Get more from AI without giving more of yourself.

You have probably done this already.

You open ChatGPT or Claude when you have a question. You type. You get an answer. It is often good. These days it even remembers you between chats, and you can read and edit what it remembers. Genuinely useful.

But notice what happens as you use more than one. Claude remembers one slice of you. ChatGPT another. Your coding assistant a third, per project. Each tool keeps its own notebook, none of them share, and none of them add up to one picture of your work. Switch tools and you start that notebook from zero.

And you still hold back the sensitive parts. The client detail, the board matter, the email you cannot afford to leak. Those live under one vendor's terms, on servers across the ocean. So the AI helps with drafts and quick questions, and never touches the work where it would matter most.

Many notebooks, no memory.

Each tool remembers its own slice. The picture of your work is scattered across vendors and apps, and no tool reads another's notes. The real leverage, one assistant-grade memory that every AI you use can draw from, never gets built.

Nothing checks what gets written. The model writes its own notes about you, and grades its own homework. When it remembers wrong, it is wrong quietly, and everything built on that note inherits the error.

No receipts. Two months later a customer, your board, or a regulator asks why you chose X. You remember using AI to help think it through. You cannot show what you actually saw, or what checked it.

The walls are vendor walls. You can read your memory, but moving your working context to a different AI is a manual job. And the confidential work still sits in a US cloud under someone else's terms, which is exactly why you keep it out.

One memory, every AI you use, on this side of the ocean.

Imagine one memory that every AI you use draws from. Claude today, whatever wins next year, several side by side. The history stays put; only the machines rotate. Your context stops being the reason you are stuck with one vendor.

Imagine the important stuff got a second opinion before it filed. When a session closes, a machine built by a different company checks the first one's work, the way a second doctor reads the same scan. It cannot approve its own homework, because it never wrote the homework.

Imagine every session left a flight recorder. What was decided, what checked it, what the machine actually said, timestamped, in a record that cannot be quietly rewritten. Three years later, "who approved this?" has an answer you can point at.

Imagine it lived in the EU and opened only with your key, the way a password manager does. Not even the system underneath can read it without you. Suddenly the sensitive work can actually come in.

Now imagine handing a whole task to an agent under that same discipline. A model from a different family reviews each step, and the hard limits are plain rules, written as code. Machine judgment where judgment helps. Plain code where rules must hold.

This is what Spegling is.

Builders know the shape of this failure.

The agent wrote the patch. You read the diff, the tests passed, you merged. Two weeks later there is a bug, and when you step into that code you realise you do not quite know why it was written that way. The code is yours, but the mental model is not. The friction that used to force understanding was traded away for speed.

That gap is the drift. Small each time, cumulative across a year. Every profession gets its version, and it is not laziness: it is what happens when every tool in your hand optimises to answer faster than you can frame the question.

In February 2026 the failure got two academic names from two disciplines arriving independently. Margaret-Anne Storey calls it cognitive surrender. Wharton's Shaw and Nave measured it: 80% of users follow AI's wrong answer when it is wrong. The diagnosis is settled. The architecture that takes it seriously is what comes next. Read the thesis.