During each research round, structured JSON extractions normalize each discovered statement into the grounded_memory_facts table with list-position confidence scores while attaching their cited URLs to grounded_memory_sources.