All articles
3 min read

Why your AI gets dumber in long chats

Your AI has a memory jar with a fixed size. When it fills up, the oldest stuff falls out. Here is what is happening and six ways to fix it.



The short version

Your AI has a memory jar. It is a fixed size. Everything you say goes in it.

When the jar fills up, the oldest stuff falls out the bottom. Usually that is the instructions you gave at the start.

So it did not get dumber. It just cannot see the beginning of your chat anymore.


What is actually taking up space

All of this fights for the same jar:

What goes inHow much space
Your custom instructionssmall, but always there
Your first messagesmall
Any file you pastedusually the biggest problem
Every reply the AI gave yougrows fast
Error logs and search resultshuge, and mostly junk
Your latest questionsmall

Here is the part people miss. The AI re reads the whole jar every single time you send a message. That is why long chats also get slower and cost more, not just worse.


6 things that fix it

1. Just start a new chat

Boring advice. Still the best thing on this list by a mile.

If the chat has gone bad, nothing you type will fix it. The bad stuff is already in the jar. Start fresh.

2. Stop pasting whole files

Paste the 20 lines that matter, not the whole 2000 line file.

One giant error log can eat half your jar and push out the instructions you actually cared about.

Bad: paste the entire stack trace Good: paste the 5 lines around the actual error

3. Put your rules in a file, not in the chat

Chat messages fall out of the jar. Files get loaded fresh every time.

  • Claude Code: use CLAUDE.md (keep it under 200 lines, longer files actually get followed less)
  • Cursor: use .cursorrules
  • ChatGPT: use Custom Instructions or Project instructions

4. Ask for a summary, then restart

When the chat is going well but starting to slip, do this:

Summarise everything we decided so far as a bullet list.
Include the decisions, the rules, and what is still left to do.

Copy that. Open a new chat. Paste it as your first message.

You keep the useful stuff and throw away the noise.

5. Do your research somewhere else

If a task means reading a lot of stuff you will never look at again, do it in a separate chat and bring back only the answer.

In Claude Code that is a subagent. It fills up its own jar instead of yours.

6. Learn the warning signs

Your jar is full when the AI:

  • repeats something it already said
  • ignores a rule you gave earlier
  • forgets a file you pasted
  • argues against something you both already agreed on

Do not try to correct it. Correcting it just adds more to the jar. Restart.


The thing nobody tells you

Bigger context windows did not fix this problem. They just made the jar bigger.

AI models still get less reliable as the jar fills up, even before it is completely full. Stuff in the middle of a long chat gets ignored more than stuff at the start or end. So a huge context window does not mean the AI actually remembers all of it properly.

The skill is not finding the AI with the biggest jar. It is keeping your jar clean.


Save this bit

CHAT GOING BAD?

repeating itself        -> start a new chat
ignoring your rules     -> move rules into a file, restart
forgot your file        -> paste only the part that matters
chat is good but slipping -> ask for a summary, restart with it
about to paste a log    -> paste 20 lines, not 2000

I post shorter versions of this on Instagram, most days.

@theprocrastihacker