Why Your AI Agent Gets Dumber After 30 Minutes

You start a chat at 9am. The agent is sharp. You ask for a small function, and it gets it right. You ask for a refactor, and it understands your files.

By 11am, the same chat feels broken. The agent forgets a setting you made an hour ago. You ask for a small change, and it rewrites the whole file. You point at a bug, and it fixes code that is not even in your project.

The model did not change. Your context window filled up.

What is a context window?

The context window is everything the agent can see right now. Your messages, the files it read, and its own past answers all live there.

Think of it as a desk. Every message adds paper to the desk. The bigger the pile, the harder it is to find the one page you need. When the pile grows too big, the agent reads junk next to gold, and the junk starts to win.

Why longer chats get worse

Most people think more context means better answers. The agent sees more code and more corrections, so it should get smarter. The opposite happens.

Every word in the window pulls on every answer. Old mistakes, dead ideas, and fixed bugs still sit there. They pull the agent in directions you did not ask for.

Engineers call this signal and noise. Signal is the useful stuff. Noise is the junk. When noise grows, the good stuff gets harder to see. From the outside, it looks like the agent lost its memory. Really, it lost the signal.

Three simple rules

Rule one: every model has a size limit. Past that limit, the chat cannot grow anymore, and you must start fresh.

Rule two: nothing sits in the window for free. Dead code and failed tries still pull at the next answer.

Rule three: a small window often beats a big one. The agent works best when it sees only what it needs.

Change how you work

If you run one long chat for the whole project, you pay in quality. The fixes are simple, but they need new habits.

Start a new chat for each task. A bug fix does not need to remember yesterday's refactor.

Cut dead ends fast. When you drop an idea, rewind the chat to before that idea. A failed try left in the window will leak into the next answer.

Move big text into files. Do not paste ten pages of docs into the chat. Put them in a project file, like AGENTS.md, and point the agent to it.

When long chats still make sense

Long chats are not always wrong. A debug session where each step needs the last one is a good case. So is exploring a new problem together.

Ask one question: is this pile signal or noise? New clues are signal. Three failed attempts are noise.

Reset when the noise wins. The agent will not say it is confused. It will give answers that sound right but are wrong. By the time you notice, you lost an hour.

The simple test

If you cannot remember the whole chat, the agent cannot either. Start a new one.