Transcription
Right now, all of us are building the most important asset of our careers in AI systems all over the place, and we're not owning any of it, and it's fragmented, right? We are building it in ChatGPT, we're building it in Perplexity, we're building it in Claude, and we are putting in our own context, our own thoughts, our own ideas, our own documents. And this is all a massive problem of fragmentation that nobody has an answer for.
We need a "bring your own context," a BYOC system for the enterprise worker in 2026. And we have none. Right now, all we have are IT departments that say, "Do not bring your personal AI in the door." We hate that. And we have corporate IT rollouts for tools that are supposed to be equivalent to what we use at home, but they're not. And the reason they're not is context. The reason they're not is they don't know us like we have our personal AIs that know us.
In other words, the bet that Sam and Daario have been making worked. The fact that we care about which AI instance we use is a function of their ability to build memory systems. And I don't just mean them, right? If you use Perplexity enough, you have this effect. I call it a honing effect. The system hones to you and your cognitive behavioral pathways the more you use it. There is really an intelligence at work here, and it adapts to you because it can look at previous context, depending on how the orchestrators of these systems have designed them. And they are deliberately designing them to make them sticky.
This is as old as Silicon Valley consumer habit loops. You build a product, you want it to be sticky, so you design things that sustain engagement. This is how Facebook is addictive. This is how Instagram is addictive. This is why people make big speeches on the floors of Congress about TikTok. Consumer products are designed to be addictive, and memory helps us make AI addictive. And so we have memory that helps us to do this work. And it turns out that when you make a product for consumers that's addictive because you use memory, you provide a really big side benefit to all of us workers. Because when we use it professionally, we benefit from all the previous work that we did. And so I don't say that just to sort of score cheap points. I'm not saying it's bad that memory exists in these tools. It's actually really helpful. It has the side effect for us of also making them very difficult to leave. And that is exactly why they're sticky, and it's why we need a way to bring our own context elsewhere with us, between tools, into the workplace, you name it. Our context needs to be much more mobile than it is.
So, in this video, I'm going to break down why that's a hard problem, what we can do about it, and how I'm building something to help make that easier. But first, let me be specific about what we're accumulating. Because it's very easy to just gesture at AI context and say, "Oh, it's a bunch of stuff. It has some economic value, etc." There are four specific layers of context. And I think if we don't understand what we mean by AI context, particularly in a professional capacity, we're going to misunderstand the problem. And the reason I'm emphasizing a professional capacity is because this is where the conflict is the sharpest.
If I want to migrate between my personal ChatGPT and my personal Claude, it's a little bit of a hassle because the company you're taking data out from doesn't want you to do that and doesn't make it super easy. Uh, but that's it, right? There's no policies in the way. The larger, more complicated question is how you actually separate out the elements of your professional context from everything else you're doing so that you can make that portable. That is a hard problem, and no one is addressing it well today.
So, what are those elements? We have to understand them in order to get into the problem space.
Number one, domain encoding. Over months of daily use, you have probably taught your AI your industry vocabulary, the company's products that you use, your market dynamics, your competitive landscape, the regulatory environment you operate in, the internal acronyms your team uses, the way your organization thinks about strategy. And if you're sitting there and saying, "I don't do that. I don't ever use my personal AI for company stuff." You are in the minority. More than 60% of workers surveyed say they use their personal AI at work. Most people do, whether or not the IT department approves, and they do partly because of the context problem. And I think we need to acknowledge that. And I rarely, rarely, rarely see people recognize it for the issue that it is.
Most people misunderstand AI. They think, "Oh, it's just a random interchangeable tool. You get Claude on this computer. You get Claude on this computer. They're going to be the same." They're not going to be the same because of the context, and domain encoding is a part of that context. And part of what makes it hard to move and recognize is that you didn't give all of that domain information to your AI by sitting there and writing it a briefing document. Instead, you gave it that information daily, in little bits and pieces, over the course of hundreds or thousands of conversations. And that means you may not fully realize all the stuff you gave. And that is part of what makes it hard. Is because if we were to sit down and to write out, "This is all the context that we have laid out over months and months and months of AI," we couldn't do it. I couldn't do it. You couldn't do it. It's too hard and too complicated. We don't know the context we know. And that is part of what makes it hard to move.
This is functionally equivalent to the institutional knowledge that used to live in a senior employee's head. It took years to build in the old model. You learned your industry through experience, through osmosis, through mentorship, through hanging out with the senior engineer at the water cooler. That was the only way to learn. With AI, that encoding is happening faster because the only way we can interact is by writing things down explicitly in conversation. So that is faster. You get years of progress in months if you're intentionally writing things down. That's part of how we see people make progress quickly. But if you start out fresh with a new AI, anybody will tell you it feels like you're talking to a stranger. You still don't have that domain information. And even if you could accumulate it quickly, because it's entirely a verbal pattern, it still takes a long time. It still sucks. And the more it sucks to use a new AI, that's a sign to you that you've done a great job encoding that domain knowledge into your existing AI. Right? Good job. Now, it's hard to move. So, let's make that easier.
The second layer of context is workflow calibration. Beyond knowing your domain, your AI has also learned how you work, how you like your research structured, how you want your code reviewed, what a good first draft looks like for your documents, the sequence of steps you follow when analyzing a new problem, a new market, the format you want your internal memos to be in, how you like your Slack summaries published. You've established these patterns through a lot of repetition, through a lot of edits you've given, through the way you've held a high bar. This is why, by the way, I emphasize holding a high bar when you talk to AI. The higher the bar, the more you are encoding that bar, so it's easier to hold it over time. When you open a new conversation with an AI, you know, the AI already knows the shape of what you're going to ask based on previous conversations. It just does. And we don't realize how much of a help that is till we try a new AI. And it feels like we're grinding in first gear on the car. And yes, that really saves you time. A calibration can save you five, six, seven, eight turns of conversation because the AI is more likely to get it right the first time because it knows what you want.
Now, the third layer is the behavioral relationship, and that's the one that matters the most, but I think it's also the hardest to articulate. So, I'm going to try to be very precise here. This is the emergent understanding your AI develops about how to work with you specifically, not your stated preferences. Those are very easy to write down. The unstated preferences. When you need to be challenged versus when you need to just execute and get stuff done, how technical to go, when you, the prompter, did not specify a level, whether a question that you ask is an invitation to talk or whether it's just rhetorical, how much preamble you tolerate before you want to get to the point. These are things where the AI is inferring how you want it to behave based on your response to its responses. And this happens very subtly. It is almost never explicitly instructed, despite what people may tell you, and it's very, very difficult for you to see yourself doing it. It's one of those things like your nose. You don't see your nose. Your nose is right in front of your eyes. Your eyes block it out. Same thing. You don't see the way the LLM is shaping itself to your patterns of response, but it is.
This layer is built through hundreds and hundreds of microcorrections, right? The times you rephrased a prompt because the first output missed the point. Uh, the times you said, "No, no, no, more like this," and provided an example. The times the AI did get it right, and you kept going, and you didn't give it a comment. It's like the compound interest on a relationship, except in this case, it is the compound behavioral difference the LLM learns from interacting with you over time. And so this is like the difference between working with a colleague you've known for a year and working with a brand new colleague. Right? That the colleague you've known for a year knows exactly how you are going to respond when faced with a particular situation. That is not true of your new colleagues. They have to learn that. Same with AI.
The fourth layer is the artifact layer or the demonstrated capability layer, and it doesn't really exist today. So right now, when we do work with AI and we produce artifacts, right? We produce documents, we produce spreadsheets, we produce PowerPoints, we produce code, all of it ends up as artifacts we copy, paste, and move somewhere else. And frankly, who knows where it all gets to. What we are missing, if you are a professional, is a way to say, "This is an artifact that I made. This is how I made it, and this is how I know it was good." And you want to bring the context around that artifact, around your intent, over with you to your next role. Not because you want to steal trade secrets, not because you want to take information, not because you want to copy it verbatim, but because you want to take the encoded thinking, the encoded rationale, the way you taught the AI to think through pros and cons, which you can take and move to a new role, and you want to bring that with you. That is how you demonstrate capability, right? And and one of the things that we look for when we're interviewing is this demonstrated capability. We don't care if someone brings their exact strategy over from the previous company. In fact, we don't want them to. That would be inappropriate. What we want them to do is show us they can build that, that they understand how to build trade-offs into that strategy, and then they can do a competent job with us. That's the whole point of interviewing. Well, increasingly, we do that work in AI, and when we want to talk about how we did it, it's buried in some chat somewhere. So, good luck finding it, right? Like, it's hard to explain what you did because it's all emergent in a series of like 800 chats, and you have to go back and find it, and oh no, did they store the document? Where did you put the document? It's a mess.
So there you have it. The four layers of the problem we have to solve. Domain encoding is one. Behavioral change is another. How you coordinate workflows is another. And then finally, understanding how to communicate your thinking process around artifacts is a fourth. These are all really important pieces, but because we don't break apart the problem, I think we look at it as just general context. And that makes it really hard to be specific about a solution.
If you're sitting here wondering, "Is this a real issue? Is this only for people walking into specific companies?" It is an almost universal issue, and it will affect all of us in the next couple of years. Not because there's some magical wave that's coming for all of us at the same time, but because on average, over the next two years, are you likely to look for another job at some point? Are you likely to be on the hiring market at some point? Are you likely to switch AIs or need to add a new AI at some point? The answer to one of those is probably yes. Sometimes it's an unforced job change. Sometimes it's a job change where we're looking for more. Sometimes, frankly, the company brings in a new AI, and it wasn't a job change, but we have a responsibility change. You moved over, now you have a new AI regime. A new agreement was signed with Anthropic and not OpenAI. The company AI changes. All of these things trigger this problem. This is a problem that I would bet you a lunch affects 90% of us in the next two years if we're in the professional workforce, one way or another. Whether we are switching AI at work because the company told us to and its policy, whether we're doing it on purpose because we got permission to use a new tool, whether we're switching roles and that led to an AI change, whether we are switching to a new company and there's a new AI regime, whether we got fired and have to use our personal AI tools, but we want to bring our work context or vice versa. All of us are in this boat together.
And to be clear, this is a genuine market failure. I don't want to put this on us as individuals. Employers want to hire AI-capable people and have no reliable way to evaluate that capability or understand what they bring. Candidates want to demonstrate AI capability and have no structured way to talk about that or show it or bring that context in with them. And the credential gap is kind of being filled by vibes, right? Like you say, "Oh, I can do all of this stuff," but how do we know you can do all of this stuff? At this point, there are major companies, including Meta, that are resorting to flying you in on their dime and locking you in a room with their laptop and their tools to see what you can do. And we are going to see more of that because we just don't understand how to use AI because bringing context with us is so hard. And by the way, in that example, you don't bring your context with you. And so we're just assuming that you can use these new tools from scratch to generate extraordinary results when really you would want to say, "I've generated these results partly because I've honed this tool as a working companion, and now I'm going to have to switch companions, and it's going to feel like I lost a leg for a few months." If I were there and I had to switch AI tools, even if I knew Claude code well, which I do, and I had to move into a new environment and it was a brand new Claude code, and I had to do magical things with it, it would take me longer. I would underperform relative to where I actually am because I don't have that context.
Why has nobody solved it on the platform side? I think the reason nobody has solved it is pretty simple. It's incentives, right? None of the model makers has an incentive to solve this problem. They all want to keep you inside, right? None of them want to lose you. Every single platform makes it easy to get context in and relatively hard to get context out. And none of them are doing any kind of a job at helping you separate personal and professional context out or helping you separate professional context from trade secrets and things that are inappropriate to bring to a new workplace.
Now, you might think, "Well, this is ripe for a third-party build, and there are well-funded VC startups going after this." And so, the answer here, the reason why we haven't broken through, I think it's more interesting, right? It's not a cash problem. I think it's that memory layer startups are failing to capture people where they feel the pain point. Right? This is a diffuse pain point from a product strategy perspective. You feel it all the time. It hurts, but it doesn't hurt in the dramatic FOMO, "my tire is out on my car" kind of way, right? It is the kind of, "there's a funky sound in the car" kind of a pain point. And it might be your engine giving out and costing you $20,000 miles, but you don't really think of it that way. You think of it as, "I'll drive it as long as I can until I have to take it to the shop." It is a diffuse pain point. Memory and how we use memory sucks when we have a new AI to use. It sucks all the time. It sucks every chat. It may affect how often we use it. It certainly affects our performance, but it is diffuse enough that are we really going to go and seek out a third-party tool? If the third-party tool exists, does it provide the linkages we need between the major platforms? Does it bring our personal context with us? Is it going to diverge trade secrets I don't want to take to the new place? Like, there are all of these questions that none of these memory providers are really answering. Plus the fact that it's not a specific, highly painful moment in a single point in time, which is usually the sign for a good product. Like, we talk in product strategy about like candy products and opium products. This kind of third-party memory tool is very much a candy product. And what I mean by that is it's great. It's nice to have. We like it. We're glad we have it. We're not going to go out and seek it because it doesn't immediately solve the thing that makes us really suffer. Whereas an opium product, pardon the crassness, is the kind of product that we must take to cut pain off because it's so, so, so painful. It's acutely painful. This is not an acutely painful product problem. And so memory startups are struggling to solve it.
So, given all of this, given the platform hostility, given the injection challenges, given the calibration gap, the diffuse pain, what would it actually mean to own your own AI working intelligence over time, to treat your professional context like an asset? That's the question I'm asking. That's the thing I'm building against. That's the thing I want to challenge you to think about. I think it starts with a shift in mindset. We need to treat our AI context as a professional working asset that we will nurture for the rest of our careers. Period. End of sentence. It's not something that happens to you inside platforms. It's not something platforms should own. It's something you should build and carry.
And at a practical level, this means your working identity does need to live somewhere that you can control. The simplest version, and the one available to you today, is a structured document. Right? A markdown file is the very simplest version of this. It's not probably good enough to get the job done, but a well-crafted markdown file, put together properly, can capture a fair bit of domain context, communication preferences, workflow patterns, recurring projects, style. And you can audit it and make sure it doesn't include any uh, inappropriate detailed information from your previous work. And you can generate that with help from an AI that knows you best, right? And you can review it, and you can edit it, and you can paste it wherever you need to go. And it's sort of a band-aid. It helps you to close the gap a little bit, and it's certainly positive ROI if you're willing to spend 30 minutes on it, and it's probably the simplest fix out there.
I would say the stronger version of this is a personal context server, right? An MCP-native memory store that any compliant AI agent can discover and query against on demand. This flips the architecture from push-based, where you paste a document in, to pull-based. The AI retrieves what it needs for the current task. The AI can query your context selectively. It can pull in your B2B pricing heuristics, for example, uh, or your editorial preferences, or whatever it is. And it doesn't just dump the entire profile into every single conversation window, which is very token-heavy. And critically, the AI can also, if it's constructed properly, write back. It can say, "You know, the user prefers this format for strategy docs," as you yourself evolve over the course of your working career. Because isn't that the point of a professional asset? Is that if it, if this is an asset that records our professional growth, it should grow with us. And I think a database does a better job of that than just a working document.
And if you're wondering, "Am I building both?" You got that right. I am 100% building and launching prompts that help you to elicit the right working information from the AI that knows you best. I am building a structured output format so you can get them into a markdown if that's the band-aid that you think you can use. And I'm also building a solution that plugs into OpenBrain that lets you pull together your professional working context into a single database that an MCP server can hit. Because I want you to have options, right? And if you want to listen to this and paste this transcript into AI and say, "Screw OpenBrain, I can do it myself." More power to you. I want you to have this as an asset that helps you to build your career over time. And shout out to MCP. Part of why this whole idea is viable is because the model context profile is effectively the USB-C connector for AI. Everything plugs into it. It is a universal connector, and that means even your work AI can plug into it. Now, I know that there are some work AIs that shut off those external MCPs as part of their special security policies. One, I think that most AIs to do their work need to be a little bit less shut down than that, and that's a policy conversation. And two, if it is that shutdown, you can still use a markdown file and get a little way with that.
So, if you want to make this yourself, or if you want to follow my guide to making this, I'm going to give you a really clear rundown here. Start with extraction. The AI that knows you best already has a rich model for how you work. You can ask it to articulate that model through a series of structured prompts. Your domain context can be extracted. Your communication preferences, ditto. Workflow patterns, style, recurring projects, behavioral patterns that it's observed about you. Maybe you've never asked them. Run a structured extraction prompt. I can provide you with one against your primary AI. Review and edit the output, and you have a portable representation of your entire working identity. It's not perfect. It's not a full behavior calibration. Think of it as a 720p video. Like, it's pretty good. It's not 4K, but the capturable layers, right? They can all be there, right? It can capture the domain encoding you do. It can capture the workflow calibration you do. It can capture both stated and observed preferences. And that's substantial. Having them in a document you own is a fundamentally better position to be in than having them locked in a platform's memory system. It is a win by itself.
The second big piece, and I think this one's really important, is to not be satisfied with a document, but instead insist on writing to evolvable infrastructure you control. You know how we had personal websites in the 2010s, and that was the thing? Like, you have to own your own domain name, and your domain name is your front door to the internet, and we'll all link it on our LinkedIn profiles, etc. Well, your personal database is kind of going to be that for the 2020s, because data is what allows you to bring this context with you reliably. You need to write to infrastructure you manage. And so write your extracted profile to a persistent database. And yes, I have a plugin for this. I'll show you how to do it with embeddings that you can own and a retrieval system you can host. This can be a local Postgres instance. It can be a Supabase project. It can be a VPS. It's whatever you want to run, right? The point is that your canonical working identity is important enough to put in a database now, because guess what? This is the way the web is going, and it should live with you, not inside a platform a company owns.
And if you're wondering, "Is it that difficult?" I have had thousands of people build the OpenBrain system. This is based on including many non-technical people. It is not that hard. You can totally do it even if you're non-technical. I think the larger point is really simple. Your assets are now complex enough to justify a database from a digital perspective. Your assets in the digital space are not just a thin paper resume. It is actually the accumulated rich pattern of working that you bring to the table whenever you do a job. And that has had no hope.
And then, of course, the next piece is exposing it via MCP. It needs an outbound, right? And this is the part that makes everything else compound, right? Because MCP is not just a de facto open standard by which AI operates. It's also a read-write standard. So, it allows you, once you expose it via MCP, to plug it into any AI you're using and to start to both read and pull in the stuff you need, and also write and evolve the database over time, which allows you to have an evolvable, portable working memory for your professional life. And that's kind of a big deal. Because the next time you get stuck with a new AI and they're like, "Good luck with your new Claude. Bless you. Off you go. Be just as productive as with the old ChatGPT that knew six months of history about you." You actually have an answer for that.
And this, this happens so much. It's so painful. And I don't know why we have not had a larger conversation about it. I've looked around, and most of the conversations that we have had have been focused on either switching between ChatGPT and Claude, or have focused on the idea that to solve this problem, you need to have an eagle eye on the screen on your personal computer or your work computer all the time. In other words, the only way to solve it is to have some sort of all-seeing photographic memory that screenshots your screen every 10 seconds, and that's how it knows what you're working on, and that's how it brings the context layer in. I find that a little creepy. I don't know about you. I don't think that's the solution that I would go for. I understand why people think that's the be-all and end-all solution because it feels complete. But I don't think it's correct. I think the more correct solution is to go back to the behavior we use to express these preferences in the first place, which is in chat with our AI, and to actually extract it from there in a reliable way. Basically, pull back from the AI you're used to working with the preferences you have implicitly encoded, and get those into a space and a format that allows you to be in control. And I think that's a much more reliable way to do things. And I think MCP allows us, as individuals, to propose a solution to the industry that doesn't force us to wait on some startup, doesn't force us to wait on the big players to make memory interchangeable, which I don't think they will do absent a congressional ruling, right? They just have no incentive to do so.
It also critically allows us to not wait on IT departments. Because if I had to say one corner of most companies that tends to be most cautious about AI, it's the IT department. They have a lot to lose. And they have a lot to lose partly because the burden for mistakes in security falls on the IT department. They're the ones at risk. They're the ones with their necks on the line. And there's a lot of risks that go with AI deployment. So, it's really natural for them to say, "No, you can't bring your AI. That's too risky. Use ours." But I don't think most of them really realize the difference in performance you get when you work with an AI that knows you well. People who work with AI a lot and who have an AI that knows them well will tell you they're two, three, four, five times more productive working with the exact same AI. Claude, if it's Claude, but it's the instance that knows them. And all of those encoded preferences help them be that much faster. And I think companies don't realize that that has real work implications, and they should. And I think something like this, having a responsible solution like this where we say, "We're not trying to port over company secrets. We are trying to just bring the working context that helps us understand how a particular employee works with a particular AI and bring that working style over in a responsible way so that they can work effectively." That feels like a really productive response to that, right? That's that's a grown-up response to that reasonable concern. And I think it opens the door to a more permissive IT policy.
So that's it. That's the architecture. You can, you can generate it with a prompt, which I'll give you. You can write it to a database you can control. You can expose it via MCP, and you're kind of off to the races. And yes, I'm building tooling to make this easier. I'm building extraction prompts. I'm building profile specs. Building implementation guides because I think the faster this practice spreads, the stronger the pressure will grow on platforms and on companies to take the idea that you deserve to bring your professional context with you seriously. And I really get on my soapbox about this because I think that this is one of those things where we are effectively taking away a large part of our professional lives and making it difficult for us to continue a seamless path of growth in the AI age when we really need to be growing fast. That's what everybody tells us, and it's harder now. And so we need to have solutions that allow us to bring our own context with us.
Look, if I step back, I think we're in the early years of a shift in what professional capital really means. And I don't think most of us have absorbed the implications, including companies, right? For decades, your value as a knowledge worker came from four things, just four. What you knew, what you could do, who you knew, and what you could prove that you'd done, right? Skills, abilities, network, and track record. All of it accumulated over the years. All of it was portable in the sense that it lived in your head, in your relationships, and in your reputation. Employers could benefit from your professional capital while you were there, and they could not really keep it when you left.
AI is creating a fifth category of professional capital, right? Your working intelligence, the accumulated context and the calibration that makes you effective with AI tools. And this category has a property that the others did not. It occurs outside your head, inside systems controlled by third parties who have a direct financial interest in keeping it there. Your skills live in your head. Nobody can lock them up. Your network lives in your relationships. LinkedIn can make it easier to maintain, but it can't prevent you from knowing people. Your track record lives in your reputation, in your references, in the memory of the people you've worked with. Your AI working intelligence, though, that lives on servers that belong to Anthropic, to OpenAI, to Google, to Microsoft, fragmented across accounts that cannot talk to each other, governed by terms of service you did not negotiate and subject to change at any time.
This is new, and the newness of it is why most of us haven't really reacted yet. I think we don't have the instinct for this that we have for the others. When you build a skill, you know you own it, right? When you build an AI context, the ownership question seems sort of unsettled. It may not even occur to you that it's an asset you should own until you try and take it somewhere or wish that it was there and it's not anymore. I think the professionals who recognize this now and act on it, building their own portable context, maintaining their working identity deliberately as if it was a career asset because it is, you're going to have a meaningful and compounding advantage as AI keeps getting better. Not because the tooling is perfect today. It's definitely not, but because the habit of ownership itself is a valuable recurring skill to have. The person who maintains a portable AI working identity moves between tools with less friction, switches jobs with less context loss, and compounds their AI effectiveness across platforms as they grow in their skills instead of starting over every time they cross a boundary.
Ultimately, I think memory has replaced models as the moat of 2026. The platforms that built their retention around your accumulated context are winning right now. The question is whether you keep pouring that context into walled gardens and hoping for the best, or whether you start building the professional asset that travels with you regardless of which AI, which account, or which employer you're working with next year. I know which bet I'm making. Best of luck.