WhyVault
Sign indeStart for free

MCP memory server

A memory MCP server: one memory for every AI tool.

Every chat with an AI starts from zero, unless something remembers for it. A memory MCP server is that something, and because it speaks MCP, it works with Claude, ChatGPT, Cursor and more. Here’s how the different kinds work, local vs hosted, and the checklist to use before you trust one with what you know.

Reading time: 10 minutesUpdated:

01

What a memory MCP server is

Language models don’t remember anything between chats on their own. Whatever you explained yesterday is gone today, unless it’s stored somewhere the model can reach. Built-in memory features do that inside one app. A memory MCP server does it as a separate service that any assistant can plug into.

MCP, the Model Context Protocol, is the shared plug: the server offers tools such as “save a note”, “search” or “read”, and the assistant decides when to use them. You tell Claude “remember that we chose weekly releases”, and it calls the server’s write tool. A week later, in Cursor, you ask why releases are weekly, and Cursor calls the search tool on the same server.

That’s the real point: one memory, many assistants. Instead of Claude knowing one thing and ChatGPT another, they all read and write the same store.

02

Three ways memory servers store what they know

Under the hood, memory servers differ mostly in how they keep things. That decides whether you can read and fix your memory yourself.

  • Knowledge graph

    People, things and the facts about them, stored as entities and relations, often in a local JSON file. The reference memory server from the MCP project works like this. Precise for facts, hard for people to read.

  • Vector store

    Snippets saved as embeddings and recalled by similarity. Scales to huge amounts of text, but you can’t easily see what’s stored or correct a wrong memory.

  • Markdown notes

    Plain files with headings and [[links]]. Readable and editable by you, portable to any notes app. Local tools like Basic Memory and hosted ones like WhyVault work this way.

03

Local or remote (hosted)?

MCP servers run in two ways. Which one fits depends on where you use AI.

–Local (on your computer)Remote (hosted, over a URL)
How it runsA process your client starts, via npx, uvx or DockerA server on the internet, reached over HTTPS
SetupNode.js or Python, plus a JSON config in every clientPaste a URL, sign in once
Works inDesktop clients on that machine: Claude Desktop, Claude Code, CursorThose too, plus Claude on the web and phone, and ChatGPT
Sign-inUsually none, or an API key in a config fileIdeally OAuth; an access key as an alternative
Shared across toolsOnly if every client points to the same fileYes, one memory for every connected assistant
Your dataStays on your diskWith the provider: check where and under which contract
Backups and historyYour jobDepends on the provider – ask before you commit
CostMostly free and open sourceOften a subscription, sometimes a free tier
How it runs
Local (on your computer)A process your client starts, via npx, uvx or Docker
Remote (hosted, over a URL)A server on the internet, reached over HTTPS
Setup
Local (on your computer)Node.js or Python, plus a JSON config in every client
Remote (hosted, over a URL)Paste a URL, sign in once
Works in
Local (on your computer)Desktop clients on that machine: Claude Desktop, Claude Code, Cursor
Remote (hosted, over a URL)Those too, plus Claude on the web and phone, and ChatGPT
Sign-in
Local (on your computer)Usually none, or an API key in a config file
Remote (hosted, over a URL)Ideally OAuth; an access key as an alternative
Shared across tools
Local (on your computer)Only if every client points to the same file
Remote (hosted, over a URL)Yes, one memory for every connected assistant
Your data
Local (on your computer)Stays on your disk
Remote (hosted, over a URL)With the provider: check where and under which contract
Backups and history
Local (on your computer)Your job
Remote (hosted, over a URL)Depends on the provider – ask before you commit
Cost
Local (on your computer)Mostly free and open source
Remote (hosted, over a URL)Often a subscription, sometimes a free tier

Local is great for privacy and tinkering on one machine. Remote is what lets memory follow you onto your phone and into ChatGPT. And then the choice of provider matters, which is what the checklist below is for.

04

What to look for in a memory MCP server

Your memory ends up holding decisions, projects and personal context. Check these before you connect an agent:

  • OAuth sign-in instead of keys pasted into config files, and access you can revoke per connection.
  • Permissions per connection: read-only for tools that only look things up, and deleting off by default.
  • Versioning: every change saved as a version, showing which agent made it, with a way to undo.
  • Careful writes: targeted edits, such as appending or replacing one section, instead of overwriting whole documents.
  • Export in an open format: Markdown or JSON you can take anywhere, even after you cancel.
  • Readable for you: a memory you can browse and fix yourself, not a black box only the model sees.
  • Where it’s hosted, and the contract: data location, a DPA for business use, no training on your data. In Europe: EU hosting and GDPR.
  • Rules for agents: a place for instructions every agent reads first, so the memory stays tidy.
  • A brake: limits that stop a runaway agent before it rewrites everything.

05

How WhyVault handles it

WhyVault is a hosted memory MCP server for plain Markdown notes. The checklist, point by point:

  • OAuth or access key

    Claude and ChatGPT sign in with OAuth 2.1. Claude Code and Cursor can use an access key instead. You revoke any connection under Connect agents.

  • Read, write, delete

    Set per connection. New connections read and write. Deleting is off until you allow it, and even then notes only go to the trash.

  • Every change a version

    With the agent’s name and a diff. Undo one change in a click, or everything one agent did since a point in time. History and control

  • Plain Markdown out

    Folders, frontmatter and [[wikilinks]], exported as a zip that opens as an Obsidian vault. Always possible, even without a subscription.

  • EU-hosted, GDPR-compliant

    Database, sign-in and files in Frankfurt, with a DPA for business use. Your notes aren’t used for training or ads.

  • Rules every agent reads

    Notes in 9_System come with vault_overview at the start of every session: who you are, how notes should be written, what goes where.

06

The tools an agent gets

Ten tools, grouped by permission. Full reference

  • Read

    vault_overview, search_notes (full text), read_note, list_notes and get_backlinks. Answers always include path, title and version.

  • Write

    create_note, append_to_note, update_section and replace_note. Every write carries a one-line summary for your history.

  • Delete, if you allow it

    delete_note moves a note to the trash, only with the version the agent read. Nothing disappears for good over MCP.

07

Set up WhyVault as your memory server

One address for every client: https://mcp.whyvault.io.

  1. 1

    Create your vault

    Sign up with your email and pick a template. It comes with the folders and the 9_System rules agents read first. Quick start

  2. 2

    Claude: web, desktop and mobile

    Under Customize → Connectors, add a custom connector with the URL above, then sign in and approve. Works on every Claude plan, including Free. Connect Claude

  3. 3

    Claude Code

    One command, then sign in with /mcp. Or add an access key as a header. Claude Code setup

    claude mcp add --transport http --scope user whyvault https://mcp.whyvault.io
  4. 4

    Cursor and other clients

    An entry in ~/.cursor/mcp.json. VS Code, Gemini CLI and any client with remote MCP work the same way. Cursor · Other clients

    {
  "mcpServers": {
    "whyvault": {
      "url": "https://mcp.whyvault.io"
    }
  }
}
  5. 5

    ChatGPT

    In developer mode on chatgpt.com (Plus and up), add WhyVault as an app with OAuth. Connect ChatGPT

  6. 6

    Your first prompt

    Ask your agent to read the overview and save something. Then open the app and see the note, marked with the agent’s name.

    Read my vault overview. Then save today’s most important decision as a note in the right room: what we decided, why, and what we rejected.

08

What it feels like

In Claude Code, in the middle of a refactor:

Claude Code

  1. You: Remember this: we’re sticking with Postgres full-text search, no vector database. Too much to run for the gain.

  2. Tool: create_note · “2026-10-12 Full-text search instead of vectors”

    Claude Code: Saved as a decision note in your project room, with the reason and the rejected option.

  3. You: What did we decide last week about rate limits?

  4. Tool: search_notes · “rate limit”

    Claude Code: “2026-10-05 Rate limits per connection”, written from Cursor: a brake per connection instead of a global one, so one runaway agent can’t block the others.

Written in Cursor last week, found in Claude Code today.

If you can’t read your AI’s memory, you can’t correct it.

?

Frequently asked questions

What is an MCP memory server?

A service that gives AI assistants tools to save and recall information across chats, using the Model Context Protocol. Because MCP is an open standard, Claude, ChatGPT, Cursor and other clients can all use the same memory.

Is the official memory server enough?

For experiments, yes. The reference memory server from the MCP project keeps a knowledge graph in a local file. It has no sign-in, no history and doesn’t reach Claude on the web or ChatGPT, so it’s fine on one machine and limited beyond that.

Local or hosted: which is safer?

Local keeps everything on your disk, but you handle backups, and anything that can run on your machine can read it. Hosted means trusting a provider, so check where it hosts, whether there’s a DPA, and whether you get history and export. Either way, the AI provider sees what the agent reads.

Doesn’t Claude have memory built in?

It does, and for many people that’s enough. Claude’s memory only works inside Claude, though. A memory server is shared across every tool you use. Claude memory compared

Can I self-host WhyVault?

No, WhyVault runs as a hosted service. If self-hosting is a must, a local memory server is the better fit. Your WhyVault notes stay plain Markdown, so you can export them anytime and take them along.

Is WhyVault GDPR-compliant?

Yes. WhyVault is hosted in the EU, in Frankfurt, and businesses get a DPA. For the agent side, check your AI provider’s terms too: what an agent reads, its provider processes. History, export and privacy

How much does it cost?

WhyVault is free for up to 20 notes, no credit card needed. Pro costs €9.99 a month or €99 a year and has no limit. Reading and exporting always work, even without a subscription.

All guides

Memory with sign-in, history and export.

Create your vault and connect your first agent in five minutes. Free, no credit card.

Start for free