Comment by camelliaPTM

The prompt went from 44k to 6k tokens, but you're making two extra model calls per round to get there (chunker + working_memory_update). What does the all-in cost comparison actually look like?