Retrieval finds relevant stored evidence; context assembly chooses what fits the next model call. RAG can do this for agent memories. An optional service that generates an answer from memory, such as Hindsight reflect, is a different operation from recall.

Reading memory: retrieval, context selection, and optional synthesis

LangChain's retrieval overview describes fetching external knowledge at query time and integrating it with generation. The source can be a document collection, structured data, or an agent's previous experience. Memory and RAG are therefore overlapping concepts, not opposing architectures. The RAG investigation concerns the evidence-access side of this boundary.

A useful memory query can constrain subject, project, and time as well as textual relevance. Hindsight recall is one documented implementation combining semantic, keyword, graph, and temporal retrieval. That is a capability example, not proof of better retrieval on our workload.

Context assembly then decides which retrieved records, standing facts, and recent messages the next model call receives. Anthropic's context-engineering guide discusses compacting history and consulting durable notes rather than retaining everything in active context. Compression can lose details, so I would preserve a path back to source evidence.

Finally, Hindsight reflect runs an LLM-driven search and synthesis loop, whereas recall returns records. A caller asking for evidence need not delegate answer generation to the memory service. The quality and cost of searching should be distinguished from those of generating a new interpretation.