Guides / Why assistants forget

Why does ChatGPT or Claude forget? And what actually fixes it

Stignova · 19 October 2026 · 6 min read

Reason 1: the model does not learn from your chats

It feels as if the assistant gets to know you over weeks. It does not. The model is trained once, before you ever use it, and your conversations do not change it. What looks like learning is the assistant re-reading the conversation so far, every time it answers.

So the knowledge you built up lives in exactly one place: the text of that conversation. Close it, start a new one, or let it get deleted, and that knowledge is gone.

Reason 2: every conversation has a limited memory

The text an assistant can consider at once is called its context window. It is large, but not endless. In a long conversation, older parts get shortened or left out to make room, and the details you gave in week one are the first to go.

Even before the limit, more text is not always better. Anthropic, the maker of Claude, describes context as a limited resource: the longer it gets, the worse the model recalls what is in it. They call this "context rot". A very long chat is not a good memory; it is a crowded one.

Reason 3: your instructions are buried or contradict each other

"Always write short emails" in message 12 and "give me more detail" in message 80 pull in opposite directions. Claude's own documentation is blunt about it: when two instructions contradict each other, Claude may pick one arbitrarily. Rules that are scattered across a long chat slip for the same reason.

Doesn't the built-in memory solve this?

Partly. Both ChatGPT and Claude now have memory features. Claude, for example, saves short notes on topics as you chat, keeps a separate memory for each project, and on paid plans can search your earlier chats.

That helps, but two things stay out of your hands:

• The assistant decides what to keep. It writes and rewrites the notes itself. You can review and delete them, but you are not the author.
• Constantly rewritten memory degrades. A May 2026 study found that when a model keeps updating and summarising its own memory, usefulness first rises and then falls, sometimes below having no memory at all. The authors recommend keeping the raw facts and summarising rarely and on purpose.

What actually fixes it

Treat your assistant's memory like any other important document: write it down, in a place you own.

1. Put the lasting context in files. Who your clients are, what you charge, how you write, what you never want. A short plain text file the assistant reads before every task beats a long chat every time.
2. Keep the files short and specific. Claude's documentation recommends staying under about 200 lines per instruction file, because longer files reduce how well the rules are followed.
3. Separate what is true now from what used to be true. Old prices and finished projects belong in an archive, not next to today's facts.
4. Remove contradictions. Read your files every now and then, and delete the rule that lost.
5. Change the files deliberately. When the assistant proposes an update, you approve it. That keeps the memory from slowly drifting.

In Claude, the standard place for this is a file called CLAUDE.md. Our guide CLAUDE.md explained for people who don't code shows how to write one, with a template.