Bring what you know AI guides in plain language

Foundations  /  How-to

Why it forgot what you told it

Long conversations quietly get worse. The tells, and how to move to a fresh one without losing the work.

You’re forty messages into a conversation. You told it near the top to keep answers under 200 words, and the last three have been essays. You ask it to fix one thing and it changes something else. It starts summarising what the two of you have discussed instead of answering the question you asked.

The usual conclusion is that the tool got worse today. It didn’t. The conversation got long.


What’s actually happening

Nothing you say is filed away somewhere it can look up later. Every time you press send, the whole conversation so far gets handed back to it, from the top, and it reads all of it before answering. That is the only memory of you it has.

There’s a ceiling on how much can be handed over at once. When a conversation grows past it, something has to give. Some tools drop the oldest parts. Some compress them into a summary. Some tell you. Many give you no signal at all.

But the dropping is the late stage. The real problem starts much earlier, and it’s simpler: the rule you set in message two is now competing with two hundred messages of newer material. It doesn’t need to be deleted to stop mattering. It only needs to be outnumbered.


The tells

None of these are proof on their own. Two or three together mean the thread is done.

What you seeWhat’s usually behind it
A rule you set early stops being followedOld instruction, outnumbered by newer text
It repeats an option you already rejectedThe rejection is buried too far back
Answers get longer and vaguerIt’s averaging across a sprawling thread
It contradicts something it told you earlierBoth versions are in there and it can see both
It summarises the conversation instead of answeringIt’s treating the thread itself as the subject
It’s lost a file you attached earlierAttachments live in the conversation, not on a shelf

The last one surprises people most. An upload is part of the thread, not a permanent attachment to your account. Long threads lose files the same way they lose instructions.


The fix is a new conversation, not a better prompt

When a thread starts behaving like this, the instinct is to push harder. Repeat the instruction in capitals. Explain that you already said this. Write a longer, firmer prompt.

That makes it worse. A longer prompt inside a full conversation is more material competing with itself.

Open a new conversation instead. The cost is the context, and the context is recoverable.


Carry the context across yourself

Before you abandon the thread, ask it to pack for you. Paste this as the last message in the old conversation:

This conversation has got long and I'm starting a fresh one.
Write me a handover note I can paste into the new conversation.

Include:
- What I'm trying to produce, and who it's for
- Any rules or preferences I've given you
- Decisions we've made, and the options we ruled out
- Where we got to, and what's still open

Facts from this conversation only. Don't add advice, and don't
summarise your own explanations back to me.

Then read it before you paste it. It’s a summary of a long document, which makes it exactly the kind of output step three is about. Check the specifics, especially the decisions. A handover note that quietly reverses one of your choices will carry that reversal into everything that follows.

The ruled-out options matter as much as the decisions. Without them, the new conversation cheerfully proposes the thing you spent an hour rejecting.


One conversation per job

The habit worth building is smaller than it sounds. Most people run one long thread per day, or per week, and pour everything into it. Run one per job instead, and start a new one when the job changes.

It feels wasteful. It isn’t. There’s nothing to preserve by staying put, because nothing accumulates in your favour. A thread is a working surface, not an account.


When it breaks

It claims to remember you from a previous conversation. Some tools do have a memory feature that carries facts between threads, which is a different mechanism and the subject of the next guide. If you haven’t set one up, treat “as you mentioned before” in a fresh window as invention and ask what it thinks you said.

The new conversation is worse than the old one. The handover note was too thin. Go back, if the old thread is still open, and ask specifically for the constraints and the rejected options. Those are the parts people leave out.

You can’t start over, because the thread contains something you need. Ask for it verbatim first. “Output the current draft in full, no commentary.” Take it with you.

It went wrong in a short conversation. Then length isn’t your problem and this guide won’t help. That’s usually a thin prompt, so go back to step two.

The thread is enormous and working fine. Leave it alone. Length raises the risk, it doesn’t guarantee the failure. There’s no reason to restart on a schedule.

It got worse right after you attached a large file. A big document takes up much of the room by itself, which can push a long conversation over the edge in one move. Start fresh and attach the file first, before the discussion.


What to take from this

Every other tool you use gets better as you go. Files accumulate, settings stick, the software learns where you keep things. The instinct that context builds up in your favour is correct almost everywhere else.

Here it’s reversed. A conversation is at its sharpest early and degrades as it fills. Knowing that turns a frustrating hour into a thirty second decision: this thread is done, pack it up, open a new one.

Next