brody
6.9K posts









guys in brooklyn like ok yeah lemme just grab my tiny shirt and enormous pants








some f/u thoughts: the active problem is that as agents start taking real actions (eg agentic robinhood — trading on your behalf, managing accounts, paying), the agent's memory of you basically *is the agent. risk tolerance, prior decisions, what worked, what didnt work last week, etc I think there’s three problems today: 1. trapped — your robinhood agent's memory of you lives in robinhood, your claude memory lives in anthropic. switching = blank slate. lock-in moat for the labs 2. unverifiable — when an agent moves real money "based on what it knows about you," you can't audit what memory it actually used or whether it got tampered with. memory poisoning is already a real attack vector 3. not yours — same data, same value accrual, sits on the lab's balance sheet the bet is agent memory becomes a portable verifiable user-owned primitive that sits *outside the lab, gets read by whichever model/agent you point at it. matters most when agents start touching things that have real consequences (money, contracts, accounts, etc) this isn't about helping the labs with their HBM problem, it's about owning the memory layer they currently own by default




















