How to give Claude lasting memory of your work (in plain files)

TL;DR: Claude doesn’t reliably remember you between conversations, and no built-in AI memory feature fully fixes that as of this writing. The way I make it remember my work is a folder of plain text files – who I am, what I’m working on, how I like things done – that Claude reads at the start of every session. That’s the entire mechanism. No plugin, no add-on, nothing to rent.

I’m Julie Kaiser, a science writer. I’ve used AI on real client work for three years, and this folder is the single thing that took me from re-explaining myself every morning to an AI that actually knows what I do. Here’s how it works, why I trust it, and how to start yours in a few minutes.

The memory folder is one layer of a bigger setup – the system-level view is in How I built an AI that actually knows my work.

Does Claude have memory?

Officially, yes – there are memory features, and Projects can hold files and instructions. In practice, the memory is scoped in ways nobody warns you about. When I started, I was using Claude chat, Cowork, and Projects all at once, and I assumed they shared what they learned. They didn’t. Chat taught nothing to Cowork. One Project couldn’t see another Project. I kept hitting the same confusion: how do I set this up so the right project knows everything it needs to know?

And there was a worse version of the problem. Somewhere in one of those projects, I’d figure something out – a genuinely smart way to do a thing. Then later, I knew I did something great. It was really cool, it was really useful, and I can’t ever find it again.

That’s the real memory problem. Not whether the AI remembers you. Whether your own work is somewhere you can reliably get it back. To be fair, I’m sure there are ways to adjust your working style and habits to make this work anyway, but I moved on to bigger and better things.

Why I keep the memory in files instead

Files answer both problems at once, and they’re mine.

A file on my computer doesn’t care which chat window I open. Claude Code, a fresh conversation, a different tool entirely – they can all read the same folder, so nothing I’ve figured out is trapped inside one product’s walls. And when I need something back, I don’t have to remember what I named it or which thread it lived in. I describe what it was about, and the AI finds it for me. My brain can’t hold everything I’ve captured over three years. The files can, and they can be searched semantically.

There’s a sturdier reason underneath the convenience. It’s just files on my computer. If something better ships next week that makes my current tool obsolete, I point the new one at the same folder and keep working.

How to set it up (start tiny)

Don’t set out to build an elaborate system from the get-go. Instead, write one file.

Open any plain text file and put three things in it: who you are, what you’re working on right now, and how you like things done. You’ll see people call this file memory.md or CLAUDE.md – the name doesn’t matter, the habit does. Mine started exactly that small. Then, every time you start a session, tell Claude to read that file first.

If you’d rather not start from a blank page: my starter file is a free download – first-file-template.txt. The same three sections, ready to fill in, with instructions for handing it to your AI.

From there, when a session produces something worth keeping – a decision, a draft, a way of doing things – save it to the folder before you walk away. I think of the chat as the workshop and the files as the warehouse. The workshop is where the messy thinking happens, and it’s disposable. The warehouse is what you keep.

In my system, most of that saving happens without me. My setup keeps a log of what each session did, and I close working sessions with a small skill I built called /harvest – it combs through what happened, pulls out the lessons worth keeping, and files them where they belong. None of that existed on day one, and none of it is required.

How does Claude remember context across chats?

Less than you’d hope – and understanding what happens inside one conversation will save you real frustration.

Every conversation starts cold. Then, as you talk, it builds a real working memory of everything said so far – within one session, it genuinely remembers. But that memory lives in the conversation window, and the window has a size. As the conversation gets long, older material gets compacted away to make room – quietly summarized, and many tools don’t show you when it happens. Performance drops with it. If you’ve ever agreed on something with an AI early in a thread, asked it to implement that agreement an hour later, and watched it act like the agreement never happened – that’s what you hit. Why the forgetting is designed in, and what a context window actually is, is its own post: Your AI forgets. Here’s why. And when you close the window and open a new one? Cold again, holding nothing, until something reloads what matters.

Across chats, there are built-in memories that try to bridge the gap – chat memory, Project memory, Cowork memory – and each one lives somewhere different, syncs differently, and can’t see the others. (It confused me to no end, until I had my AI write me a map of where every memory actually lives.) What they hold is a summary of you, refreshed on a cadence Anthropic documents as roughly every 24 hours – though telling Claude to remember something applies right away. Useful, and not something I’d build my work on.

The files flip who’s in charge of that. When I start a new session, I’m the one deciding what its memory is – and it’s the same every time, because the important things live in the files, and every session begins by reading them. I don’t waste context loading up a pile of stuff for it to pick through, guessing what might be important and what might not. The conversation stays free to be what it’s good at: the messy middle. Save aggressively (or let Claude do it in an LLM wiki like mine) and the forgetting stops mattering.

Can I move my memory from ChatGPT (or anywhere else)?

Yes, and this is where the file approach quietly beats every built-in feature. A product’s memory belongs to the product. A folder belongs to you.

If you’ve been using ChatGPT for months, it knows things about you – so ask it. Have it write out what it knows: who you are, your projects, your preferences, the decisions you’ve made together. Paste the useful parts into your files, fix what it gets wrong (there will be some), and you’ve done a memory transfer no export button gives you. The same move works in reverse, and it’ll work for whatever tool exists next year. Portable memory is a folder you already have.

What this doesn’t fix

Honesty section: the files only know what you save to them. Working in a chat window with one file, the saving is on you – the AI can’t reach into your computer and update the file, so you’re the one who has to notice the thing worth keeping and put it there. I’m not naturally disciplined – I’m a smart person, but I get distracted. And in the early weeks I didn’t trust it: I kept asking, over and over, did you save that? Did you save that? Willpower didn’t fix that; rules did. I made rules that make my setup save constantly as it goes – they grew into the logging and the harvest skill from earlier in this post – and I stopped having to ask.

And the AI still needs your judgment on every important thing it does with all that context. Memory makes it know your work. It doesn’t make it right. You cannot automate getting AI to produce work that meets your standards – at least not yet.

Start today, ten minutes

One file. Who you are, what you’re working on, how you like things done. Have Claude read it first, every session, and save what matters back to the folder before you walk away. Everything else I teach grows from there.