Vector-backed conversation memory: messages are embedded and indexed, so search/2 recalls the most relevant past messages (not just the most recent). Wraps an AgentSea.Embeddings handle (any embedder + vector store).

load/1 still returns a conversation's messages in order; search/2 does semantic retrieval across stored memory. A singleton (like AgentSea.Memory.Buffer); start it with an :embeddings handle.