[andriy_koval]

For each LLM interaction they likely have bunch of thoughts traces, tool calls, etc, which don't go to context, but still can be retrieved.

But I obviously don't know for sure.