01
a stanford team taught a small model
to keep a diary
02
not a bigger system
not more memory bolted on from outside
just the skill of filing things properly
03
they tested it in nethhack
a 1987 game that kills you for forgetting
04
performance doubled
and the 32b model pulled even
with claude opus 4.5
05
a cohort of well-funded startups
built entire companies selling memory as infrastructure
serena yeung-levy's grad students just trained it away