loading
loading
The quality drop when a context window gets stuffed too full and the model loses track of what matters.
Long conversations and big dumped documents fill the window with stale or irrelevant tokens, and the model starts ignoring key instructions, contradicting itself, or losing earlier facts. It matters because more context isn't always better — past a point it actively degrades answers and wastes tokens. Example: an agent 40 messages deep forgets a constraint from message 2, so you fix it by summarizing old turns and pruning dead tool output instead of carrying everything forward.
Plainly
Think of Context Window Rot as a simple recipe for doing the work better. The quality drop when a context window gets stuffed too full and the model loses track of what matters.
In practice
Use it when you need a repeatable method instead of guessing from vibes. In practice, define the owner, input, output, and failure mode before you rely on it.