My AI gets worse the more I use it.

I use AI every day.

I dictate my thoughts into Letterly, then run them through Gemini, Claude, and ChatGPT.

And I've noticed something. The longer I use the same session, the worse the output gets. It starts sharp. By the end it's sloppy.

Then I noticed something else. Even when I start a fresh chat, the quality is still off. Every few months I do a hard reset of my memory and saved instructions. Suddenly it's good again.

I thought I was imagining it.

I wasn't.

There's a name for it. Context rot.

In 2023, Liu et al published a paper called "Lost in the Middle." They found that as a model's context window fills up, it stops paying equal attention to everything you've given it. It favours the beginning and the end. Everything in the middle gets lost.

By 2025, multiple research teams confirmed it goes further than that.

Paulsen (2025) showed it happens across all kinds of tasks, not just simple ones, and kicks in faster when the work is complex.

Du et al (2025) ran a clever test. They stripped out everything except the key information to remove any retrieval difficulty. The degradation still happened. It's not that the model can't find what you gave it. It's that the more you give it, the worse it performs. Full stop.

Here's what that means for you.

Long sessions work against you. Bloated memory and saved instructions work against you too.

Reset your sessions regularly. Clear your memory every few months.

Treat it like a tool that needs servicing, not a colleague who remembers everything.

The longer you leave it, the worse it gets.