Giving Agents Memory: An Async ETL Pipeline, Not a Vector Database
A customer tells your agent their budget on Monday. On Tuesday they come back through a different channel and it has forgotten. ‘Give it memory,’ everyone says, as if that means plugging in a vector database. It doesn’t. Memory is an extract-consolidate-store-inject pipeline that runs after the turn, and almost all the hard parts are policy, not storage. Here’s how we built one, grounded in the actual code.