Context Rot
Context ManagementDefinition
Attention degradation caused by a context window filling with irrelevant, stale, or low-signal content, even when technically there is still room left in the window. Distinct from running out of space (a hard context-window limit) and from position effects (attention bias by location within the window) -- context rot is specifically about signal-to-noise degrading as low-value tokens accumulate.
Example Usage
An agent session that runs for hours accumulates verbose tool output and completed side-quests; even though the context window isn't full, quality drops because low-signal history is crowding out what the model needs to attend to -- the fix is pruning and compaction, not a bigger window.
In Depth
Not the Same as Running Out of Room
Context rot is often misdiagnosed as a capacity problem, but the two are distinct: running out of room means the context window is literally full and the next token can't be added. Context rot happens *before* that point -- the window still has technical space left, but attention degrades because it is dominated by irrelevant, stale, or low-signal content the model has to sift through to find what actually matters.
Context Drift: The Symptom Over Time
A closely related phenomenon, context drift, is the gradual loss of track of instructions or facts over a long interaction. The telltale symptom is a long agent session that starts ignoring earlier instructions or degrading in quality, even though nothing about the underlying task changed. Context rot is the mechanism; context drift is what it looks like from the outside over the course of a session.
Distinguishing Rot from Position Effects
Context rot is related to but distinct from the lost-in-the-middle effect: position effects describe *where* in the window content gets less attention (the middle, versus the start and end); context rot describes *how much* of the window's attention budget is wasted on low-value content regardless of position. A window can suffer from both simultaneously, but the fixes differ -- placement fixes position effects, while pruning and compaction fix rot.
Countering It
Compaction, pruning, isolation (delegating heavy work to a subagent's own context window), and re-stating key constraints all counter context rot and drift by keeping the active window's signal-to-noise ratio high. None of these is a one-time fix; they are continuously applied techniques for any long-running session.
Common Pitfalls
- Reaching for a bigger context window as the fix, when the actual problem is signal-to-noise, not raw capacity -- a bigger window filled with the same low-signal content still rots.
- Responding to a session that's started ignoring earlier instructions by raising temperature or adding more prohibitions to the prompt -- neither addresses a crowded, low-signal context window.
- Assuming quality degradation late in a long session is a model limitation rather than checking whether the context has simply rotted.