Context Degradation
Degradasi Konteks
The slow loss of quality as a session grows, even though nothing broke.
Context degradation is the gradual decline in instruction-following and accuracy as a session gets longer. No error appears, no tool fails, but the model starts ignoring earlier instructions, losing track of decisions it made, and contradicting itself.
The mechanism is competition for attention. The context window is a fixed budget, and every token in it spends some of the model's capacity. When the window fills with old files, long tool outputs, and repeated summaries, the current instruction is one voice among many.
The practical response is to keep the context small and current. Load only what the step needs, summarise what has already been decided, and start a fresh session when the topic changes. Compaction, the harness doing that summarisation automatically, is the mechanism most agents now use.
Related terms
Context Window
Everything the model sees on one request. Finite, and it decides what fits.
Compaction
Summarising the old conversation so the new part fits in the window.
Attention
The mechanism that lets each token weigh the relevance of every other token.
Context Engineering
Deciding what goes into the model's context, and what stays out.
Memory
A system's ability to carry information from one request to the next.
Want this working inside your team?
It starts with one conversation about your workflow.