All topics
AI memory

What Is an MCP Memory Server?

An MCP server that gives AI assistants read (and sometimes write) access to a persistent knowledge base.

By Ilian Raykov, founder of BrainTube

Updated October 2026

An MCP memory server is a Model Context Protocol server that gives AI assistants read (and sometimes write) access to a persistent knowledge base — so any MCP-compatible AI, like Claude or Cursor, can query your notes, documents and saved content directly.

What MCP is, in one line

MCP (Model Context Protocol) is the open standard that lets any AI client connect to any tool or data source through a single uniform interface. See our full explainer, What is MCP (Model Context Protocol)?. An MCP memory server is one specific use of the protocol: exposing a persistent knowledge base as something the AI can browse and query.

Why it beats copy-pasting context

Without a memory server, giving an AI context means pasting the same doc, transcript or note into every new chat — every time, in every app. An MCP memory server flips that: connect once, and every future conversation in every MCP-capable client can pull from the same store on demand. No paste, no re-upload, no drift between what you told Claude and what you told Cursor.

What you can connect

Notes, documents, saved videos and podcasts (with transcripts), articles, PDFs, EPUBs — anything the server has indexed. A good MCP memory server exposes both resources (browseable items) and tools (query endpoints), returns citations back to the original source, and handles auth cleanly so you're not pasting API keys into every client.

Two kinds of MCP memory server

Servers sold as MCP memory do one of two jobs, and they are easy to confuse:

KindWhat it storesWho writes to itExample
Fact storeShort facts about you and your work, kept as entities, relations and observationsThe assistant, during chatsThe official reference server, @modelcontextprotocol/server-memory
LibraryThe sources you saved: transcripts, articles, PDFs, notesYou, when you save; some servers let the assistant save tooBrainTube

The reference server in the official modelcontextprotocol/servers repository describes itself as "a basic implementation of persistent memory using a local knowledge graph". It keeps a JSONL file on your machine and gives the assistant tools such as create_entities, add_observations and search_nodes. That is enough for an assistant to remember that you work at a company or prefer TypeScript. It is not built to hold a two-hour podcast or a 300-page PDF.

A library server holds the material itself and answers questions from it with citations back to the source. If what you want remembered is what you watched and read, you need the second kind.

Read-only vs read-write

Some MCP memory servers only let the assistant search. Others also let it save: add a source, write a note, tag an item. Read-only is safer to try; read-write means the knowledge an agent finds while it works lands in the same store you search later, instead of disappearing when the chat ends. Check the server's tool list before you connect it: tools named like search or read are read paths, tools named like add, ingest or create are write paths. More on the trade-off in read-only vs read-write MCP servers.

How to choose one

  • What should it remember? Facts about you: a fact store. The content you learn from: a library.
  • Where does it run? Local servers keep data on your disk and work in one desktop client at a time. Remote servers (an HTTPS endpoint with OAuth) can work in web clients such as Claude.ai and across devices.
  • Can the assistant write? Decide whether you want agents saving into it.
  • Can you leave? Check that the data exports in an open format.

The MCP documentation calls the protocol "a USB-C port for AI applications": the point of picking an MCP server over an in-app memory feature is that the same store keeps working when you change assistants. A side-by-side of six servers is in how to give Claude long-term memory with MCP.

BrainTube as your MCP memory server

BrainTube runs a managed MCP memory server over your compiled knowledge base — everything you've saved from YouTube, podcasts, articles and documents, transcribed and indexed. One-click connect to Claude.ai and Claude Desktop as a custom connector via OAuth, or drop a config into Cursor. The same store answers every AI you use, with cited-to-source recall you can verify.

Frequently asked

How is an MCP memory server different from a plugin?
Plugins are per-app: a ChatGPT plugin only works in ChatGPT, a Claude tool only works in Claude. MCP is a shared protocol — one server, many clients. Build (or connect) it once and every MCP-capable AI can use it.
Can Claude and Cursor use the same MCP memory server?
Yes. That's the whole point of MCP. Add the server URL to Claude Desktop as a custom connector and to Cursor with an API key, and both get the same view of your knowledge base.
What is the official MCP memory server?
The MCP project's reference server, published on npm as @modelcontextprotocol/server-memory. It stores entities, relations and observations in a local knowledge graph file so an assistant can remember facts about you across chats. It is a fact store, not a library for documents or transcripts.
Can an MCP memory server save things, or only search?
Both exist. Read-only servers expose search and read tools; read-write servers also let the assistant add sources, notes or tags. BrainTube's server is read-write on every plan: your AI can save what it finds, add notes and tag items, as well as search.
What are the best MCP memory servers?
It depends on the job. For facts about you inside one desktop client, the official reference memory server is the simplest start. For a library of what you watch and read that several assistants share, use a hosted library server such as BrainTube. A comparison of six servers is in the MCP setup guide at brain-tube.com/guide/give-your-ai-memory-mcp.
Is my data private?
With BrainTube, you own your corpus — the knowledge base is yours, exportable as Markdown or Anki cards from Settings or in full through MCP, and only the AI client you explicitly connect gets access. See our data pledge for details.

Try BrainTube on your own corpus

Free tier, no card. Export anytime.

Start free

Video notes on this topic

All video notes

More to read