Why AI tutors forget

Ask any AI tutor the same question on Monday and again on Wednesday, and you’ll get the same answer both times — word for word, in most cases. Not because the model is stubborn. Because it genuinely doesn’t know it already answered that question once.

This isn’t a bug in any particular product. It’s a property of how these tools are built. A chat completion call sees exactly one thing: the conversation history you hand it in that request. Close the tab, and that history is gone unless something outside the model call chose to keep it.

The gap nobody notices until it matters

For a single, self-contained question — “what’s the capital of France” — this doesn’t matter. The model doesn’t need to remember you asked something similar last week.

Tutoring is different. A student’s understanding of a topic is built over many sessions, not one. Whether they’ve already been shown a particular explanation, which one worked, what they got wrong last time and how — none of that lives inside a single conversation. It lives in the pattern across conversations. And a model that only sees one conversation at a time has no way to see that pattern.

The result is a specific, recognizable failure mode: a tutor that re-explains from scratch every time, repeats a hint that already failed once, or asks a student to re-describe a confusion they already described clearly a few days ago. None of this looks broken in the moment. It just quietly signals that nothing is being learned about the student — only about the question.

Why the obvious fixes don’t fully solve it

The first instinct is usually “just give it a bigger context window” — paste more history into the prompt. That buys time, not a solution. Conversations still end. Sessions still get truncated once they’re long enough. And a context window is, by definition, temporary — it exists for the duration of one call, not across the weeks a student is actually learning something.

The second instinct is retrieval — pull in relevant documents so the model has more to work with. That helps when the missing information lives in a document. It doesn’t help when the missing information is “what does this specific student already understand,” because that was never written down anywhere to retrieve.

What’s actually missing is a place for that understanding to live between conversations — updated after each one, read back before the next. Not a bigger window into the current conversation, but a persistent, structured record of the student that survives after the tab closes. That’s a different kind of infrastructure than a chat feature usually ships with, which is exactly why so many AI tutors — good ones, technically — still start every session from zero.