Every time I open a new chat with ChatGPT, it forgets I exist.
Not in a dramatic way. It's polite. It's helpful. But it has no idea who I am, what I'm working on, how I like my answers, or what we figured out together last week. So I re-explain myself. Again. "I'm a writer, I prefer short answers, I'm working on a book about X, please don't use bullet points for everything." Five minutes of throat-clearing before I can actually ask my question.
I got tired of being a stranger to my own assistant. So I built a fix that takes about ten minutes to set up and needs zero code, zero plugins, and zero subscriptions. It's just one short document. Here's exactly how it works.
Why it forgets in the first place
Quick reality check, because it matters: chat AIs are designed to start blank. Each new conversation is a clean slate. That's not a bug, it's how they keep your chats separate and private.
ChatGPT does have a "memory" feature now, and Claude and Gemini have their own versions. They're fine. But they're a black box — you can't really see what the AI saved, you can't edit it cleanly, and it absolutely does not travel between apps. What ChatGPT remembers, Claude has no clue about.
I wanted something I could see, control, and move around. So instead of relying on the AI's hidden memory, I keep my own.
The whole system: one doc and two habits
The trick is embarrassingly simple. I keep a single short document — I call it my Memory Doc — in a notes app. That's it. One file.
Then I do two tiny things:
- At the start of a chat, I paste the doc in and say "Here's our memory. Continue."
- At the end (or whenever something important happens), I say "Update memory," and the AI hands me a fresh version to save over the old one.
The AI does the actual work of keeping it tidy. I just save the file. That loop is the entire system.
What the doc actually looks like
Keep it to one page. Mine has three sections:
# Memory Doc — Sarah
## Durable
- Writer, based in Lisbon. Prefer short, direct answers. No bullet points unless I ask.
- I'm blunt; don't soften things or add "great question!" filler.
## Projects
- Book ("Tidewater"): literary novel, ~60k words drafted. Stuck on the middle act.
- Newsletter: weekly, ships Fridays.
## Recent
- Jun 24: decided to cut the second POV character from the book.
Durable is the stuff that stays true for months — who you are, how you like answers, your standing preferences. Projects is what you're actively working on. Recent is a short log of what just happened that might matter soon.
The key is keeping it short. A tight, accurate one-pager beats a long stale one every time. Which is exactly why you let the AI maintain it.
The part that makes it self-cleaning
When I say "Update memory," I don't want the AI to just dump everything we said into the file. I want it to curate. So I give it a few rules up front (paste these into the same instructions, once):
- Promote: if something from our chat will still matter in a month, move it into Durable.
- Reconcile, don't pile up: if a new fact contradicts an old one, replace the old line. If I switch from coffee to tea, change the line — don't keep both.
- Prune: drop Recent items that stopped mattering. Keep it under a page.
With those rules, the doc stays alive instead of bloating into a junk drawer. A week in, it quietly fills up with the things that actually matter, and it stays readable because the AI is constantly throwing out the trivia.
Where you paste all this
You only set the rules up once, in the AI's settings:
- ChatGPT: Settings → Personalization → Custom Instructions. Or make a Custom GPT and put it in the Instructions field.
- Claude: Make a Project and paste it into the project's custom instructions.
- Gemini: Make a Gem and paste it into the Instructions.
After that, you never touch the settings again. The only ongoing habit is paste-at-start, "update memory"-at-end. Thirty seconds.
The quiet superpower: it's portable
Here's my favorite part. Because the Memory Doc is just text you own, you can paste it into any of them. I draft in ChatGPT, but when I want a second opinion I drop the same doc into Claude and it instantly knows my whole situation. One memory, three AIs, no lock-in. Try doing that with the built-in memory features.
Does it actually work?
Honestly, the first day it feels like extra steps. The payoff shows up around day three or four, when you open a chat, paste, and the AI just... gets it. No re-introduction. It remembers you cut the second POV character. It keeps its answers short because that's in the doc.
It's not magic and it's not perfect — you have to actually do the save-and-reload habit, and if you get lazy, the memory goes stale. But it's the cheapest, most portable way I've found to make a chat AI feel like it knows me. And it costs nothing.
Steal it
That's the whole thing — one doc, two habits, three rules. You can copy everything above and build it yourself in an afternoon.
If you'd rather skip the setup, I packaged the fill-in-the-blank persona block, the full memory protocol, and a worked example into a no-code kit: AI Soul Kit (No-Code) — Core ¥980, Plus ¥3,800. But the idea above is free. Steal it.
Top comments (0)