Making Your AI Remember You
You explain yourself. It gets it. You open a new chat the next day and it has no idea who you are. Again.
That is not you being bad at this. It is a real, solvable design problem, and I went digging into the research from Claude on exactly how it gets fixed.
What I found was six moves. I put them in order for you here, from the thing you can do in the next ten minutes to the habit that pays off over a year. No coding. No engineer. Just you and the tool you already pay for.
What are you trying to fix?
Come back to this page any time. Find the sentence that sounds like your problem.
The amnesia problem
Every time you open a new chat, your AI walks in brand new. It does not remember yesterday. It does not remember you.
Here is the part most people miss, and it is the part that changed how I use these tools. The AI is not getting dumber. It is arriving smart and blank. Smart is useless if it does not know that you write in short sentences, that your audience is women in their thirties, or that you cannot stand the word delve.
That missing piece has a name. It is called , and it is the whole difference between a brilliant stranger and someone who has worked with you for a year.
When I went through the research, what struck me was how fast this all moved. The whole fix developed over about one year. It went from a sticky note to a filing cabinet.
You do not have to be technical to walk it. Every stage has a version you can do yourself, and I am going to take you through them in order.
Leave it a note
The very first fix anybody tried was almost embarrassingly simple. One plain text file sitting at the top of every session, telling the AI how you like things done. Your preferences, your way around a project, your rules.
The phrase used to describe how well it worked was unreasonably effective. Meaning nobody expected a sticky note to do that much.
There is a catch, and the research names it directly. That note keeps growing. Add every preference you ever had and eventually you are handing your assistant a novel to read before she can start work.
Write your one page
- Open a new Google Doc. Not a note on your phone. You will see why in move five, and it costs nothing to start in the right place.
- Title it “How I work.” Boring on purpose. You want to find it in six months.
- Answer five questions in plain sentences. Who am I. What do I make. Who am I talking to. How do I write. What do I never want.
- Keep it to one page. If it runs long, cut the nice-to-haves and keep the rules you would correct someone for breaking.
- Paste it at the top of a new chat before you ask for anything. Then ask your question. Watch the difference in the first answer.
- Add a line every time you correct the same thing twice. Twice is a pattern. Patterns belong in the doc.
The steps above are mine, not the research. I wrote them for how we actually work.
Coming in the full build
Let it take its own notes. Give it the bookshelf, not the whole library. Keep it in a filing cabinet it can search.
These three pages follow the same shape as move one, so you can already tell whether the format works. Skip ahead to move five and move six to see the parts you asked about specifically.
Grown folks’ guardrails
Everything up to here works beautifully until you scale it. Once you have many assistants running at once, or many people touching the same notes, things break in ways that are hard to see. So the research names the protections that keep it from falling apart.
One example should make you sit up. Picture a single confused assistant deciding to rewrite the rules every other assistant reads from. If what it wrote was wrong, that mistake now belongs to everybody.
Here is where I want to be straight with you. Four of these you can set up yourself, step by step, in an afternoon. The fifth one is deeper. I am not going to skip it, and I am not going to pretend it is simple.
Versioning
Being able to see what changed, and undo it.
Your version: keep your “How I work” doc in Google Docs. Version history is built in and free. File, then Version history. You can restore any earlier version.
Permissions
Deciding what can be changed and what stays locked.
Your version: keep two files. A master only you edit, holding your voice rules and brand facts. A working file for everything else.
Portability
Being able to take your notes with you.
Your version: never let your master file live only inside one AI’s memory. Keep it somewhere you own. Then switching tools costs you nothing.
Concurrency
This one only shows up once you are not working alone.
Your version: two people editing the same file at the same time will overwrite each other. One owner per document. Everybody else comments.
More on the deep one
Concurrency sounds the most technical of the four, and honestly, working by yourself you will never hit it. The engineered fix is a check the system runs before saving. Look at the file, write the change, look again, and if anything moved in between, start over.
You do not need to build that. But the day you hand your brand voice doc to a VA, you have the problem. My fix is the same idea done by hand. One owner per document, everyone else suggests, the owner accepts. That is it.
Lock it down in four steps
- Move your “How I work” doc into Google Docs if it is not there yet. That single move gives you versioning for free.
- Split it into two. Master file for the rules that must not drift. Working file for notes and drafts.
- Open your AI’s memory settings and read what it saved about you. Delete anything wrong or out of date. You are allowed to.
- Name one owner per document. If nobody has been named, it is you.
The four guardrails come from the research. The everyday versions and the steps are mine.
Let it dream
Here is the problem with everything you have built so far. All that note-taking happens inside one chat. So your AI never sees that it has botched the same thing fifty times across fifty different sessions. It walks in fresh every time, which means it cannot possibly know how annoying it is being.
The fix is a second process that runs separately, on its own time, with its own budget. It reads a pile of past sessions plus everything currently saved, and asks one question. What do these all have in common?
The name for it is .
The analogy that unlocked this for me is a school. Students hand in work. Teachers grade it. But a head teacher reads across every class at once. If every geography student bombs the same question, that is not fifty bad students. That is a missing lesson in the curriculum. And when every student answers in radians instead of degrees, nobody is dumb. Everybody’s calculator is set wrong.
That second one is a broken setting, not a broken brain. The pattern is only visible from above.
Be your own head teacher
- Put a recurring reminder on your calendar. Once a month, thirty minutes. Call it “AI cleanup.” If it is not scheduled it will not happen.
- Scroll your recent chats looking for one thing. Not mistakes in general. The correction you typed more than twice.
- Write those repeat corrections down. Three or four is a normal month. Each one is a hole in your instructions.
- Add each one to your master file as a rule. Not “stop doing that.” Write what you want instead.
- Delete anything in the file that is no longer true. Old client, old offer, old voice. Stale notes make it worse, not better.
- Check what your AI has saved about you and clean that too. Same pass, same thirty minutes.
What you get out of it
Three results were reported from doing this well. Better accuracy the second time around, because the mistakes are written down. Lower cost and less waiting, because it stops fumbling and gets things right the first time. And your own capacity back, because you are no longer correcting the same thing forever.
One honest note from me. There were no numbers attached to any of that. No percentages, no benchmarks. I am telling you what was reported, not what was measured, and I want you holding it that way.
The concept, the school analogy and the reported results come from the research. The monthly routine is mine.