We’re, in impact, standing up a second knowledge stack particularly for brokers, then questioning why nobody in safety feels comfy letting these brokers close to something essential. We shouldn’t be doing this. In case your brokers are going to carry recollections that have an effect on actual choices, that reminiscence belongs inside the identical governed-data infrastructure that already handles your buyer information, HR knowledge, and financials. Brokers are new. The way in which to safe them just isn’t.
Revenge of the incumbents
The trade is slowly waking as much as the truth that “agent reminiscence” is only a rebrand of “persistence.” In case you squint, what the massive cloud suppliers are doing already seems like database design. Amazon’s Bedrock AgentCore, for instance, introduces a “reminiscence useful resource” as a logical container. It explicitly defines retention durations, safety boundaries, and the way uncooked interactions are reworked into sturdy insights. That’s database language, even when it comes wrapped in AI branding.
It makes little sense to deal with vector embeddings as some distinct, separate class of knowledge that sits exterior your core database. What’s the purpose in case your core transactional engine can deal with vector search, JSON, and graph queries natively? By converging reminiscence into the database that already holds your buyer information, you inherit a long time of safety hardening without spending a dime. As Brij Pandey notes, databases have been on the middle of software structure for years, and agentic AI doesn’t change that gravity—it reinforces it.
