Vector databases and long-term memory systems for LLMs address how models store, retrieve and maintain contextual knowledge across sessions. This topic covers the Model Context Protocol (MCP) approach to integrating vector search, knowledge graphs, and localized storage so assistants can persist facts, project context, and temporal state without re-prompting. Relevance in 2026 stems from widespread LLM use in multi-session agents, where reliable retrieval, privacy, and operational scale are critical. Current trends favor hybrid architectures (local fast reads with cloud sync), graph+vector hybrid search, temporal awareness, and open protocols for interoperability. These patterns reduce latency, improve relevance for Retrieval-Augmented Generation (RAG), and enable per-workspace context management and access controls. Key tools illustrate common patterns: Qdrant and Chroma provide vector search and embedding-backed document stores for semantic memory layers; cognee-mcp combines graph RAG with customizable ingestion and search; memento-mcp and Neo4j focus on knowledge-graph memory with ontological structure and temporal retrieval; mcp-memory-service targets production needs with hybrid backends and lock-free local reads; context-portal (ConPort) offers structured project context in per-workspace SQLite; Basic Memory emphasizes local-first Markdown knowledge graphs. Practitioners choosing integrations should weigh trade-offs: vector stores excel at fuzzy semantic retrieval and scale, graphs capture relations and temporal reasoning, and local-first or hybrid services improve privacy and responsiveness. MCP-compatible connectors and database/storage integrations ease interoperability across Knowledge Base Connectors, Database Connectors, and Storage Management Integrations, allowing systems to combine vector search, graph reasoning, and operational concerns into durable LLM memory solutions.