From the source
Laminar implemented trace deduplication: instead of storing every LLM call's full message history (which grows quadratically with turn count), the system hashes each message and stores unique messages only once per trace, storing a compact array of hashes in spans.
At query time, the original input array is reconstructed byte-identically via the SQL engine.
On realistic long-running agents the approach reduces data from 5× to 10×, and on a 50-turn coding agent with stable system prompt and long tool outputs it approaches 50×.
