Video18/18models did better on the short version
Your AI didn't get dumber. Your chat got longer.
TL;DR Every model tested did worse on a long chat than on a short one. Start fresh and bring a summary, not the whole history.
Your AI didn't get dumber. Your chat got longer. Researchers gave 18 models the same question two ways: the full 113,000-token chat history, or just the 300 tokens that mattered. Every single model did better on the short one.
Why? More context means two jobs instead of one: find what matters, then think. The finding part is where it fails. And AI still has no common sense about which details are important.
Try this
When a chat starts compressing, start a new one. Bring a short summary, not the whole history.
Sources: Chroma, “Context Rot” (2025); NoLiMa, ICML 2025.
