AI memory
Why AI memory isn’t the same as project memory.
ChatGPT and Claude already remember things about you. That’s not the same as remembering where a project stands, and it isn’t supposed to be.
ChatGPT’s and Claude’s own memory is built to remember things about you, your preferences, your job, the way you like answers phrased, pieced together from everything you’ve ever said to it. That’s personalization, and both are already good at it. It’s a different job from remembering where one specific undertaking actually stands: what’s decided, what changed, what’s still open. Native memory was never built to hold that shape, and piecing it together after the fact is exactly the hard, unreliable part. A project holds it because it’s named and written down as true the moment you land on it, not reconstructed later by guessing.
What personalization memory is for
Ask ChatGPT or Claude to remember something about you, and it will: a preference, a fact, the way you like to work. Over time it builds a running impression of you out of everything you’ve said, across every conversation, on every topic. That’s genuinely useful, and it’s exactly what that memory is for: making the assistant feel like it knows you. It isn’t organized around any one thing you’re doing. It’s organized around you.
The question it was never built to answer
The trouble starts when what you actually want isn’t “remember that I like directness” but “remember where the Q3 launch decision landed.” That’s not a fact about you, it’s the state of a specific undertaking, and nothing in a personalization memory was built to hold that shape. To even attempt it, the system has to comb back through everything you’ve said, guess which parts belong together, and guess which one is still true if two of them disagree. Ask either assistant to summarize where things stand on something you discussed weeks and several conversations ago, and you’ll get a best guess, assembled from fragments, not a record.
A record of an undertaking, not a profile of a person
A project is organized around the undertaking, not around you. When you tell your assistant what you’ve landed on, that becomes part of a named, specific record, this trip, this decision, this launch, not one more fact folded into a general impression of who you are. Nothing has to be reconstructed later, because it was already organized at the moment you said it.
Where Giraffe sits
Giraffe is a project outside of either assistant: the current facts, the decisions, what’s still open, for one specific thing you’re doing, not the conversation that produced them. You tell your assistant what you landed on, it saves that to the project, and the project stays organized around the undertaking instead of around you. ChatGPT and Claude keep being good at knowing you. Giraffe keeps track of what you’re doing.
What a record does that a memory cannot
Being organized around the project also means it can be precise about what kind of thing changed. A fact updates in place, with the old value kept as history, true right up until it wasn’t. A decision that gets reversed stays visible instead of vanishing, marked reversed, with the reason attached. And a question you’re still turning over stays open on purpose, right there in the project, until you actually settle it, at which point your assistant marks it resolved, with the answer, rather than quietly forgetting you’d ever been unsure.
One open question, before and after
Say you’re hiring for a small business. Early on you tell your assistant: “Open question for the hiring project. Full-time or contract to start?” You genuinely don’t know yet, so it stays logged as open rather than decided either way. Three weeks and several candidates later: “Resolve that. We’re going contract-to-hire, three-month trial. Two of the three people we liked only wanted to test the fit first anyway.” The question moves from open to answered with your reasoning attached. When you come back to draft the job listing, you ask it to check the project first, and it reads contract-to-hire as settled. Not because it remembers a conversation from three weeks ago, but because there was an answer waiting instead of an old open question it would have had to guess about.
Read it without opening an assistant at all
Open the project on its own and the open questions aren’t buried in search results next to everything else you’ve ever said. They’re just there, next to the facts and the decisions, in a project that reads like a page: what’s settled, what’s still open, in plain sentences. That’s the Simple view. Structured view shows the same project broken into its labeled parts. Either way, it’s a record of the undertaking, not a profile of you.
Memory is what an assistant knows about you. A project is what is true about the thing you are doing.
Starting one
- Connect Giraffe in ChatGPT’s or Claude’s own settings, or Gemini through Gemini Spark. Once each, and what it takes depends on your plan: Claude works on every plan, ChatGPT needs a paid one on the web with Developer mode turned on, and Gemini goes through Spark rather than an ordinary chat.
- Next time something is genuinely undecided, say so: “Log an open question on the [project]: we haven’t settled this yet.”
- When you actually settle it, tell whichever assistant you’re using to resolve it, and why.
- Ask the other assistant to check the project before it assumes anything. It reads the settled answer instead of guessing at one.
Before any of that, it helps to see the thing itself. What a project holds, and how it is read.
Connect ChatGPT or Claude to Giraffe: setup steps for each assistant, kept current as their own settings change. Connect Gemini through Gemini Spark is a separate setup path: Gemini’s custom apps only work inside Spark, not an ordinary chat.
A few more questions
Isn't this just what ChatGPT's or Claude's memory already does?
Related, but built for a different job. ChatGPT's memory is a running list of things it's noticed about you, built up across every conversation. Claude keeps memory too: an account-wide one, plus a separate memory space inside each project that does not draw on it. A Giraffe project isn't about you at all: it's a record of one specific undertaking, that both can read and add to.
Does Giraffe replace ChatGPT's or Claude's own memory?
No, and it isn't trying to. Keep whatever personalization you've already got: Giraffe isn't where "I prefer short answers" belongs. It's where "here's what we've actually decided about the Q3 launch" belongs.
Why track an open question at all, instead of just answering it later?
Because "we haven't decided yet" is information too. Without it recorded, the next assistant you ask either guesses at an answer or makes you explain the whole thing again from scratch. An open question sitting in the project means it does neither.
Can a resolved question turn out to be wrong?
Yes: resolving isn't final. If it turns out you were wrong to settle it, tell your assistant to reopen it, and it goes back to open, same as before.
Does Giraffe pull things in automatically, or do I have to say so?
You say so, or your assistant does, when you've told it to. Nothing is pulled in just because you talked about it.
Read next
So what does a record of one undertaking look like?
The argument here is that a project has a shape a profile of you cannot hold. The clearest case is a number that keeps moving and drags its own decisions with it, which is what a major purchase is for ten months.