How a chat remembers

What does a chat actually send each time?

An AI assistant has no memory of its own between messages. A chat tool makes it look like it remembers by sending the earlier messages along with your new one. When you type a follow-up, the assistant receives the whole conversation so far, reads it again from the top and writes the next reply.

That works well for short chats. Your follow-up can say “now make them look like buttons”, and the earlier messages explain what “them” means. Problems start when the earlier details are many, far back or easy to outweigh. The newest message gets a lot of attention, and a short follow-up gives the assistant little to hold on to.

Why do details get lost in a follow-up?

There are two common reasons. The first is weight. Your newest request is short and about buttons, and the assistant’s idea of a typical button, bright, pill shaped, maybe with a library for effects, can pull harder than a colour you mentioned three messages ago.

The second is space. Every model has a limit on how much text it can take in at once, often called its context window. In a long chat, especially one with big pasted files, the oldest messages may be shortened or dropped to make room. When that happens, your early instructions are gone, and nothing on the screen tells you.

Both lead to the same result: a reply that ignores something you said earlier, delivered with full confidence.

What happens in the practice area?

The Code tab holds the result of an earlier message in this chat, where you asked for your links in plain HTML and CSS, no JavaScript, using your green #2f6b4f. Now send a short follow-up: “now make the links look like buttons please”.

The reply is bright blue pill buttons. It loads a button library from another server and runs a script, though you said no JavaScript. It also writes the links from scratch and drops the email link. Nothing in the reply mentions any of this. The assistant answered the question in front of it, and the question in front of it did not mention colour, limits or email.

The pretend assistant models this in a plain way: it only gives the good reply when your follow-up itself carries the important details. Real tools are less predictable, sometimes keeping everything and sometimes losing it, which is exactly why the habit helps. You cannot tell in advance which follow-up will lose something, so the safe move is to make every important follow-up able to stand on its own.

How do you write a follow-up that keeps the details?

Restate the things that must not change. You do not need to paste the whole first message again. One sentence covering the colour, the limits and anything the reply must keep is usually enough. If the change touches code, paste the current version of that code too, because the code may have changed since the start of the chat.

Send it and open **Prompt check**. The Recap row is ticked because you repeated the details that matter. The reply keeps your green, stays in plain CSS and leaves all three links in place.

Now please make the links in my Find me list look like buttons. Keep everything from before: plain HTML and CSS only, no JavaScript, my green #2f6b4f, and keep all three links including the email one. Here is my current CSS: .links a { color: #2f6b4f; }

What is still wrong with the good reply?

Read the CSS after you press **Use this code**. The last rule is .links a:focus { outline: none; }, and the reply describes it as making the buttons look cleaner. It hides the ring that shows keyboard users which link they are on. Someone who moves through the page with the Tab key now has no idea where they are.

Click into the Preview and press Tab to see it. With the rule in place, focus moves invisibly. Delete the line and the outline comes back. Use the **Review** view to confirm that this deletion is the only change you made to the assistant’s code. Assistants remove focus outlines often, because many older examples do it for looks.

When should you start a fresh chat?

Start fresh when the chat has grown long, when it has wandered across several unrelated problems, or when the assistant keeps repeating an old mistake. A new chat has no stale messages pulling replies in the wrong direction.

Begin the new chat with a short summary: what you are building, the limits, the current code for the part you are working on, and what you want next. Writing that summary takes a minute and usually gets a better reply than the tenth follow-up in a crowded chat. Keep a small text note of your standing limits so you can paste them each time.

I'm building a one-page profile site in plain HTML and CSS, no JavaScript. My accent colour is green #2f6b4f. Here is my current links section: ... Next I want to make the links look like buttons.

Does the assistant remember me between chats?

Usually not, unless the tool has a memory feature that saves notes about you. Some tools do, and you can often view or delete what they saved. Treat any such memory as a convenience, never as a guarantee. If this reply depends on a detail, put it in this request.

The same goes for your files. An assistant does not see your project folder just because you talked about it last week. Paste the current code each time the reply depends on it, especially after you have edited the file yourself, since the assistant’s last version is now out of date.

How is the pretend assistant different from a real chat?

The assistant on this page does not read earlier messages at all. It picks a reply from the request you just sent, which is why it can show the effect of a lost detail so clearly. A real chat does send earlier messages, so it keeps details more often than this one, and still not always.

Real tools also vary their answers, write longer replies, sometimes refuse, and may keep what you paste. The habit you practised works with all of them: in any important follow-up, repeat the details that must not change.

Field notes

• A chat sends the earlier messages along with your new one each time you ask.
• A short follow-up can lose details, because the newest request pulls hardest.
• Restate what must not change in one sentence, and paste the current code if it changes.
• Never hide the keyboard focus outline, since keyboard users need it to see where they are.

Easy to get wrong

A short follow-up gives the assistant little to hold on to, and earlier limits can be outweighed or dropped. One sentence of recap protects them.

Old messages can be trimmed or can keep steering replies the wrong way. A clean start with the facts that matter is often faster.

The assistant only knows what is in the chat. Your edits since then are invisible to it, so it may undo them.

Questions people ask

How long can a chat get before details are lost?

It depends on the tool and on how much code you paste. Long files fill the space quickly. Instead of counting, watch for replies that ignore earlier instructions and recap or restart when you see one.

Do I have to repeat everything in every message?

No. Repeat the details the next change could break, such as your colour, your limits and anything that must stay. Small questions about the last reply rarely need a recap.

Why does removing the focus outline matter so much?

People who use a keyboard rely on it to see which link or button is active. Without it, the page is close to unusable for them.

Can I edit an earlier message instead?

Many tools let you edit and resend an earlier message, which starts the conversation again from that point. That is a handy way to fix a weak first request.