Should I use memory or knowledge?

Use memory for the conversation the agent should remember across runs. Use knowledge for source material the agent should retrieve because it is too large, dynamic, or selective to include in every prompt.

Question	Capability
“What did this user and assistant say earlier?”	Memory
“Which handbook section answers this question?”	Knowledge
“What is this account's current balance?”	An authorized tool, not a vector store
“What small policy should every run receive?”	Static agent context

Can an agent use both?

Yes. A run can load prior messages from memory and retrieve relevant documents from knowledge. Keep their lifecycles separate: conversation retention and deletion often differ from corpus ingestion and re-indexing.

Does memory scope enforce access?

No. A scope key selects stored history. The application must authenticate the request and verify that the caller may access that session. Choose an adapter in Memory store adapters.

Do knowledge filters enforce authorization?

Not by themselves. Metadata filters narrow retrieval, but application authorization must determine which tenant, visibility, and corpus constraints are allowed. See Metadata filters.

Knowledge quality also depends on document loading, chunking, embeddings, vector-store configuration, and the selected model. Test retrieval and answer grounding together rather than treating the vector adapter as a complete knowledge system.