Compaction
Replacing accumulated conversation state with a smaller representation so work can continue within context limits.
Illustrative conversation
- Why did the agent forget an edge case after a very long session?
- Compaction preserved the main plan but may have omitted that detail, so the agent needed to reload the source.
Definition
Compaction compresses earlier conversation state into a smaller representation that can be carried into later requests. It extends a long-running workflow by reducing the token footprint of its history, but compression can discard details that later become important. A representation may be a readable summary or provider-specific opaque state, as in OpenAI’s compaction API. Preserve exact constraints in source artifacts; do not assume a compressed item is an editable transcript.
Reviewed