LLMs forget. This MCP server gives them a Qdrant database.

What it does This official server acts as a semantic memory layer on top of the Qdrant vector database. It offers two tools—qdrant-store and qdrant-find—that let an LLM client save arbitrary text (with optional metadata) and later recall it via vector similarity search. Embedding generation is handled internally using FastEmbed models, so callers just pass plain text.

The interesting bit The repository doubles as a reference implementation for building MCP servers on Qdrant. A neat secondary use case is repurposing the same two tools as a code-search backend for Cursor or Windsurf by rewriting the tool descriptions to talk about snippets instead of memories.

Key highlights

• Supports both local Qdrant instances and remote servers authenticated via API key
• Can lock the server to read-only mode, disabling the store tool entirely
• Speaks stdio, SSE, or streamable HTTP depending on how the client connects
• Automatically creates the target collection if it does not already exist
• Currently supports only FastEmbed models for embeddings

Caveats

• Embedding provider support is limited to FastEmbed; bringing your own model or provider is not yet an option
• FastMCP server settings rely on a FASTMCP_SERVER_ prefix that the README warns may change in future versions
• Repurposing it as a code search tool for Cursor or Windsurf requires manually customizing tool descriptions and may need additional Cursor rules to behave reliably

Verdict Useful if you are building an MCP client and need persistent, semantically searchable context without managing vectorization yourself. Less compelling if you require pluggable embedding providers or want a standalone search API without the MCP protocol overhead.