Attention Degradation
Degradasi Atensi
Each token has a limited influence to spend, and later tokens crowd out earlier ones.
Attention degradation is the gradual loss of influence that happens as a context grows. Each token competes for a finite share of the model's attention, so as more tokens arrive, the earlier ones get less and less of it.
The effect is visible in behaviour. Instructions given at the start of a long context get followed less reliably than the same instruction given at the end. Details stated hours ago get forgotten. Two parts of a document start to contradict each other because neither held enough attention to override the other.
This is the mechanism behind context rot, and it is why the fix is subtraction rather than addition. Keeping the context small, current, and free of superseded information preserves more of the model's attention for the thing you are actually asking about.
Related terms
Attention
The mechanism that lets each token weigh the relevance of every other token.
Context Window
Everything the model sees on one request. Finite, and it decides what fits.
Context Degradation
The slow loss of quality as a session grows, even though nothing broke.
Compaction
Summarising the old conversation so the new part fits in the window.
Context Rot
What happens when a context keeps everything: it grows, and the useful parts get harder to find.
Want this working inside your team?
It starts with one conversation about your workflow.