Your AI's Memory Isn't Yours
It forgets what matters, keeps what doesn't, and never asks. The fix is a file you own.
Your AI remembers you now. Open a new chat and it already knows a few things about you — your name, what you’ve been working on, how you like your answers. A year ago every session started cold; today ChatGPT and Gemini keep notes by default, and Claude will if you let it. Sounds like the fix finally arrived.
It didn’t. What the AI kept is a thin, secondhand sketch of you, and it’s usually missing the one thing you needed it to hold. So you’re still re-explaining yourself before you can get to the question. Now you’re also correcting what it decided you were.
Three things wearing the same word
“Memory” gets used for three different things, and only one of them is what you actually want.
The first is the conversation you’re already inside. The model sees the whole thread, so within a single chat it tracks you perfectly. That’s just the current window, though. Close it and it’s gone.
The second is the cross-chat feature, the toggle people mean when they say the AI “remembers” them. This is the one that burns you. Technically, it works… which is what you say about software you don’t trust.
The third gets lumped in but isn’t memory at all — tool access. Connectors, integrations, the plumbing that lets an AI reach into your inbox or your files and pull something live. It’ll grab a fact for you, but holding onto one is something else entirely. That’s reach, not recall.
Pull those apart and you’re left with something bleak. The only durable memory most people have is the lossy one they can’t steer.
The one that burns you
The trouble is structural. It isn’t filing away your facts; it’s writing itself a short summary of you, a few lines in its words, that it pulls into later chats. What comes back is the model’s gist of you, written by the model, for the model.
That one move — summarize, then trust the summary — is where every problem comes from. The model decides what’s worth summarizing, and it leans toward what’s easy to pull out over what actually carries weight for you, so it’ll recite your dog’s name (and does, every chance it gets), but the project you’ve walked it through a dozen times? Gone. The summary is lossy on purpose; a few lines can’t hold what matters, and the specifics flatten into something generic.
Some tools let you look and prune: open the list, cut what’s wrong. Others barely let you see it at all. Either way, you’re editing the model’s draft of you. It decides what goes in, in its words, on its schedule, and your part is cleanup. When it drops something, nothing warns you; you find out mid-task, when the thing you assumed was safe simply isn’t there. You’re not wrong that it forgets. You’re wrong to expect it won’t.
The fix is a file you own
So stop relying on it. Hand the AI your context yourself, every time, out of a file you keep.
That’s the whole move. It sounds too dumb to work, and it’s the most reliable way I’ve found to keep an AI useful from one session to the next. The file gets read at the start and rewritten at the end. Everything the AI knows about you sits in it, in plain words you can open and read — because you put it there, or watched it go in.
What goes in the file is the part worth getting right.
The everyday version
Most of what you’d want your AI to remember splits into two speeds.
One part changes constantly — what you’re actually in the middle of:
## Working on
- Inner Dialogue — v2.7 shipped, repo cleanup before I point more people at it
- Inner Dialogue — protocol layer refinements (people, places, concepts, events)
- Vendor eval — 2 of 4 demos done, scorecard half-built, security review still out
- Drip irrigation for the backyard beds — parts in, still need to map the zones
The other part barely changes — who you are and how you like things done:
## How I work
- Lead with the recommendation, then the why
- Tables for comparisons, not paragraphs of prose
- Challenge my timelines, don't just agree with me
- Show me the plan before any big change
## What stays true
- West Coast / Pacific time — assume it for anything scheduled
- Former audio engineer and musician; technical analogies land, skip the 101s
- I still gig on weekends, so they're not as free as they look
Keep those two apart. Mash them together and the daily churn buries the things that stay true, or the durable stuff goes stale because updating it means scrolling past today’s noise every time. One file you rewrite constantly, one you barely touch.
That’s the structure — a handful of text files plus the rules that keep them current. Underneath the tooling it’s almost embarrassingly plain: you read the notes at the start, you update them at the end. Mine’s on GitHub if you want a head start.
The same mechanism runs everything I do with AI — even a therapy tool I built, Inner Dialogue, where the file holds a profile of a person instead of a project list. The content changes completely; the file doesn’t.
The catch
There’s a cost. A memory you own is yours to keep up — by hand, or with a setup that reads and writes the file for you (mine does, on rules I wrote). Automation isn’t the catch; ownership is. The loop is yours, which also means a neglected file goes stale fast, and that’s worse than no file, because you’ll trust it anyway. But that responsibility is the same thing that makes it work: it runs on rules you can see and a file you can change, so nothing gets decided behind a setting you never open. That’s why it holds: you’re the one keeping it true.
You can start today
The crudest version works in any tool right now. Keep one note about yourself and what you’re up to, and paste it at the top of a new chat. Clumsy, and still miles ahead of trusting the toggle.
A step up is to put that note where the tool grabs it for you. A ChatGPT Project holds instructions and a few files that ride along in every chat inside it, so you set it up once and stop repeating yourself.
The most hands-off version has the tool read and write the file for you, so you never paste anything: notes in a plain text app, an AI coding tool that loads them when you start and updates them when you close out. It’s more setup than most people need — if you want it, I’ve written up the architecture separately.
The rung you pick doesn’t matter. Owning the file does.
Before you build anything
Try one thing first. Open your AI’s memory settings and read what it’s actually saved about you. Most people never have.
You’ll find one of a few things. Either it’s thin and generic, which means you’ve been trusting a memory that was barely there. Or it’s full of oddly specific things you don’t remember handing over — the dog again, your home address, a guess about your income you never confirmed — which is a fun thing to learn about a piece of software. Or it won’t show you anything at all. Some won’t, and that’s the loudest answer of the three. They all land in the same place: the context that makes the AI useful to you is too important to leave to a feature you can’t see.
So don’t. Keep it where you can read it, in words you chose, updated when you say so. Mine isn’t a feature. It’s a file — and the reason that matters isn’t that it’s clever. It’s that I’m the one who decides what it keeps.