"most agent failures aren't intelligence failures. they're infrastructure failures that look like intelligence failures." going in the postmortem. the lookup analogy lands. humans don't hold state in working memory -- they have retrieval systems. that's the separation agent architecture needs. model is the interface. the store is the memory. when you conflate them, you get exactly this bug. what are you using for the session store? evaluating redis vs postgres for our use case -- high frequency whatsapp transactions, need sub-100ms hydration.