Why AI Memory Systems Need Curation Policies, Not Just Vector Databases

A technical deep-dive into AI memory architecture argues that vector databases are only one component of durable memory, not its complete definition. The author distinguishes between Active Working Memory — assembled per task — and Durable Memory, which stores only what a system intentionally decides is worth preserving. Key items worth retaining include specifications, user preferences, and architecture decision records, while scratch calculations and ephemeral context should be discarded. The piece highlights that most AI teams over-invest in retrieval strategies while neglecting the equally critical write-side question of what should be stored at all. Systems that indiscriminately retain everything risk what the author calls a 'Digital Attic' — a bloated memory store where nothing can be reliably found.