Why ChatGPT forgets long conversations — and what actually helps

By Niko Nishimine · Updated July 2026

You're deep into a long ChatGPT conversation — a project you've been building for weeks, a draft you've revised twenty times — and it happens: the model contradicts a decision you made together forty messages ago. Or asks you a question you already answered. Or the whole thread gets so slow you dread opening it.

This isn't a bug, and it isn't your imagination. Here's what's actually happening, and every real option for dealing with it.

The context window: the model can only "see" so much

Every AI model has a context window — a hard limit on how much text it can consider at once. Your entire conversation (plus the model's instructions) has to fit inside it. Long before your thread feels "old," it stops fitting.

When that happens, ChatGPT doesn't warn you. It quietly drops or compresses the earliest parts of the conversation. The messages are still on your screen — you can scroll back to them — but the model can no longer see them. You're reading a shared history; the model is reading the last chapter.

The symptoms, in the order they usually appear

• Slowness. Very long threads take noticeably longer to respond, and the browser tab itself starts to lag.
• Repetition. The model re-asks things you settled earlier.
• Contradiction. It violates constraints or decisions from the early conversation.
• Personality drift. Style and tone instructions from the start fade away.

What people usually try (and the trade-offs)

1. Start a new chat and re-explain. Works, but you pay a re-explanation tax every time — and details always get lost in the retelling.
2. Ask for a summary, paste it into a new chat. The best manual technique. Ask the model to summarize all decisions, constraints, and open threads; start fresh with that summary. Downsides: it's manual, you have to remember to do it before quality degrades, and the old thread becomes a graveyard you can't search well.
3. Use ChatGPT's built-in memory. Helpful for small durable facts ("I prefer concise answers"), but it isn't designed to hold a project's worth of context, and you don't control exactly what it retains.
4. Custom instructions / Projects. Good for standing preferences and grouping, but neither prevents an individual conversation from outgrowing the window.

The structural fix: carry the context, not the thread

The manual summary technique points at the real answer: what matters isn't keeping one endless thread alive — it's making sure the context survives the hand-off to a fresh session. Done well, a "new" conversation with the right summary plus your key reference material behaves like the old one, minus the slowness and forgetting.

That's tedious to do by hand every time. It's exactly the thing software should do for you.