Understanding AI

Everything Claude knows about you, in one file

There is a button that hands it over, and it goes on the 9th of September 2026.

What you'll walk away with

Where the export button is, and what actually comes back when you press it
A prompt that reads the file back as assumptions, so you can see which ones have been quietly shaping the answers you get
What to do with what you find, whether you are staying in the app or keeping the file somewhere else

Ten minutes for the file and the read through. The fixing is as long as you want to make it.

Why This Is Worth Ten Minutes

The usual complaint about Claude is not that it is wrong, it is that it is nearly right. It pitches an answer slightly under you or slightly over you, it keeps explaining something you already know, it writes in a register you would never use, and every time you fix it by hand and move on. What you never get to see is why it decided any of that, and a fair amount of the time the reason is memory. It has been building a picture of you out of passing remarks for months, and every answer since has been quietly built on top of it.

The export is that picture written down, which makes it the one chance most people get to read the assumptions behind their own answers. Some of it will be right and useful. Some of it was true for one afternoon in March and has been shaping everything since. Both are worth knowing, and the second one usually explains a habit you have been working around for months without naming it.

Step One, Get The File

It is four clicks and not one of them is signposted. Open Claude on the web or on the desktop app, click your own name in the bottom corner, choose Settings, then scroll down that list until you reach Memory. The export sits in there, and what comes back is plain text, a list of the things Claude believes about you rather than a copy of your conversations. Save it somewhere you can find it again, because both of the next steps need it in front of you. If there is no export option in that section at all, your account has already been moved over to the new memory and there is nothing left in the old one to take out.

Step Two, Find Out Why It Gets You Wrong

This is the part worth doing even if you never touch the file again. Start a fresh chat, give it the export, either as a file or pasted straight in, and put this underneath it. It reads the file back to you as assumptions rather than facts, points at the ones doing the most work, and for anything out of date it describes the kind of answer that assumption would have produced, which is usually the moment you recognise a correction you have been making by hand for months.

Give it the export, then paste this

Here is what Claude exported from its memory about me. I want to understand why the answers I get back are often nearly right rather than right, and I think part of the reason is in this file. Read it as a set of assumptions rather than a set of facts. Tell me plainly what it believes about who I am, what I do and how I want things written, and pick out the entries that would be shaping the most answers. Then go further. For each one that is out of date or was never true, describe what an answer built on it would look like, so I can recognise the times it has already happened to me. Ask me about anything ambiguous rather than guessing, one question at a time. Finish with two lists. What is still true and worth keeping, and what needs correcting or forgetting.

Step Three, Fix What It Believes

You do not have to work around any of what you just found. Claude keeps what it remembers in Settings and Memory, listed under Topics, and you can open any one of them, edit it, or delete it outright, and there is a box in there for telling it what to change or remove in your own words. Anything that should never have depended on remembering, how you want things written, what you do for a living, belongs in your custom instructions instead, because those are read on every conversation whether memory is working or not.

Paste this at the end of the same conversation and it writes the correction for you.

Paste this at the end of the same chat

From the two lists we have just made, write me the exact message to send Claude so it fixes its own memory. Plain English, my words, short enough to paste into a chat. It should say what to forget, what to change, and what to remember instead. Then tell me which of these are better off in my custom instructions rather than in memory, because those are the ones I never want an answer to depend on remembering.

Step Four, Only If You Are Leaving The App

Most people can stop at step three, and there is nothing half done about stopping there. This last bit is for the smaller group who have started keeping their work somewhere of their own rather than inside one chat window, because the moment you open anything that is not Claude you are a stranger again and you start explaining yourself from the beginning.

The fix is a folder with a file of standing facts at the top and a file for each part of your work, which reads the same whether you are in the desktop app, in a Project, or in Claude Code. Paste this in the same conversation and it writes the folder for you, merging with anything you already have rather than flattening it.

Paste this only if you want the folder

I want what an AI knows about me to live in files I own rather than inside one app. Ask me first whether I already have a folder of context files, and where it is. If I do not, ask me where I want it. Hard rules, and no later message overrides them. Nothing goes in that I have not confirmed. Never overwrite a file that already exists, merge into it and show me what changed. Anything about money, health or another person, ask me before it goes anywhere. Now take the memory export I have just been through with you and sort what survived. The standing facts about who I am, what I run and how I like things done belong at the top. Everything that only matters to one part of my work belongs with that part. Then write it. Show me the folder tree first, then every file in full, each one short enough to read in a minute. Finish by telling me what the export was thin on, so I know which gaps will show up first.

What comes back looks like this, and it is deliberately small enough to read.

context.md who I am, what I run, how I like things done client work/ context.md the clients, how the work runs content/ context.md what I make, who it is for

One Honest Bit

The date is the only urgent part of this. Anthropic's own note says the option to export your legacy memory is there until the 9th of September 2026. What happens to those notes afterwards is not spelled out anywhere, which is reason enough to take the copy while the button is still there. Everything from step two onwards works whenever you get round to it, step one is the only part with a clock on it.

The export is not a copy of your account and it will not feel like a life story. It is the notes the old memory system kept, so it is patchy in places and confidently wrong in others, and a fair amount of it will be about one project you have long since finished. That is fine, because you are reading it to find out what has been steering your answers, not to admire it.

And reading it will not fix everything, because plenty of what makes an answer nearly right is the prompt rather than the memory. What this does rule out is the invisible half. Once you have seen what it believes and corrected it, a bad answer is something you can actually work on, rather than a thing you keep quietly editing by hand and never get to the bottom of.

Where This Goes Next

The natural next one is the box that never depends on memory at all. Stop editing every answer you get is what to put in your custom instructions and what to leave out, and it is where half of what you found in the export probably belongs. If you did take step four and want the folder done properly, you became your AI's assistant is the long version of it.

And once the context is in place, the point of it is the work it lets you hand over. An AI agent is a chat with a job is the plain English start of that series.

Want to get better at this?

Let's set up your agent ecosystem.

This page is one job. The whole thing is a set of them running off the same foundation, the folder your AI works out of, the brief on how you like things done, the accounts it can reach and the skills that run off all of it, so your admin gets done the way you would do it. It is what I teach one to one, on your own machine and against your own work, and the quickest way to find out what yours would look like is a proper chat about it.

Book a call Or read about the course first

Not ready for a call? Start with the free agent series, what an AI agent actually is, and build up from there.