A memory MCP server gives your AI app a memory that lasts beyond a single conversation. Claude, ChatGPT or Cursor connect via MCP and get tools like remember this and search what you know. That sounds the same with every provider. Underneath sit very different things: a JSON file on your machine, a folder of Markdown notes, a vector database in the US, or a knowledge graph built from your own documents.
Here five memory servers stand side by side: the three that show up most often in search results and AI answers, the MCP project’s reference server, and Memexa. Memexa is our own product. The other four we filled in from their own pages and repositories, with sources at the end.
The comparison
| Server | Runs | How it stores | Documents | Open source | Price |
|---|---|---|---|---|---|
| Reference server (memory) | Locally | Small knowledge graph in a JSONL file | No | Yes (MIT) | Free |
| Basic Memory | Locally, cloud optional | Markdown files with an SQLite index, full-text and vector search | PDF, Word, PowerPoint, CSV | Yes (AGPL-3.0) | Free locally, cloud from $19 a month |
| Mem0 MCP | Hosted | At Mem0, with vector search and a graph over extracted entities | PDF on the platform only by URL, via MCP not stated | Library yes (Apache 2.0), hosted server not stated | Free plan with quota, usage-based above it |
| Supermemory MCP | Hosted, self-hostable | Cloud not stated; self-hosted with a built-in graph engine | PDF, including scans with OCR | Repo yes (MIT), server source not public | Free with $5 in credits, Pro $19 a month |
| Memexa | Hosted (EU) | Notes and documents as a knowledge graph, semantic and keyword search | PDF, scans, mail, audio, cited down to the page | No (export as Markdown) | From €2.99 a month, no free plan |
How to connect them
| Server | Setup |
|---|---|
| Reference server | npx -y @modelcontextprotocol/server-memory (Node.js 24 or later) or Docker |
| Basic Memory | Locally via uvx, cloud at https://cloud.basicmemory.com/mcp with browser sign-in |
| Mem0 MCP | https://mcp.mem0.ai/mcp, browser sign-in or API key |
| Supermemory MCP | https://mcp.supermemory.ai/mcp, OAuth sign-in |
| Memexa | https://brain.memexa.ai/mcp, OAuth sign-in |
Local servers go into your AI app’s configuration file and only run while your machine runs. Hosted servers you add in Claude as a custom connector, in ChatGPT as a custom MCP server under Plugins; how that works is in MCP servers explained.
The servers one by one
The MCP project’s reference server
The official MCP project maintains a memory server as an example: it stores entities (people, projects), relations between them and observations about each entity in a local file, now memory.jsonl. Search is text-based over names and observations. It doesn’t take in documents.
Fits if you want to try out how an MCP memory works, or keep a few facts about yourself and your projects. Free, open source, set up in five minutes. It isn’t built for an archive.
Basic Memory
Basic Memory stores everything as Markdown files in a folder on your machine, plus an SQLite index for full-text and vector search. You can open the files in any text editor. Locally it’s free and licensed under AGPL-3.0. The cloud starts at $19 a month per workspace, with usage credits (1 credit = 1 cent), and according to its privacy policy runs on servers in the US.
Version 1.0 brought documents (PDF, Word, PowerPoint, CSV) and agents, still in beta: you describe in plain language what an agent should do, such as linking related notes or tidying folders. Among the ready-made agents, “Cleanup” merges duplicates, archives finished work and splits notes that grew too big. Every change can be reviewed and undone. Agents run only on the cloud plans and use credits, according to Basic Memory usually under 10 per run.
Fits if local files in an open format matter to you and you like deciding yourself how things get tidied.
Mem0 MCP
Mem0 runs its MCP server hosted only: nothing runs on your machine, the connection goes to https://mcp.mem0.ai/mcp. Mem0 extracts facts and entities from what the AI stores and searches them later with vector search and a graph. For the first 30 days the free plan includes 10,000 adds and 1,000 retrievals a month; after that you need your own key for a language model to keep adding. Mem0 is headquartered in the US.
There used to be a local variant, OpenMemory. The matching folder in the repository is no longer reachable, and the OpenMemory page points to the hosted server. Whether OpenMemory is still maintained for local use, we couldn’t confirm.
Fits if you mainly want to collect conversation facts about yourself and a US service is fine for you. Mem0 is also widely used as a library by developers who build memory into their own applications.
Supermemory MCP
You connect Supermemory via OAuth at https://mcp.supermemory.ai/mcp, no API key needed. The service takes documents too; PDFs are read on all plans, with OCR for scans. The free plan includes $5 in credits a month, Pro costs $19. You can also run Supermemory yourself: the local version is free up to 10,000 documents, but its server source is not in the public repository. Supermemory doesn’t state where the hosted version stores your data.
Fits if you want a hosted service with documents and a certification counts: Supermemory lists SOC 2 as certified.
Memexa
Memexa runs hosted only, at Scaleway in Paris, with AI processing (OCR, embeddings, condensing) at Mistral, also in Paris. You put in notes, PDFs, scans, mail and voice recordings; everything is read, split into sections and condensed into linked notes. Every search hit carries its citation down to the page. New material is linked to what’s already there automatically: no rules for you to write, no credits. Memexa brings no language model of its own; your AI app provides that.
The limits: Memexa isn’t open source and doesn’t run locally. There’s no free plan; from €2.99 a month or €29.90 a year (all plans), and the 14-day trial requires a card. You can leave any time with a complete export as Markdown, JSON and original files.
Fits if your knowledge isn’t just what you tell the AI but documents, and later you want to know where an answer came from. Where the other providers keep your data is compared in EU-hosted AI memory.
Quick decision
- Just trying it out, free, local: the reference server.
- Markdown files on your own machine, tidying by your own rules: Basic Memory.
- Collecting conversation facts, hosted, also inside your own applications: Mem0.
- Hosted with documents and a certification: Supermemory.
- Documents with citations, hosted in the EU, upkeep automatic: Memexa.
Connecting several memory servers at once rarely helps: the AI then has to guess on every question where to look. Pick one and give it a few weeks.
Frequently asked questions
What is a memory MCP server?
An MCP server whose tools are about remembering and finding again. The AI app calls them itself: it stores what it considers important and searches there before answering. The memory then no longer belongs to a single AI app but is available to every one that speaks MCP.
Do I need a memory server if Claude and ChatGPT already have memory?
The built-in memories keep preferences and conversation facts inside one app. A memory server holds your state for all AI apps at once, and depending on the server, your documents too. What Claude’s own memory can do is in the overview of Claude memory.
Which memory server is free?
The reference server and Basic Memory locally are free and open source. Mem0 and Supermemory have free plans with quotas. Memexa has no free plan.
Sources
Retrieved November 3, 2026, Basic Memory’s agents page on November 7, 2026:
- Reference server: README in the MCP project’s repository
- Basic Memory: repository, pricing, agents, agent docs, what’s new in 1.0, privacy policy
- Mem0: MCP server, pricing, platform vs. open source, file formats, repository, OpenMemory
- Supermemory: MCP server, pricing, self-hosting, security, repository
- Memexa: How it works, trust and hosting