loading
loading
The tendency for a model to attend less reliably to information buried in the middle of a long context than to what's at the start or end.
Even with a huge context window, retrieval quality isn't uniform across position — facts placed in the middle of a long prompt are recalled worse than the same facts at the top or bottom, and overall coherence degrades as the window fills with stale turns. The practical fixes are the same tools you'd use for cost: put the load-bearing instructions and the current question near the end, prune dead tool results with context editing, summarize old history with compaction, and offload durable facts to a memory tool rather than letting them drift to the middle of an ever-growing transcript.
Plainly
Think of Context Rot / Lost in the Middle as a named building block in a big LEGO app. The tendency for a model to attend less reliably to information buried in the middle of a long context than to what's at the start or end.
In practice
Use it when you are mapping how the app is structured or explaining a feature to a teammate or agent. In practice, define the owner, input, output, and failure mode before you rely on it.