Reuse similar answers to cut LLM cost and latency.
Recent revisions are kept. Restoring one creates a new revision.
