Where agentic memory sits in the agent stack ## Where agentic memory sits in the agent stack When Salesforce's Silvio Savarese decomposes an agent into its four working parts, memory is the first one he names — alongside the brain (the reasoning engine), actuators (API and function calls), and the interface. In his telling, memory is typically implemented as RAG: a retrieval layer that stores data and context and feeds the right pieces back to the agent when it needs them. [^50a3b3fa-9ad2-59a3-9c76-ddfe74635066@220] The reason memory matters so much is Savarese's own argument: an agent without connected data and context is useless. Enterprise deployments stumble on siloed data — sales and marketing manage data differently, and the agent cannot operate well when the dots aren't connected. [^50a3b3fa-9ad2-59a3-9c76-ddfe74635066@342] [^50a3b3fa-9ad2-59a3-9c76-ddfe74635066@676] ### Memory as shipped infrastructure Anthropic's managed, cloud-hosted agent platform ships with memory and orchestration built in — evidence that memory has moved from a research idea to a packaged production component. The hosts' contrast is sharp: Anthropic offers memory and orchestration in a scalable hosted platform, while OpenAI, at the time of the conversation, had only an Agents SDK with no cloud version. [^ed15d1a6-8e60-53ca-9dc0-bd0a82a2ddf5@211] ### The usage pattern across the material Across the channel's captured conversations, agentic memory appears in two connected senses: - Architectural sense — one of the four canonical parts of an agent, the component that holds and retrieves context so the brain can reason over it. - Product sense — a feature of managed agent platforms, bundled alongside orchestration so teams don't have to build persistence and retrieval themselves. Both senses share a core claim: agents need to act on your behalf over long, asynchronous tasks, and that only works if something persists and recalls the relevant context between steps. [^ed15d1a6-8e60-53ca-9dc0-bd0a82a2ddf5@113] - In the agent decomposition used in this material, memory is one of four parts: memory, brain, actuators, interface — and it is typically built with RAG. - Savarese's justification: an agent without connected data and context is useless; siloed enterprise data is where deployments fail. - Managed agent platforms (e.g. Anthropic's) ship memory and orchestration as infrastructure, signalling it is a solved-enough problem to be packaged.