The Memory in ChatGPT and Claude, and Where It Stops

Sooner or later everyone who has used ChatGPT for a few months says it: it knows me by now.

That is half true. It is worth knowing which half, because that decides what you can rely on and what you had better not.

Three Different Things Are Called Memory#

The first misunderstanding starts with the word. Three things are at work inside an AI tool, and all three look like remembering.

The first is the conversation you are in. As long as the window is open, the AI has everything you said in it. That feels like an excellent memory, but it is really a desk: everything is within reach, and when you close it, the desk is cleared.

The second is notes the provider keeps about you. Those outlive the conversation. When ChatGPT greets you tomorrow in the right tone, that is where it comes from.

The third is a memory of your own: a store that belongs to you, holding what you decided to keep, that several tools can reach. That is a different animal from the first two, and the difference is what the rest of this is about.

Most of the confusion comes from the fact that number one feels far more impressive than number two, so people credit number two with what number one did.

How ChatGPT Remembers#

There are two routes by which something sticks.

The first is the list of saved memories. You say "remember this", it lands there, you can read it in settings and delete it. Traceable, manageable, and limited for exactly that reason: the space is not infinite. When it gets tight, things are cleared out, and for a while now the system has done that itself, by its own judgement of what still matters.

By the way, you no longer need to say "remember this" at all. ChatGPT decides on its own what to write down, then tells you it has remembered something. Anyone still assuming it only saves on request is a few versions behind.

The second route is the more interesting one. ChatGPT draws connections out of your earlier conversations that never appeared on any list. Since the summer a background process handles this, which OpenAI calls Dreaming: it reads across many conversations and writes a summarised picture from them. It also updates itself, so "you are flying to Singapore in July" becomes "you went to Singapore in July" once the trip is over. The reason is stated in the announcement: saved notes go stale, and this is meant to compensate.

It is cleverly built. It just has two properties worth knowing.

The first: what ChatGPT knows about you is no longer what you put there, but what a system made of it. You can look at the summary. The raw notes it came from, not in that form any more.

The second is in OpenAI's own help pages and matters more than it sounds: saved memories are sent along with every single message, until you delete them. Not the relevant ones. All of them. On top of that comes the summarised profile, also in full, and also regardless of what you happen to be asking. Whether you ask for a recipe or an invoice, your entire stored self travels along.

And one more thing that surprises people: deleting a conversation does not delete what ChatGPT learned about you from it. That lives somewhere else and stays there.

How Claude Remembers#

Claude went the other way in the same summer, and that is not a detail for specialists. You feel it in daily use.

It, too, used to produce a summary once a day. In July that was dropped. The announcement says, word for word, that memory now works as a set of individual, categorised entries that Claude reads and updates during your conversations, replacing the previous daily summary.

Since August, everything remembered sits under the heading Topics in settings, where each item can be edited or deleted on its own. And a second point that matters for client work: a project in Claude has its own memory area. What is created inside a project stays there. ChatGPT has that separation too, but as a setting you have to switch on.

One more detail is worth having, because it sets up the next chapter. Claude can do two distinct things. It can search your old conversations, and you see in the chat that it is searching. That is a real search with hits. Separately it holds the remembered topics, which are not searched but loaded in automatically. So the searching happens in the conversation archive. Inside the memory itself, nothing is searched. It simply sits there.

Put plainly: one service turned individual notes into a diary, the other turned the diary into a card index. Five weeks apart, in opposite directions. Both have good reasons. But if you want to know what is stored about you, the card index is the more honest build.

Both Run Into the Same Wall#

Whichever of the two you use, there are things a built-in memory cannot do by construction.

It does not know since when something has been true. It remembers that your main client is such and such a firm. It does not remember that this has been the case since March and that someone else held the spot before. So you cannot ask it how things stood in spring.

It does not notice when something new supersedes something old. That was studied in May, across four hundred everyday situations. The failure is called implicit conflict: there is no explicit contradiction, the new information simply makes the old one obsolete. The best model tested caught it in roughly half of the cases. In daily life that means: if you move, change a price or lose a member of staff and never say so outright, the AI keeps working from the old picture.

It stores sentences, not connections. That is the most important of the limits and the hardest to see. More on that in a moment.

And it cannot be moved. Getting your data out is entirely possible, both providers offer a data export, and in Europe that is a legal requirement anyway. But such an export is a stack of paper: a file to keep, that no other tool can do anything with. Moving does not mean getting it out, moving means being able to read it back in somewhere else. And your colleague sees none of it regardless.

Sentences or Connections, That Is the Real Difference#

Picture two ways of keeping track of your clients.

The first is a stack of notes. One sentence on each. "Ms Berger is at Sonnenhof." "Sonnenhof belongs to the Nordlicht group." "Nordlicht never pays before the second reminder." All correct, all present.

The second is a pinboard. Ms Berger is up there as a photo, Sonnenhof beside her, the Nordlicht group above, and strings run between them. Works at. Belongs to. Pays late.

Ask both whether to expect a reminder with Ms Berger. The pinboard shows the route: two strings, done. The stack of notes does not know. The answer is on none of the notes. It has to be inferred from three of them, and that works sometimes and sometimes not.

That is exactly how a built-in memory operates. It has sentences. When you ask about a connection nobody ever wrote down as a sentence, the model has to reconstruct it while reading. Usually that goes well, and then it seems almost clairvoyant. Sometimes it does not, and the answer sounds every bit as confident. From the outside the two cases are indistinguishable, and that is the actual problem: not that it guesses, but that you cannot tell from the result that it did.

A memory with a graph, which is the technical name for the pinboard, stores the strings as strings. The connection is looked up, not worked out.

The difference starts with the name. On a note, "Ms Berger" is nothing but a word inside a sentence. It is not a thing that anything could hang off. Which is why "Ms Berger", "Berger" and "Sabine Berger" on three different notes are three different words, and whether the same person is behind them has to be guessed anew every time.

On the pinboard she is a photo. A thing in her own right, carrying every spelling she has ever appeared under, and everything ever said about her hangs off her. You can point at it and say: show me everything about her. On a stack of notes you cannot. There you can only search for the word and hope it was spelled the same way everywhere.

That is why a built-in memory shines with a handful of preferences and stops being enough for a grown client list. Preferences are sentences. Clients are things that connect.

And Every String Carries a Date#

Now the part that turns a nice idea into a tool.

Every string carries the date it was tied. And when it is cut, because Ms Berger changes firms, it does not vanish. It gets a second date and becomes a string that was true until then.

That lets you ask something impossible with notes: what did this look like in March. Not what do I know today about March, but what was the state of things in March. For a complaint, an old invoice, or the question of why you decided the way you did back then, that is the difference between an answer and a shrug.

Incidentally the same date solves the problem from the previous section. When a new string replaces an old one, that is a visible change with a timestamp, not a silent contradiction that somebody notices eventually or does not.

And the whole thing can be looked at. A pinboard you can take in at a glance. A stack of notes you have to read through. If someone can show you what their system knows about your firm by drawing it for you, then you also know what is missing.

The Whole Diary, or Just the Right Page#

Here is where it is decided whether such a memory is still worth anything after two years.

A list and a diary share the same underlying problem: every question brings everything along. Whether you ask about an appointment or an invoice, the memory puts its entire contents on the table and the AI picks out what fits.

While there is little in it, that works. It works beautifully, in fact, which is why the beginning with such a system is always the best time. But every line that gets added makes it a little harder. Not because space runs out, but because attention gets spread thinner. This has been studied, and the result is uncomfortable: accuracy drops noticeably well before the technical limit is anywhere near. A person asked to skim fifty pages remembers the beginning and the end and misses the middle. A machine is no different.

How this plays out in practice you can actually read up on with the developer tools, because there everything is in the open. The memory consists of text files. One of them is a table of contents with a line per note, and beside it sits a separate file per topic. At startup the table of contents is loaded, but only up to a fixed length. The maker states that anything beyond it is simply dropped on the next load.

The rest lives in the topic files, which are opened only when needed. That is the right way to build it. It also means the AI has to decide for itself which file to open, and that its only guide is one line in the table of contents.

Then there is what happens when the same thing is written in several places. Notes can sit at different levels: the company, you personally, the project, subfolders. The maker's own guide says what follows: if two entries contradict each other, the AI may pick one of them arbitrarily. And it recommends going through the files regularly to remove anything outdated or conflicting.

That is an honest sentence, and it describes the problem exactly. There is no guardian. There is a recommendation that a human tidies up regularly. In a business where everyone already has plenty to do, nobody does that after the third week.

A memory with a graph works the other way round. It carries nothing along. It is searched on every question, and only what fits the question is pulled out and passed on. The rest stays put without getting in the way.

That is the difference between a diary and a filing cabinet with an index. The diary gets heavier to carry every month. The cabinet can have grown for thirty years and you still pull out the one folder you need in seconds. Which is why a memory built this way can grow for years without getting worse. It gets better, because there is more in it that can be looked up.

What a Memory of Your Own Does Differently#

At this point I will use ours as the specimen, because I know what it looks like inside. This is not about that one system, it is about the way of building.

It throws nothing away when space gets tight. Something you have not needed for a long time slides towards the back. It does not disappear. That sounds like a small thing and it is the single most important difference, because "not looked at in a while" and "no longer true" are two entirely different statements. We had that very confusion built in for a time and took it back out, because knowledge became silently unfindable. Treat the two as the same and you lose correct facts without noticing.

It spots contradictions and flags them instead of resolving them quietly. When two entries do not fit together, that is a case for a person, not for an automatism. Only where it is unambiguous does it tidy up by itself; everything else lands on a list to be looked at.

It tells two similar events apart from one corrected statement. That is the subtlest part of it and the most underrated. Two quotes to the same client are two quotes, even if they read almost identically. A memory that goes by similarity alone takes the second for a correction of the first and discards one. We ran into exactly that and built in a catch that exempts events.

People and client projects do not decay with us at all, however long nothing has happened. A client you have not dealt with for two years is not a stale record. He is a client who gets back in touch.

It stays separated by area without you having to set anything up. What arises with one client does not surface with another.

And it can take in what you already have. That is the other half of the moving question, and it is the half people forget. The export files from ChatGPT, Claude, Gemini and a few others can be read in, and they become a starting stock. So you do not begin at zero, you begin with whatever has already accumulated over the past months.

That is what I find important about the moving question. Getting your data out works everywhere. The question is whether anyone out there will take it.

You Can Look at It, and You Can Look Back#

This deserves its own section, because it is what makes the whole thing tangible.

With the built-in memories you get a list. Sentences under each other, sorted by topic in Claude, as a dated list in ChatGPT, similar at Google. That is better than nothing, and with all three it is all there is. None of the big providers shows you what connects to what. There is no map, no timeline, no picture. Just text, one line under the next.

With ChatGPT there is the added point that the overview you see there may not match what the AI actually has in front of it. That is not an official statement but the finding of people who examined it from the outside, so I will write it as what it is: a well-founded suspicion, not a proof. It does fit the design, though, because a summary is regenerated each time.

A pinboard, by contrast, can be drawn. With us that is a map, with the things as dots and the connections as lines between them. You see at a glance where a lot hangs together and where something sits on its own. Anyone who has once looked at such a map understands immediately what a graph is and needs no further explanation.

And there is a slider for time. Push it back and you see the same holdings as they stood three months ago. Beside it, what has been added since and what has fallen away. That is not a gimmick. It is the only honest way to check whether a memory is genuinely growing with you or merely getting fuller.

For that second part I found nothing at the three big services. None of them lets you call up an earlier state. You see what is in there today, and what was in there yesterday is not a question the system can answer.

The Comparison at a Glance#

Where None of This Helps#

Now the part a vendor rarely writes.

For occasional use none of it is worth it. Someone who has a text reworded twice a week has everything he needs in the built-in notes, and any extra machinery only slows things down.

A memory also does not help when you want a fresh opinion. If you have spent months saying a particular approach is the right one, a well-kept memory will confirm exactly that. For an honest counter-argument an empty window beats a full one.

And writing more into it does not make it better. There is a study from February on precisely the kind of notes file most people set up first. The result was inconvenient: instructions in it are followed reliably, but general descriptions of your own company achieve nothing and cost noticeably more. Accuracy did not rise, the bill did. Translated: "write informally" works. "We are a family business in the second generation" does not.

What Happens After a Year#

Most of us have only been working seriously with these tools for a few months. So hardly anyone has seen what such a memory looks like further down the line. From what has been measured, though, it is easy enough to predict.

It does not break suddenly. It gets vague. Old entries sit beside new ones without anybody having decided which one holds. Part of it has quietly gone because space was needed. And the more that is loaded in automatically with every question, the likelier the model is to miss the one thing that mattered. More is not better. The right thing is better.

Anyone who wants to keep a grip on that needs two things. He has to be able to see what is in there. And he has to be able to decide what stays. With a built-in memory both are partly true. With one of your own, fully.

What I Would Tell You to Do#

Go and look. Both services show you what they have noted about you, in settings under memory. It takes two minutes and tells you more than any article about it.

If there are things in there that are no longer true, delete them. If nothing is in there that would matter to you, then the AI knows you less well than it feels, and what you took for memory was the open conversation window.

And when you notice that you are still explaining the same things over and over, that is the point at which a memory of your own pays off. Not before.