RAG decouples what a model can talk about from what it was trained on: instead of hoping a fact is stored in the weights, the system looks it up and shows it to the model at inference time. The standard pipeline embeds documents into vectors, retrieves the nearest neighbors of the query embedding, and prepends them to the prompt.
The same mechanism generalized beyond documents becomes agent memory: retrieval over an agent’s own past experiences, notes, and skills, typically scored by relevance plus recency and importance rather than similarity alone. Retrieval quality is the usual bottleneck in both cases; a store only helps if the right entry comes back at the right moment.