Why AI "Forgets" Things During Long Conversations

There is no persistent memory to begin with

It helps to start from the actual mechanism rather than the word forgets, which implies something was stored and then lost. A model has no memory between requests at all; each response is generated fresh from whatever text is currently in its context window, and nothing else.

What feels like forgetting is really one of two things: the earlier information got pushed out of the context window as the conversation grew past the limit, or the information is still technically present but far enough from the model's current point of attention that it gets weighted less heavily when generating a response.

Why long conversations specifically make this worse

Every turn adds tokens: your messages, the model's replies, and in an agent setting, tool outputs like file contents and command results, which are often the largest contributors of all. A single file read or test run can add more tokens than several turns of conversation, so agent sessions fill their context window faster than plain chat.

As the window fills, tools typically start dropping or summarizing older content to make room for new turns, and an instruction you gave early on can be exactly what gets cut. This is a structural consequence of a fixed-size window, not the model choosing to disregard something.

• No persistent memory exists between requests; only current context matters
• Tool outputs in agent sessions often consume more tokens than chat turns
• Older context gets dropped or summarized once the window fills up
• Information far from the model's current focus gets weighted less heavily

Working around it

Restate important constraints periodically in long sessions rather than assuming an instruction from twenty turns ago is still in effect, especially for anything you genuinely need respected consistently, such as style rules or hard constraints.

For genuinely long-running work, it is often more reliable to start a fresh session with a concise, explicit summary of what matters than to keep extending one conversation indefinitely and hoping the important parts survive.