What your AI already remembers, and what it doesn’t.

Open Claude’s memory panel and read what’s in it. Your stack. How you like to be answered. The constraint you mentioned twice and shouldn’t have to mention again. ChatGPT’s saved memories are the same kind of thing. This is a profile, and it’s built to do one job well: make the next answer fit you without you having to set it up each time.

Now try to find this in there. Contract, $48k a year. Renews Mar 31. Health, at risk. Decision — offer a two-year renewal at 6% uplift. Open question — does Raj own the budget now, or has it moved to finance? None of that is a fact about you. It’s the state of a thing you’re working on, and most of it will be wrong in a month, which is exactly why it needs somewhere that expects it to change.

There’s a second reason it won’t be in there. Open any ChatGPT project, click its name and look at the memory setting: project-only means it uses what’s inside that project and deliberately not your account-wide memories. Claude does the equivalent — chats inside a project remember separately from your ordinary ones. Both are sensible; you’d ask for work context to stay out of your general profile if it weren’t already there. But the moment you ask the quick question outside the project, which everyone does, the answer lands somewhere the project can’t see.

Giraffe has one place per thing you’re working on. It doesn’t matter which assistant you had open, whether the chat was long or two lines, or whether you typed it in yourself at eleven at night. It lands in the same project, and it’s there in the morning.

The practical test is the ten minutes before a call. You cannot open a memory panel and be ready. You can open a project and be ready, because a project is written to be read cold by the person who wrote it.

Claude’s memory entries are updated in place. ChatGPT’s memories are replaced. That’s the right behaviour for a profile — you don’t want your assistant working from a job you left in 2023 — but it means the earlier answer is no longer in the profile. It is still in the chat where you said it, and in a ChatGPT project that chat is still sitting in the folder. So the old number isn’t destroyed so much as scattered: to get it back you have to work out which of forty conversations said $56k, open it, and read it — and nothing in there will tell you that number was superseded three weeks later. Notion is closer: page version history is real, it’s block-level, and it’s retained by plan tier — seven days on Free, thirty on Plus, ninety on Business, longer again on Enterprise. What it tells you is that a page changed.

What you usually need to know is narrower and older than that. Not that the page changed. That the Contract went from $56k to $48k on 12 March, saved by Claude. In Giraffe that record is per item — what it said before, next to what it says now, dated, attributed, never overwritten — and it ships in your export as history.jsonl.

The thing you lose isn’t the answer. It’s why it was the answer.

A number you replaced doesn’t sit on any current row — it survives inside the change log, and the change log is searchable too.

The contract was $56k. Now it’s $48k. Turn History on and search 56k: the old number comes back struck through, labelled Past value, with not current · replaced by a newer value beside it, and opening it lands you in the Acme renewal project’s history at the change that replaced it. A decision you reversed works the same way. So does a question you closed. None of it is deleted, because deleting it would make the record lie about what happened.

Two rules keep the old number out of your way. History is off unless you turn it on, so an ordinary search only ever returns what’s true now. And a past value never arrives unmarked — it comes back struck through, labelled, and told you what happened to it. Neither of those is a dial we set to a sensible default — there is no dial. Nothing here is edited over the top of anything: a new value goes in, the old one stays behind it with the date and which assistant saved the change.

Notion’s strength and its cost are the same thing: you decide everything. Which database, which properties, what counts as a decision, and where a stray finding goes when it doesn’t fit any of the columns you made. If you’ve built that and you maintain it, Notion will hold your work beautifully — point the connector at a database you designed and the writes are genuinely structured, filling in the properties you set up, and the number lands in a column you can filter on later.

The cost is upstream of that. It only ever writes into columns you designed first and keep maintaining, and the finding that fits none of them lands as a paragraph on a page — which is prose you have to read again to find the number in. If you haven’t built the database yet, that’s every finding.

Giraffe ships the structure and doesn’t let you change it. Four kinds of item — what’s true now, the decisions, the open questions and the notes — one live value per fact, one level of Spaces for grouping projects, and nothing to configure. That’s a real loss of flexibility and it’s the whole trade: “save what we concluded” lands sorted, because there was only ever one place for each part of it to go.

The second difference is the one nobody brings up. Notion’s connector holds read and write across the workspace you sign it into. A Giraffe connection is allowed to do much less, on purpose: it can read your projects and update them, and it cannot delete a project, create or rename a Space, manage your connections or pin anything. Not as a policy we promise to follow — the database doesn’t grant it those. Disconnect it and it stops immediately, and the block is tied to the connection, not to what it calls itself.