Fixed facts about your business
What you sell, to whom, in which region, at what rates, on what terms. Things a new employee needs to know on day one and that stay the same for six months.
We only save your display preference in a preference cookie, for up to 180 days on this device. No CRM or advertising data. Your tracking consent stays unchanged. If cookies are blocked, your choice may not be saved.
You share one memory between ChatGPT, Claude and Codex by putting it outside the AI apps: in a text file you bring to every conversation yourself, or in a memory server that every AI reads and updates over MCP. Importing and browser extensions move your memory from one app to another, but they do not share it: after the copy, the two grow apart again.
Below you can read why every session starts empty, which four options there are and what they ask of you, what belongs in a shared memory and what must never go in, and how we do it ourselves at TheSEO with Jarvis. All sources were read on 11 November 2026.
~/.claude/skills/ (or .claude/skills/ in your project)./ai-geheugen-opzetten-en.AGENTS.md in your repo.AGENTS.md.
Every AI session starts empty. What you explain to ChatGPT in the morning about how your quotes are structured, the tone you take with clients and the margin you never give away, Claude does not know in the afternoon. Open Codex after that for a piece of code and that conversation starts from zero as well. None of the three is at fault. They are built that way: each conversation is a room of its own, and what is said in one room does not carry through to the next.
At TheSEO we ran into this every day. We work with Claude, with Claude Code in the terminal, with Codex and now and then with ChatGPT, each for the work it is good at. The same facts about our business, the same agreements and the same style rules went out of the door four times, and each version drifted a little further from the others. How that feels in practice, and why we built Jarvis for it, is on the page on how Jarvis works.
The memory that ChatGPT and Claude now have themselves does not solve this. That memory lives inside the app. ChatGPT remembers what you say in ChatGPT; Claude remembers what you say in Claude. The same goes for the tools developers use: Claude Code does not automatically load what Codex remembered, and Codex does not read what Claude Code recorded (One context: MCP shared memory, 2026). Each tool keeps its own half of the story.
So the question is not how to give each AI a better memory. The question is where you put the memory so that all of them can reach it. Once you look at it that way, only a few answers remain, and we walk through them below.
Partly. A good opening paragraph per conversation helps, and many people keep such a paragraph in a note somewhere and paste it in each time. That is in effect already a shared memory, only without structure and without a date. The trouble starts once that note lives in three places and changes in three ways. This article is about turning that loose note into one place that is correct.
Since March 2026 Claude has been able to import memory from ChatGPT: you retrieve what ChatGPT remembered about you and place it in Claude (PC World, 3 March 2026). That saves an afternoon of typing. But it is a one-off copy, not a continuous sync. What you change in ChatGPT afterwards does not reach Claude, and the reverse is true as well (memx.app, on Claude and ChatGPT memory).
Browser extensions such as Migrato, MemoryPlugin and MemoryBase do essentially the same thing, only between more apps: they read what one AI remembers and put it into another (MemoryPlugin, MemoryBase). Useful if you are switching, or if you want to start in two apps from the same base. But this too is moving, not sharing. Afterwards you have two memories that are identical at that moment and start growing apart again the next day.
The difference sounds small and is large in practice. Sharing means: there is one place, and whoever reads that place always sees the latest state. Moving means: there are two places, and you are the one who has to keep them in step. If you take a decision on your prices in Claude on Tuesday, ChatGPT knows it on Wednesday only if you copy it across again. Forget that once and your AIs are working from two different truths, and you only notice when a quote comes out wrong.
On top of that, an app's memory consists mainly of what the app itself thought worth keeping. You do not control which sentences are stored, or in what form. A business wants it the other way round: you decide what is settled, and the AI reads it.
No. If you have worked with ChatGPT for months and are now adding Claude, the import is a good starting point: you do not have to begin again. Think of it as a removal box, not a shared cupboard. Afterwards, take out what you really want to keep and put it in the one place you choose below.
There are four ways to work with more than one AI without explaining everything again each time. Two of them move your memory, two genuinely share it. The table shows what each option asks of you.
| Option | Sharing or moving | Works in | What you have to do yourself |
|---|---|---|---|
| The memory inside the app itself (ChatGPT, Claude) | Neither: each app remembers only its own conversations | Only that one app | Nothing, but you explain it again in every app |
| Importing or copying (Claude import from ChatGPT, extensions such as Migrato, MemoryPlugin, MemoryBase) | Moving: a one-off copy | The app you copy into | Copy again after every change, in both directions |
| A text file you carry with you (such as our business brain template) | Sharing, by hand | Any AI you paste it into or attach it to | Keep it up to date yourself and bring it to every conversation |
| A memory server over MCP (Mem0, Zep, Supermemory, Memlo, Membase, Jarvis) | Sharing, continuously: every AI reads and writes the same place | Any AI that can connect to an MCP server | Connect once, then maintain what is in it |
Source: PC World and memx.app on the Claude import, MemoryPlugin and MemoryBase on extensions, HydraDB and Membase as overviews of memory layers over MCP. Read on 11 November 2026. The business brain template and Jarvis are made by TheSEO.
MCP stands for Model Context Protocol. It is the open standard that lets an AI assistant use external tools: a server offers tools, the AI calls them, and the result comes back into the conversation. Claude, Claude Code, Codex, Cursor and ChatGPT on the business plans can connect to such a server. What MCP is exactly and how you connect to it, we explain in plain language on our MCP page.
For a shared memory, that is the key. If the memory is an MCP server, it does not matter which app you are in: every AI connected to that server reads the same facts and can add the same decisions to them. That is the difference from importing. There is no copy, there is one place.
A lot of work is going into memory layers of this kind, mostly for developers. In October 2025 Mem0 raised 24 million dollars to build, in the words of the TechCrunch headline, the memory layer for AI apps (TechCrunch, 28 October 2025). Zep and Membase work in the same corner. Supermemory shed its consumer product and became infrastructure for other apps (Supermemory, on alternatives to Mem0). Memlo is a smaller player. On 11 November 2026 the directory site mcpservers.org listed 405 MCP servers in its memory category (mcpservers.org, memory category). Overviews that set the approaches side by side can be found at HydraDB and Membase.
What these servers have in common: they are built to be read and updated by an AI, and they sit outside the app. In that respect they all differ from the memory in ChatGPT or Claude. What they also have in common: most are aimed at people who are not put off by a configuration file. For a business owner without a technical background, connecting is often the highest hurdle, not using it afterwards.
For anyone working with Claude Code and Codex, an MCP memory server is at present the only way to have the two know the same things, because they do not read each other's memory (One context, 2026). That is where we started too: not with a plan for a product, but with two tools working past each other.
Not to use one, though often to connect one. In Claude and ChatGPT you add a server through the settings, with an address and a single sign-in. In Claude Code, Codex and Cursor you put a few lines in a configuration file. That is one-off work. After that you simply talk to your AI and it calls the tools itself.
A shared memory is only valuable if you are strict about what goes into it. This is the structure we keep to ourselves, learnt by getting it wrong first: we put everything in and then read nothing back out.
What you sell, to whom, in which region, at what rates, on what terms. Things a new employee needs to know on day one and that stay the same for six months.
How a quote is structured, what is asked in an intake, who approves what. The rules your AI has to keep to, even when nobody is watching.
Not "from now on we do X", but "decided on 11 November 2026: X, because Y". Three months on, a decision without a date cannot be told apart from an old idea.
Formal or informal address, which words you avoid, how you speak to a client, whether you give figures with a source. Everything you would otherwise correct again in every text.
Never. A memory is read by every AI connected to it and ends up in prompts and logs. A key belongs in a password manager or a secrets vault, nowhere else.
No names, email addresses or files of clients without a reason and without an agreement; that is also a GDPR question. And no "today I am working on the quote for the roofer": tomorrow that is noise.
Source: our own way of working at TheSEO, laid down in the rules for our own AIs. What we do and do not record is also on the Jarvis security page (in Dutch).
The hardest part of a shared memory is not setting it up but keeping it up. A memory nobody looks after becomes a pile of half-truths within a few months, and an AI reading from that is worse off than an AI that knows nothing. Four rules made the difference for us.
One owner. One person is responsible for what is in it. Not a team, not 'everyone', and not the AI either. At TheSEO that is Gianluca. An AI may propose an addition, but the owner approves it. That stops three AIs overwriting one another with what they thought they had understood.
A date on every decision. Every decision, every agreement and every fact that can change gets the date on which it was recorded. That costs a few characters and saves every argument about which version is the newest. An AI that sees two conflicting rules then picks the more recent one, and you can see at a glance what is due for replacement.
Old facts out, not added to. If your rate changes, you replace the line. You do not put a new line underneath it saying 'from now on'. Otherwise a memory only ever grows, and every outdated line is a chance that an AI will use it once more. Small and correct beats complete and half true.
Ask it back. You only know an AI reads your memory once you test it. Start a new conversation and ask: what do you know about our margin, about our style, about last week's decision? If the right answer comes back, with the right date, it works. If you get a generic answer or something made up, the AI is not reading the memory, or it is reading an old version. Do this not once during set-up, but every time you have added something important.
There is one more rule that we learnt the hard way: a stored text is not automatically a verified fact. An AI that puts something into the memory may have misunderstood it. That is why we have a person approve additions before they count, and why every fact states where it comes from.
Smaller than you think. Everything an AI reads at the start of each conversation costs attention and space in that conversation. A memory so long that you no longer read it through in one go yourself is one the AI will not read well either. What you rarely need goes in a separate document the AI can ask for; the memory itself keeps only what applies to every conversation.
At TheSEO we solved it with Jarvis, an MCP server at mcp.theseo.nl/mcp that you sign in to through the browser. In plain language it offers four things: Memory (remembering and recalling, plus fixed profile fields about your business), Inbox (messages between your AIs), Tasks and Log. Claude, Claude Code, Codex, Cursor and ChatGPT on a business plan all connect to it in the same way.
What that does in practice: at the start of a session Claude reads what is in the memory, so what Codex recorded yesterday is there too. On 11 November 2026 Claude Code and Codex sent each other messages through that inbox for the first time, to coordinate a rollout of our website: who had which version and who was allowed to go first. Two tools that a week earlier were still working past each other and overwriting each other's work now settled it between themselves. That was the moment it went from an idea to a way of working for us.
A few things we say plainly. Jarvis is young: Start and Solo can be ordered, Team and Business are still in development. With Jarvis you pay for your brain, not per AI query; your own ChatGPT or Claude subscription stays separate from that. Messages and memory stay within the space of a single account; export and deletion are available on the account page. What we do and do not record, and how we handle access, is on the security page (in Dutch). Promises that are not made there, we do not make here either.
And most important of all: you do not need Jarvis to get started. The first step is a text file. Our business brain template (the page is in Dutch) is exactly that: a file with the structure from the cards above, which you fill in yourself and bring to every conversation, in Claude, ChatGPT, Codex or Cursor. It remembers nothing by itself and it does not grow with you, but you need nothing from us to use it. Only when you notice that you forget to bring the file, or that your AIs need to tell each other something, do you need a server.
Technically yes: an AI can connect to more than one MCP server. In practice we advise against it. Two memories means two truths again, exactly the problem you set out to solve. Choose one place and keep it up to date.
This is the order that worked for us. You can do the first three steps today with a text file. The last two are what keep a shared memory alive.
You can do steps 1 to 3 with the skill at the top of this page or with the business brain template. For step 4 with a server: Jarvis or one of the servers from the table.
What it gives you is easy to name: you explain things once. The ten minutes at the start of each conversation in which you say who you are and how you work disappear. More important still: the answers from your different AIs start to resemble one another, because they work from the same facts. A quote from ChatGPT and an email from Claude then use the same rates and the same tone.
What it does not give you, we state just as clearly. A shared memory does not make an AI smarter or more reliable. It can still make things up, including about what is in the memory. It does not replace a briefing per task: what you want for this one quote, you still have to say. And the maintenance is real work. Anyone who thinks a memory is filled once and then finished will, six months on, have an AI calculating with outdated rates.
There is a privacy side as well. If you put client data into a memory, you are processing personal data, and the ordinary rules apply: a reason, a lawful basis, and a data processing agreement with whoever runs the server for you. With a text file on your own computer that is easy to oversee. With a third party's server, you first read what that party records, where it is stored and how you remove it again. For Jarvis that is on the security page (in Dutch); for other servers, ask before you start.
If you would rather work out for yourself first how to set up this kind of thing, everything we have written on the subject is in the knowledge base. Start small, with one file and one owner. The rest follows once the first conversation no longer begins with explaining.
Then it is not the AI that wins, but the owner. With us an AI may propose an addition, but it only counts after approval. If two proposals contradict each other, Gianluca looks at the source of both and chooses. Without that step you get exactly the problem you started with: two truths, only now in one place.
No. ChatGPT's memory lives in ChatGPT and Claude's in Claude. Since March 2026 Claude has been able to import memory from ChatGPT, but that is a one-off copy. If you want both to see the same state, you put the memory outside the apps: in a file you carry with you or in a memory server over MCP.
Importing makes a copy at one moment; after that the two memories drift apart again and you have to keep them in step. Sharing means there is one place that every AI reads and updates, so everyone always sees the latest state. Browser extensions such as Migrato, MemoryPlugin and MemoryBase import; an MCP memory server shares.
An open standard that lets an AI assistant use tools from outside. A server offers tools, the AI calls them and the result arrives in the conversation. Claude, Claude Code, Codex, Cursor and ChatGPT on business plans can connect to it. A memory server is such a server, with tools like remember and recall.
Passwords, API keys and tokens: they end up in prompts and logs. Client data without a reason and without an agreement: that is a GDPR question. And temporary context, such as what you are working on today: tomorrow that is noise. What does belong in it are fixed facts, agreements, decisions with a date and style rules.
Ask it back. Start a new conversation and ask what the AI knows about your margin, your style or last week's decision. If the right answer comes with the right date, it works. If you get a generic answer or something made up, the AI is not reading it, or is reading an old version. Repeat this after every important addition.
No. Start with a text file, for example our free business brain template (the page is in Dutch), which you bring to every conversation yourself. You only need a server such as Jarvis once you forget to bring the file or your AIs need to tell each other something. The Start and Solo plans of Jarvis can be ordered online.
Yes. Both can connect to an MCP server through a configuration file, and both read a text file you place in your project. Claude Code and Codex do not read each other's own memory; an MCP memory server is the only way to have the two know the same things.
One person, never the AI. An AI may propose an addition; the owner approves it. Put a date on every decision, replace old facts rather than stacking them up, and never treat a stored text as a verified fact.
The first shared memory is a text file with what you keep telling every AI, without passwords and without client data, with a date on every decision and one owner. You can make that this afternoon, with the skill at the top of this page or with the free business brain template.
If you want your AIs to read and update the same memory without you carrying it around, Jarvis is our version of that; Start and Solo can be ordered online. In a call we show you how we use it ourselves.
Updated on 12 November 2026