not, regulators will eventually draw a hard line somewhere, and whoever is already deployed will wear it.

Whether someone builds a layer that sits between you and whatever AI you are using, holds your memory in a format you control, and survives when you want to switch. That is both a defense against lock-in and one of the bigger business opportunities I can see forming right now.

Which industries break through first. My money is on financial services and healthcare, which sounds counterintuitive because those are the most regulated. But the tacit knowledge in those fields is also the most valuable. If real learning works in a trading desk or a radiology department, it works anywhere.

One last thing

My book argues that memory is the component that makes the POL possible. a16z just published the research case for why today’s memory is not enough, and where the frontier must go.

If you are running a company, advising one, or investing in one, the question is not whether this is coming. It is how fast, and what happens to everyone who is still buying AI based on last year’s architecture.

The AI that forgets you every time you open it is a product. The AI that remembers you across three years of work is an operating layer.

Those are not the same business. And they are not going to have the same winners.

References

[1] Aubakirova, M. and Bornstein, M. “Why We Need Continual Learning.” Andreessen Horowitz. April 22, 2026. https://a16z.com/why-we-need-continual-learning/

[2] Glorikian, H. The Invisible Interface: How AI Turns Intentions into Actions and Who Wins. Ideapress Publishing, distributed by Simon & Schuster. June 30, 2026. https://www.simonandschuster.com/books/The-Invisible-Interface/Harry-Glorikian/9781646872480

[3] Polanyi, M. The Tacit Dimension. University of Chicago Press, 1966.

[4] Li, Q. et al. “A Comprehensive Survey of Machine Unlearning Techniques for Large Language Models.” arXiv. Updated May 2025. https://arxiv.org/abs/2503.01854

[5] Xu, X. et al. “Unlearning Isn’t Deletion: Investigating Reversibility of Machine Unlearning in LLMs.” arXiv. Updated September 2025. https://arxiv.org/abs/2505.16831