How to keep ChatGPT from losing track of a project.

A short question fits in one conversation. A real project doesn’t — it runs for weeks, across dozens of sessions, and what’s true about it keeps changing. The budget moves. A decision gets made, then revisited. A question you couldn’t answer in week one gets answered in week four, in a completely different chat.

ChatGPT is genuinely good within a single conversation. The trouble starts at the edges of it — the next chat, the next day, the version of the project that exists in your head but not in anything ChatGPT can read. You either re-explain the whole thing from the top, or you plough on and hope you remembered the last update correctly.

ChatGPT Projects. A project groups the chats and files about one thing, and lets you set instructions that apply to all of them. That’s real, and worth using on its own terms. What it doesn’t do is hold a single answer to “what’s true right now” — that answer is still spread across however many chats sit inside the project, each one a snapshot of the moment it was written.

ChatGPT’s memory. It saves things it notices about you across every conversation — your role, how you like answers phrased, a preference you mentioned once. It’s a profile, and it’s actually good at that job. A project isn’t a fact about you, though, so memory has nowhere natural to put “the tile is back-ordered until March, and the budget’s already down to $30k” — it isn’t organized around a single undertaking at all.

A manual context document, pasted in. The most reliable of the three, and the most work: you keep a doc — decisions, open threads, next actions — and paste or attach it at the start of each new chat. It works exactly as well as you keep it current, and keeping it current is a separate chore from the conversation that just changed it.

Different mechanisms, same gap. None of the three holds a project’s current state as a thing that updates itself the moment something changes. Projects hold chats, not conclusions. Memory holds you, not the undertaking. The document holds whatever you last remembered to type into it. In every case, the actual state of the project still lives in your head, and everything else is a more or less lossy copy of it.

That’s also the honest limit of asking ChatGPT to “summarize where we left off.” It will try, by combing back through the conversation and guessing which parts still matter. That’s a reconstruction, done fresh each time — not a record that was kept.

The project’s current state needs a home of its own — outside the chat, updated as an action rather than rebuilt by scrolling. Not a transcript, and not a profile of you: four kinds of thing, kept current. What’s true right now. The calls you’ve made and why. The questions still open. The odd note worth keeping. When one of them changes, the old value doesn’t vanish — it’s replaced, with the previous one kept and dated, so you can see what moved and when.

That’s the whole idea behind how Giraffe works: you keep talking to ChatGPT the way you already do, and when something is worth keeping, you say so — “save that to the kitchen project” — instead of maintaining a second copy of the project by hand.

Say you’re redoing a kitchen. Month one, ChatGPT helps you think through layout — island or no island, where the plumbing can actually move to without opening a wall. Month two, a different set of chats compares two contractors’ quotes. Month three, the tile you wanted is out of stock and you’re picking a replacement at 9pm from a supplier’s website. By month four you have perhaps thirty conversations, and the honest answer to “what have we actually decided” takes real effort to reconstruct.

With a project kept alongside all of it: the budget sits at one current number, with the two it replaced still visible underneath, dated. “No island — it would block the door swing” is written down as a decision, not buried in a chat from February. “Whether the contractor can start before the in-laws visit in June” sits as an open question until it’s actually answered, instead of getting silently dropped. Ask ChatGPT to check the project before the next call with the contractor, and it reads the current numbers instead of the ones from two revisions ago.

None of that required switching tools mid-project or exporting anything. It required one sentence, said out loud, whenever something worth keeping happened.

It’s worth being precise about what this is, because it looks adjacent to ChatGPT’s own memory and isn’t really competing with it. Memory keeps getting better at knowing you, and that’s worth having on regardless. A project is organized around the kitchen renovation, or the client account, or whatever the undertaking is — not around you — which is a structural difference, not a size difference. Why AI memory isn’t the same as project memory goes into that distinction on its own.