One chat, one job: why long AI chats get worse

You opened one chat last Sunday to understand a chapter. Since then you have used that same chat for a job form, a recipe and your cousin’s phone problem. Now it answers your chapter question with a piece of the job form attached.

Nothing is broken. Your phone is fine and your free plan is fine. Long chats simply get worse, and this lesson shows you what to do about that.

The signs that a chat has gone stale

You will recognise most of these.

• It forgets what you told it. You gave your class and district at the start. Later it asks again, or guesses wrong.
• It contradicts itself. What it says now does not match what it said earlier.
• The answers get vaguer. They grow longer and safer. They stop being about your actual question.
• It mixes your topics. Your maths doubt comes back with a line about your job letter.
• It ignores part of your instruction. You asked for short Hindi sentences. You get long English ones.

This has been tested properly. One 2025 study ran more than 200,000 test conversations across 15 AI models. Spreading a task over many turns of chat made the answers about 39 percent worse on average. The same study found that once the model takes a wrong turn, it does not find its way back on its own.

Why a long chat gets worse

Here is the part nobody tells you. The app is not sitting there remembering you between messages.

Every time you press send, the whole chat is sent again from the beginning. Your first question, its first answer, your second question, all of it, and then your new line.

So the pile grows with every turn. Before it writes one word, the model has to read that whole pile again.

Reading it is not free. Every piece of text gets compared with every other piece. Double the text and the work goes up about four times. Its attention spreads thin over old material you stopped caring about days ago.

This pile has a name. It is the context window, the working memory of one chat. It holds your messages, its own answers, and any photo or document you added. It is not the same as what the model learned during training. Module 3 has a full lesson on why it forgets what you said.

A bigger memory does not mean better reading

Apps now advertise very large memories. That number is a size, not a promise.

In July 2025 a research team tested 18 leading models. Accuracy dropped as the input grew, long before the stated limit was reached. A short focused message beat a huge one carrying the whole history. That held true for every family of models they tried.

There is a second thing. When the pile finally will not fit, something has to go. Many chat apps quietly drop the oldest messages first. Others replace the early part with a short summary.

Either way, what you said at the start is gone or blurred. You did nothing wrong. The chat outgrew its window.

One chat, one job

So here is the rule. One chat, one job. A new topic means a new chat.

One chat for the physics chapter. A separate chat for the job application. A separate one for the electricity bill question at home.

Two things make this work. Old topics sit in the pile as distractions, and even one unrelated distraction lowers accuracy. Each chat also stays short, so the model reads it properly.

What drift actually looks like

Take a class 10 student in a village in Uttar Pradesh. Call him Suresh.

On Monday he opened a chat and typed his situation. Class 10, UP Board, explain light and reflection in simple Hindi. For three days the answers were good.

On Thursday, in that same chat, he asked for a job application for his elder brother. It wrote a decent one. He asked for two more drafts of it.

On Friday he went back to physics in the same chat. He asked for five practice questions on mirrors.

What came back was three questions on mirrors, one question about the job application, and an offer to write a covering letter. Two of the mirror questions were far above class 10. When he asked which class he was in, it said class 12.

He did not argue with it. He moved the work into a fresh chat, and the next answers were right again. That move is the most useful thing in this lesson.

How to carry your work into a new chat

You met this move in the lesson on correcting it when the answer is wrong. It is a handover. Use it whenever a chat has gone stale.

In the old chat, type this and send it.

Write everything I have told you in this chat as one clear request.
Do not answer it. Just write the request.

Copy what it gives you. Open a new chat. Paste it in, add anything it left out, and send.

A good handover is short and specific. It should look close to this.

I am in class 10, UP Board, in a village in Uttar Pradesh.
I am studying light and reflection.
Explain in simple Hindi, with sentences under 15 words.
Then give me 5 practice questions at class 10 level.
Put the answers at the end.

Do not paste the whole old conversation into the new chat. That rebuilds the same problem on day one. Carry only the facts the new chat needs. The lesson on how to ask so you get a useful answer tells you which facts those are.

Name your chats so you can find them again

One chat, one job only works if you can find the right chat tomorrow.

Most apps keep a list of your old chats. Look for an icon that slides a panel in from the side, or a menu with a word like history or recent in it. If you are not signed in, that list is usually empty. Logged out, nothing is saved for you. So copy any answer you want to keep before you close the chat.

The app names each chat by itself, using your first message. Those names are usually useless. You end up with five chats called something close to “Explain this”.

Look for a way to rename. In most apps you press and hold on a chat in that list. In others, a small row of extra options appears beside the name. Rename it to words you would actually search for. “Class 10 light chapter” beats “Explain this”.

Many apps also let you pin or star one chat so it stays at the top. Look for the word pin or star in the same set of options.

If there is no rename option at all, use a simple trick. Make the topic the first line of your first message. Then the name the app picks is at least about the right thing.

Short chats save your data and your battery

There is a second reason to keep chats short. It saves you money today.

Every send carries the whole chat to a computer far away, and brings an answer back. A long chat sends more text on every single turn. On a small daily data pack that adds up fast. Every extra send also costs battery. A long chat sends more, so it drains your phone faster.

Two things fill the window quickest. Photos are one. A clear photo of a page can cost the model about as much as a thousand words of English. So giving it your notes and photos is best done in a chat kept for that one job.

Hindi typed in Devanagari is the other. The same sentence in Devanagari costs the model several times more than in English. This is getting better as apps update, but it is still true today.

So one focused chat is cheaper as well as more accurate. If you like, ask in short English and add a line asking for the answer in Hindi. Using AI in Hindi covers that choice, and using AI without finishing your data covers the rest of your pack.

Where this leaves you

Carry one idea away from here. The chat is a table, not a brain. Everything you put on that table has to be read again before every single answer.

So keep the table clear. One chat, one job. When it goes stale, take the handover and start again.

Next comes a lesson that takes work away from you. Many popular tricks for talking to AI have now been tested, and they do nothing at all. Prompt tricks that do not work tells you which ones to drop.