A multi-agent investigative workflow needs turn-by-turn dialogue context, a shared case file that several agents update during a session, and durable knowledge of prior cases for semantic recall across weeks. Tenant isolation and lifecycle policies are mandatory. Which memory architecture should you specify?
Select an answer to reveal the explanation.
Short Explanation
You need a multi-tier design: short-term session context, shared team/case state, and long-term semantic memory with retention, tenant isolation, and lifecycle policies. That gives agents turn-by-turn context, a shared case file, and durable recall without leaking data across tenants.
Full Explanation
The exam calls out designing short-term and long-term memory architectures with context sharing, and specifying multi-tier state persistence for session state, shared team state, and long-term semantic memory with lifecycle and tenant-isolation policies. Map technologies accordingly, for example in-memory or Redis for session, durable database for case files, vector store for semantic long-term memory, without treating the language model context as a database. Context-window-only storage fails length limits and cross-session needs. Single public blob container storage without tenant prefixes or time-to-live violates tenancy and retention. Disabling memory so every agent turn starts with zero context sacrifices solution goals and breaks multi-step collaboration. Include compaction strategies, what is shared versus private per agent, and encryption or role-based access control on each store.