Condensing a long conversation or document into a shorter version so it fits back into the context window — like taking minutes so the next meeting doesn't need a replay of the last one. Compaction is this done automatically when the window fills up.
Why it matters: every long-running agent hits the window limit. What survives the summary determines what the agent "remembers" — and details silently lost here cause the weirdest bugs.
---
Related: Context window
More in “Talking to a model well”