I read a piece by Nate on his Substack this week that I haven't been able to put down. It's called The Six-Month AI Context You Lose Every Time You Switch Tools, Jobs, Or Employers, and the thesis lands in the first line: the working relationship you've built up with your AI tools is a new category of professional capital, and it's sitting on someone else's servers.
I want to summarise the argument, because it deserves to travel further than a single newsletter, and then say where I think he's right, where I'd push back, and what I've personally started doing about it.
The thesis
Every day you use ChatGPT, Claude, Gemini or whatever else, you're building something underneath the outputs. Not the documents and the code themselves — the thing that makes the documents and code come out faster and sharper than they did six months ago. Nate breaks it into four layers:
- Domain encoding. The AI has learned your industry, your clients, your project names, your recurring problems. It stops asking what "the nursery rebuild" means.
- Workflow calibration. It's learned how you like things formatted, how long your emails should be, when you want a one-liner versus a full breakdown. Each conversation starts five to eight turns ahead of a cold one.
- Behavioural relationship. It knows when to push back and when to just execute. It knows "tighter" means fewer words, not more formality. It's internalised preferences you never actually stated — learned by watching which drafts you accepted and which you rewrote.
- Artifact history. It's seen what you've shipped. It has a working model of your actual capability rather than what you claim on LinkedIn.
Stack those together and you get something with real economic value — the difference between a first draft that's 40% right and one that's 80% right, repeated across every task, every day, for months. Nate calls it working intelligence, and argues it's a fifth category of professional capital alongside skills, network, credentials and reputation.
The uncomfortable part is that you don't own any of it. It's distributed across five different AI tools, locked inside accounts you can't merge, invisible to anyone evaluating you, and largely abandoned the day your employer switches vendors or you change jobs. The flywheel you've been turning spins inside someone else's house.
Where I think he's dead right
The honing effect is real. I've felt it personally. The AI I talk to every day meaningfully outperforms a freshly-logged-in instance on my own work — not because the underlying model is different, but because a few months of context give it a running start. When I open a new tool and it doesn't know what my projects are, which of my domains is which, or why I care about a particular stack decision, I feel the drag immediately. Nate's claim that switching AIs currently feels like "losing a leg" is slightly dramatic, but directionally correct.
The ownership point is also uncomfortable and true. The terms of service on every major AI platform reserve the right to change what gets remembered, how it gets remembered, and whether it'll keep getting remembered at all. You are building professional infrastructure on rented land. If you're a freelancer or small agency owner — as I am — that's a strategic risk you probably haven't priced in. If you work for a large company that might mandate a single enterprise vendor tomorrow, it's a risk you can't price in even if you wanted to.
And the framing that memory replaced the model as the moat is, I think, the single most important shift of the last twelve months. We stopped arguing about which model is smartest some time ago; the frontier models are close enough that the delta between them matters less than the delta between "a model that knows me" and "a model that doesn't". The moat is the relationship, not the reasoning.
Where I'd push back
Two small things.
First, the four layers aren't equally portable. Domain encoding — the facts about your work — is relatively easy to lift and shift; you can dump it into a plain text file and paste it into any model. Workflow calibration is harder but still achievable with good prompting. Behavioural relationship is the genuinely tricky one, because a lot of it is implicit — the model has learned your preferences by watching thousands of tiny corrections you didn't even register giving. You can't extract that with a prompt, because you don't know what the prompt should say. Some portion of what you've built is honestly trapped, and no MCP server is going to un-trap it.
Second, the "memory startups have struggled" framing slightly underestimates what's happening on the open-source and standards side. MCP as "USB-C for AI" is a glib comparison, but the direction of travel is genuinely encouraging. A year ago there was no serious plumbing for portable context. Today you can stand up a personal memory server that multiple major clients will speak to. That's a long way from solved — but it's not a dead end either.
What I've actually done about it
Nothing particularly clever. Mostly this:
- Export everything, on a schedule. Both ChatGPT and Claude let you download your entire history from settings in about two clicks. Whatever happens to the platforms tomorrow, I have the record.
- Keep a plain-text "about me" file. Nothing fancy — projects, clients, domains, stack preferences, a few hard rules about how I like things written. I paste it into new tools on day one. It takes ten minutes and skips about a month of onboarding.
- Notice what the model does right. When a response lands perfectly first time, I try to work out why and write the rule down. That's the behavioural layer — the only way to capture it is to catch yourself being happy and reverse-engineer it.
- Read my own exports occasionally. This is why I built Export Reader in the first place, but I'd have said it anyway: skimming six months of your own AI conversations is a surprisingly good way to see patterns you didn't know you had, and to extract the implicit stuff into something explicit you can carry.
None of that is a full solution. It's about 70% of one. But 70% of a portable working identity is considerably better than 100% of one locked in a box you don't have the key to.
The bigger point
The most useful thing about Nate's piece, for me, is that it names something I'd half-noticed but hadn't put into words. I knew switching AIs felt annoying. I hadn't articulated that what I was feeling was the friction of leaving capital behind.
If you use AI seriously every day, the question isn't whether you've built something valuable. You have. The question is whether it'll still be yours in three years, or whether someone will have quietly changed the terms while you weren't looking.
Worth a read, and worth thinking about what you'd lose tomorrow if your current setup disappeared. Link again: Nate's Substack.