Memory MCP servers

Long-term recall and knowledge graphs.

Memory MCP servers give an assistant something it does not have by default: continuity. A conversation ends and its context is gone. A memory server writes facts to durable storage so the next session can start from what the last one established.

The common implementation is a knowledge graph — entities, relations and observations rather than a transcript. Storing 'this project uses Postgres' as a fact about a named entity means it can be retrieved later by a question that shares none of the original wording.

Most run locally and store to a file you own. That makes this one of the easiest categories to try: no API key, no service, no account, and you can read the storage yourself to see exactly what has been remembered.

Memory MCP servers in this directory

What people build with memory MCP servers

Not re-explaining your project every session

Stack, conventions, deployment target, the reason for that odd workaround — stated once and available afterwards.

Remembering preferences

How you like code formatted, which libraries you avoid, what you want the agent to stop suggesting.

Carrying context across long work

Multi-day work survives the session boundary instead of restarting from a summary you have to write.

Building a personal knowledge graph

Entities and relations accumulate into something queryable, rather than a pile of notes.

Choosing a memory MCP server

  • Check where storage lives. A local file is auditable and portable; a hosted service is neither, though it does follow you between machines.
  • Prefer a readable storage format. Being able to open the file and see what was stored is what keeps a memory server trustworthy.
  • Look for a delete tool, and use it. Memory that can only grow eventually fills with things that used to be true.
  • Check how retrieval works. A server that only returns exact matches will not find what you meant.

Before you connect one

A memory server stores whatever it is told, so assume anything mentioned in conversation could end up on disk. Point it at a path you control, look at what it has written occasionally, and delete facts that have gone stale — a confidently remembered out-of-date fact is worse than no memory at all.

Frequently asked questions

Where is the memory stored?

For most servers in this category, in a local file whose path you configure. Nothing is uploaded, and you can open it and read exactly what has been kept.

Does memory carry across different AI clients?

Yes, if you point them at the same storage. That is one of the better arguments for a memory server over a client's own built-in memory feature.

How do I make it forget something?

Ask the agent to delete the entity or observation, if the server exposes a delete tool — or edit the storage file directly, which is always available with a local server.

Is this the same as a vector database?

No. A vector store retrieves passages by embedding similarity. A knowledge-graph memory server stores discrete facts and relations. They solve different problems and are often used together.

Install any of these

Every server on this page carries a one-line install command on its own listing. Paste it into your client config and restart — the memory tools appear in your next session.

Or keep browsing: all MCP server categories · the full MCP server list · Agent Skills.