Local vs hosted memory MCP server: which one to choose?

Updated 30 October 2026. Facts about MCP clients come from their official documentation, and facts about each project from its own repository or pricing page, all read on 30 October 2026.

A local memory MCP server runs on your own computer: the AI app starts it as a small program and your memories end up in a file on your disk. A hosted one is a web address the app signs in to. Choose local if the memory must stay on your machine and you work in desktop apps such as Claude Desktop, Claude Code or Cursor. Choose hosted if you also want it in claude.ai in the browser, in the Claude mobile app or in ChatGPT, which connect to a server by its address, not to a program on your computer.

What a local MCP server is

The Model Context Protocol defines two standard transports. With stdio, the client launches the server as a subprocess on the same machine and talks to it through standard input and output; nothing goes over the network between them. With Streamable HTTP, the server is a separate process at a URL, usually protected by OAuth. “Local” almost always means stdio.

In practice, you add a local server in one of three ways:

• Claude Desktop: a command declared in its configuration file, claude_desktop_config.json, or a desktop extension: a single .mcpb file that bundles the server and what it needs, installed from Settings → Desktop app → Extensions (Settings → Extensions in the older layout).
• Claude Code: `claude mcp add` followed by the command that starts the server.
• Cursor: a `command` entry in ~/.cursor/mcp.json (or .cursor/mcp.json in a project).

Sources: local servers and desktop extensions in Claude Desktop, Claude Code, Cursor.

What you gain with a local server

• The memory file is yours. It sits on your disk, in a format you can open, copy and back up. No account, no third-party operator for the storage.
• It needs no outside server. Reading and writing the memory needs no network. The assistant itself is another matter: Claude, ChatGPT or Cursor's models still run in their vendor's cloud.
• It is usually free and open source. You can read the code, change it and pin a version.

One point people miss: local storage does not mean the AI vendor never sees the content. Whatever the assistant reads from your memory to answer you becomes part of the conversation, which is processed by the vendor like any other message.

What a local server cannot do

• Be reached from the web or a phone. Anthropic says it plainly: desktop extensions run locally and are only available in Claude Desktop and Claude Code, not on the web or mobile. ChatGPT connects to custom MCP servers by URL, over SSE or streaming HTTP, not stdio. Perplexity runs local servers only in its Mac app, through a helper app.
• Follow you to another device. The memory lives on one computer. A second laptop starts empty unless you sync the files yourself.
• Look after itself. Updates, the Node.js or Python runtime it needs, backups and the occasional broken dependency are your job.
• Stay sandboxed. A stdio server runs with your user's permissions. Install only code you trust.

Sources: Anthropic, desktop and web connectors, OpenAI, custom MCP servers, Perplexity, local and remote MCPs.

Local memory servers you can use today

Each fact below was read on the project's own page on 30 October 2026.

• The reference “memory” server (MIT, per its README), from the official MCP servers repository: a knowledge graph of entities, relations and observations stored in a JSONL file, started with `npx -y @modelcontextprotocol/server-memory`. Its search matches your query against names, types and observation text; there is no semantic search. The best way to see what an MCP memory is in ten minutes.
• Basic Memory (AGPL-3.0): your memory as plain Markdown files in a folder, readable in any editor. Free to run locally; its cloud, from $19 a month per workspace, adds sync and remote access.
• mcp-memory-service (Apache 2.0): a self-hosted service with SQLite storage for a single user on one machine, which can also be served over HTTPS with OAuth 2.0 if you want to expose it yourself.
• claude-mem (Apache 2.0): built to give Claude Code memory between sessions, stored in a local SQLite database with full-text search; cloud sync is optional.

Sources: reference server, Basic Memory, mcp-memory-service, claude-mem. One warning: OpenMemory, Mem0's local MCP server that many tutorials still recommend, was removed from the Mem0 repository on 29 July 2026; check what that name points to before building on it. The wider landscape, hosted options included, is in our comparison of memory MCP servers.

The middle way: self-hosting

Some of these projects can run on a server you rent and be exposed over HTTPS with OAuth. You then get remote access from any client while keeping control of the machine. The price is that you become the operator: TLS certificates, OAuth configuration, updates, backups and uptime.

When a hosted memory like Pryzm makes sense

Pryzm Memory is a hosted MCP memory: one address (https://pryzm-memory.com/mcp), an OAuth sign-in, nothing to install. It keeps what you tell your assistant, files it by project, finds it again by meaning rather than exact words (keyword and semantic search, re-ranked by a multilingual model), and when a decision changes, it marks the old one as replaced so the assistant works from the one in force.

Because it is an internet address, the same memory also answers in claude.ai, the Claude mobile app, ChatGPT, Perplexity, Cursor and Claude Code. That is a bonus of being hosted, not the reason to choose a memory.

The trade-off is the one described above, reversed: someone else runs it. Pryzm runs on Hetzner servers in Falkenstein, Germany. Each account has its own encryption key; while you are signed in, the server decrypts your memories in memory to answer your assistant. The free plan holds 100 memories in one project; the Solo plan is €9.99 a month. Where that lives and who can access it is detailed in where does AI memory live.

Pryzm is the wrong choice if your memory must never leave your computer, if you want to read or modify the code, or if you only ever use one desktop app: a local server will serve you better.

Decide in thirty seconds

• Only Claude Desktop, Claude Code or Cursor, on one computer, and the data must stay there → a local server.
• You want to open the memory files yourself, in Markdown → Basic Memory, locally.
• You want control and remote access, and can run a server → self-host a project that supports HTTP and OAuth.
• You also use claude.ai on the web, your phone or ChatGPT, and don't want to run anything → a hosted memory.