The video explains how modern language models manage external context through retrieval and update mechanisms. Retrieval involves embedding the current query, searching a vector store, and pulling the most relevant documents to augment the model's prompt. Various indexing structures—such as flat, IVF, and HNSW—balance speed and accuracy, while hybrid approaches combine lexical and semantic scores. Update mechanisms keep the knowledge base current: static updates replace entire indexes, while incremental updates add or delete vectors without rebuilding. Techniques like cache‑based short‑term memory, recurrent attention, and differentiable neural memory allow models to incorporate newly retrieved information across turns. The speaker compares lazy retrieval (fetch on demand) with eager retrieval (pre‑fetching) and discusses trade‑offs in latency, token budget, and relevance. They also cover consistency checks, grounding, and the role of metadata for filtering. Finally, best practices include periodic re‑embedding, versioning, and monitoring retrieval quality to prevent drift, ensuring the model remains both knowledgeable and adaptable.