Context & memory infrastructure for multi-turn AI focuses on persistent, queryable stores that let agents remember, retrieve and update conversational state across sessions. With the Model Context Protocol (MCP) emerging as a practical interoperability layer, teams are separating “memory” from models: dedicated memory servers provide semantic search, structured relations and transaction-safe storage so assistants can maintain long-term personalization, task state, and provenance. Contemporary implementations follow a few clear patterns. Vector search engines (Chroma, Qdrant) provide dense-embedding retrieval and full-text/document storage for semantic lookup; graph-enhanced approaches (cognee-mcp) combine relation-aware graph stores with vector search for richer RAG-style reasoning. Production MCP services (mcp-memory-service) emphasize hybrid, local-first architectures—fast local reads (SQLite or embedded stores) plus cloud synchronization—avoiding DB locks and reducing latency. Domain-specific connectors (obsidian-mcp) expose personal knowledge bases and note systems as MCP-accessible memories so assistants can read, write, and organize real user content. This topic is timely in 2026 because widespread multi-turn assistants and embedded agent workflows demand scalable, privacy-conscious memory that survives model updates and token limits. Key trade-offs include latency vs. durability, semantic recall vs. schema precision, and local-first privacy vs. centralized analytics. For practitioners evaluating options, the relevant categories are Knowledge Base Connectors (connectors to Obsidian, docs, wikis) and Storage Management Integrations (vector stores, graph layers, hybrid sync engines). Understanding MCP-style protocols and semantic memory servers helps teams choose architectures that balance personalization, consistency, and operational reliability for long-lived AI assistants.