Project memory tools
Which kind of AI project memory do you actually need?
“AI project memory” gets used for at least four different jobs: giving an agent enough context to do a task, remembering engineering decisions across coding sessions, tracking who owns what on a team, and keeping a plain record of what’s currently true about something you’re working on. They’re not interchangeable, and most of the tools that show up for this search only do one of them.
Four jobs, one phrase
Agent context. Give an AI agent enough background to perform a task well right now: files, prior attempts, relevant facts, assembled or retrieved on the fly. This is the job most developer-facing “AI memory” infrastructure is built for.
Coding-project memory. Remember what happened while building something: the approach that failed, the decision behind a design choice, the bug and its fix, so a coding agent doesn’t repeat the same dead end in a later session.
Team/project-management state. Tasks, owners, deadlines, status. Whether a human or an AI assistant is doing the updating, this is fundamentally execution tracking: the shape of a project board.
Human-readable undertaking state. What’s currently true, what was decided and why, what changed, what’s still unresolved: written so a person can read it cold, for one specific thing they’re doing. This doesn’t require the undertaking to involve software at all: an Italy trip, an Acme renewal, a car purchase, a career change all fit this shape exactly as well as a codebase does. This is where Giraffe sits.
Two of those four have a market this page is not the right guide to. If you need agent context, the answer is whatever your framework already uses, and it is a developer decision rather than a product one. If you need coding-project memory, the coding tools have built their own and it lives beside your repository.
The table below is the other two: team and project state, and the readable state of one undertaking. Those are the jobs where a person genuinely has a choice to make, and where the wrong choice costs months.
Five approaches, compared
Five ways people actually keep this, not five brands. The Giraffe column describes what Giraffe does; the other four describe how that kind of tool generally behaves, which is the level at which the choice is really made.
The gap is reading it cold: not what a tool remembers, but whether you can find out without asking it.
| Native project context | Knowledge workspace | Team project system | Meeting memory | Giraffe | |
|---|---|---|---|---|---|
| Built primarily for | Anyone already inside one assistant, working on one thing | People who keep notes and documents, and want an assistant to read them | Teams with work to assign and ship | People whose work happens in calls | One person, one undertaking, whichever assistant they are in |
| What it remembers | What was said in that project's chats, plus what the assistant chose to note | Whatever you wrote down, as documents and links between them | Tasks, owners, status and dates | What was said in the meeting, and what follows from it | Current facts, decisions, superseded values, open questions and notes |
| Where you actually read it | By asking, or as a list of what it noticed about you | In the documents, which you maintain | On a board, by ticket | In the transcript or its summary | On a project page, written to be read cold with no assistant open |
| Works across assistants | No. A project's memory belongs to the assistant it lives in | Yes, if each one can read the workspace | Not the point of it | Depends on the tool | Yes. The same project reads and updates from ChatGPT, Claude and Gemini |
| When a value changes | The new answer arrives; the old one stays in the history to be found | You edit the document, and the previous version is in file history if you look | The field changes and the old value moves to an activity log | The next meeting supersedes the last one, in a different transcript | The value updates in place and the one it replaced stays visible, dated and attributed |
| Center of gravity | The conversation | The document | The task | The meeting | The undertaking |
What the table can’t show cleanly
Native project memory is further along than it gets credit for. Each Claude project has its own memory space, kept separate from other projects and from ordinary chats, and it is on by default on Free, Pro and Max. ChatGPT Projects can be set to project-only memory, so the project’s chats stop drawing on your saved memories from outside it and nothing carries back out. If your work lives in one assistant and you are content to ask it questions, that may be all you need.
What none of them do is leave the assistant. A project’s memory belongs to the product it lives in. Claude offers a manual import: you paste in what another provider had learned and it is added to memory. Anthropic calls the feature experimental, and says Claude may not always incorporate what you paste. Either way it is a copy, not a link, and it is not the same as one record two assistants keep current between them.
The gap is reading it cold. Most of the approaches above are read by asking: you open the assistant, or the board, or the transcript, and put a question to it. A workspace is the exception, and the price is that you maintain it by hand. What Giraffe is doing differently is narrow: the state is written down as a page, so it can be read with nothing open, and the value a number replaced stays under it, dated and attributed, rather than living in a history you would have to go looking for.
What that looks like for something that isn’t software
Say the undertaking is replacing a car, not a codebase or a client engagement. There’s no repository, no pull request, no ticket, no team to loop in: the professional workspaces in the table above are built for work with more than one person attached to it, and setting one up for a car purchase would be more structure than the job needs. There’s still a current budget, a decision to buy hybrid rather than electric and why, and one open question about financing, and that state is exactly as real, and exactly as easy to lose across a month of scattered ChatGPT sessions, as engineering context is across a month of coding sessions. Giraffe treats a car purchase the same way it treats an Acme renewal or a career change, because the underlying shape of the problem is identical: what’s true now, what was decided, what changed, what’s still open. See the actual project this example is drawn from, or read how Giraffe works more generally.
A few more questions
What should I use to keep project context across ChatGPT and Claude?
It depends which job you mean. To give a coding agent enough background to finish a task, you want a coding-project memory tool. To keep tasks and owners moving, you want a project system. To keep the current state of one undertaking readable in both assistants, you want a record that lives outside either of them, which is the job Giraffe does.
Do ChatGPT Projects and Claude Projects already solve this?
For one assistant, they solve a lot of it. Each Claude project has its own memory space, separate from other projects and from ordinary chats. ChatGPT Projects can be set to project-only memory, so a project's chats do not draw on your saved memories from outside it and nothing carries back out. Both keep that memory inside the assistant it belongs to, so the state does not follow you to the other one, and neither is written as the state of one undertaking, with the value a number replaced still under it.
Do these approaches take the same amount of setup?
No, and the split is not technical against non-technical. A knowledge workspace is the most work, because you maintain it. Native project memory is the least, because it is already on. Giraffe sits between them: connecting an assistant is a settings step, and after that the record is written by asking your assistant to write it.
Why does this page compare approaches instead of products?
Because the question underneath the search is which kind of thing you need, and a row of product names does not answer it. Once you know whether your problem is agent context, coding memory, team execution or the readable state of one undertaking, the shortlist is short and you can judge products yourself.
The right tool depends on which of the four jobs is actually yours. Most people asking this question have never had it named as one of four.
Read next
You already pay for one of these.
Six columns here, and the one most readers already have is not among them. What ChatGPT’s memory and Claude’s genuinely cover is worth settling before you add anything, and the part they don’t is specific enough to name.