Five Years of Memory: What ChatGPT and Claude Now Keep on You

You open a new conversation with your assistant. You do not mention that you write in English, that you have been working on the same project for three weeks, that your daughter is called Lou, or that you avoid gluten. You do not have to: it already knows. The page is blank, but the memory no longer is.

It is one of the quietest, and deepest, changes to hit conversational assistants in a year. For a long time, every exchange started from scratch, like meeting a polite but amnesiac stranger. Now the machine remembers. And that shift, trivial as it looks, moves a great deal: the time you spend, the effort you put in, and what you leave behind.

The End of Amnesia

The turn came in two stages. In April 2025, OpenAI let ChatGPT draw on all past conversations, not just a short list of saved facts. A mechanism called "dreaming" sorts in the background what is worth keeping. First reserved for paying users in the United States, the feature reached free accounts by June of the same year.

Anthropic followed, then went further. In early 2026, persistent memory was switched on for every Claude account, free and paid alike. The system works by "synthesis": roughly every twenty-four hours, it distills recent conversations into durable summaries, filed in a profile and reloaded automatically into each new exchange. The assistant does not reread your whole history on the fly; it keeps a condensed version, which it consults without being asked.

The difference from the old model is not one of degree but of kind. A list of preferences you fill in yourself remains a tool. A memory that feeds itself, continuously, from everything you say, becomes a portrait. And that portrait the assistant carries from one conversation to the next, with no visible effort.

No Longer Repeating Yourself

The benefit is tangible, and it fits in a phrase: never explain again. The developer picks up a project without pasting the context back in. The translator no longer restates a preferred style. The parent asking for a dinner idea need not repeat, every evening, what their child refuses to eat. Each conversation resumes where the last one stopped, and the time saved, minute after minute, adds up.

This comfort has a market value. The AI agent memory sector was worth about 6.3 billion dollars in 2026, with projections above 28 billion by 2030. Memory is no longer a selling point: it has become a baseline expectation. An assistant that forgets everything between two exchanges is judged, in 2026, like a phone with no contact list.

At bottom, what memory restores is a form of autonomy. You hand the machine the job of holding the thread, recalling the details, picking up an idea left hanging. The user no longer has to be the keeper of their own context. That is exactly what makes the tool more useful, and also what makes it harder to leave.

A File That Does Not Erase

For this memory has a flip side, written into how it works. To remember you, the system must retain. What it distills from your conversations does not live only for the length of an exchange: for accounts whose data trains the models, the default retention period rose, in 2025, from thirty days to five years. The profile that serves you is also a file, and that file stays.

There is also a question of visibility. The two leading assistants handle their memory differently. Claude flags when it draws on something it stored ("based on what you told me last week about your project..."), so you know what is shaping its answer. ChatGPT folds the remembered context in silently, with no sign that a fragment of your past is steering what it tells you. In one case the memory is legible; in the other it works behind the scenes.

Dependence, finally, shifts. The more the assistant knows about you, the more it costs to leave: switching tools means going back to the polite, amnesiac stranger of before. Value no longer rests on the model alone, but on everything it has gathered about you, and that you cannot take elsewhere.

When Memory Gets It Wrong

Then there is the most insidious risk: that the memory is false. A memory is only worth as much as it is accurate, and the assistants' memory is not always so. A fact that was true six months ago may be stale today, and the machine serves it up with the same poise. Specialists speak of "stale memory served with confidence": not a spectacular hallucination, but yesterday's truth presented as today's.

The danger does not stop at the machine. Work presented at the CHI 2026 conference shows that users gradually confuse what they produced themselves with what the AI suggested to them. Researchers at the University of Exeter describe how these prolonged exchanges can distort memories and blur the story we tell about ourselves. When an outside memory, fluent and self-assured, mixes with our own, the line fades.

The confident tone makes it worse. We are built to believe what is expressed calmly and with certainty. An assistant that errs while stammering invites suspicion; an assistant that errs in a steady voice, leaning on a "memory" of your exchanges, lulls it to sleep.

Taking Back Control

None of this is suffered without recourse. You can pause the memory without erasing what exists. You can reset it, and delete for good what was kept. For a sensitive conversation, temporary mode leaves no trace. The controls are there; you just have to know it, and get into the habit of looking.

The assistants' memory is neither a trap nor a gift: it is a trade. It gives back time and attention by retaining, by archiving, sometimes by getting it wrong. The real move, in 2026, is no longer learning how to talk to the machine. It is deciding, now and then, what you let it keep, and what you would rather it forget.