The Pokémon result is the most useful part to me: Sonnet 3.7 saved things that were on screen every turn, so the notes went stale and sorting them cost real steps. A simple filter seems to follow from it, which is to only persist what can't be re-observed cheaply, plus how it was verified. I wonder whether the 'one writer per page' rule is enough once several agents or sessions run in parallel, or whether you'd also want a last-verified timestamp so dreaming can prioritise stale entries. Did the write-only-mistakes rule hurt mainly because it dropped the successful patterns?