The 9 Best MCP Memory Servers in 2026, Compared

An MCP memory server is a Model Context Protocol server that gives an AI client such as Claude, Cursor or ChatGPT a store it can write to and search, so what it learns survives the end of a session. The best MCP memory servers in 2026 are the reference Knowledge Graph Memory Server, Basic Memory, mcp-memory-service, Mem0 MCP, Supermemory MCP, the Graphiti MCP server, Zep's Context MCP Server, Cognee MCP and Coworker MCP.

The right pick depends on whose memory it is. Memory for one developer on one laptop is largely solved by free open-source servers. Memory shared by a company's agents and people has to come from the systems where work happens, and it has to respect who may see what.

I read each project's repository, docs and pricing on October 6, 2026, and star counts are from GitHub that day. Two recent changes matter: Mem0 removed its local OpenMemory server on July 29, 2026, and ByteRover's newer V4 runs through a skill, and its docs do not mention MCP. For basics and wider shortlists, see what MCP is, the Claude Code MCP servers list, how to choose MCP servers and the best MCP servers for databases.

Do you need an MCP memory server, or is Claude's built-in memory enough?

Before adding a Claude memory MCP server, check what Claude already does. Claude Code has CLAUDE.md files that you write, plus auto memory, where Claude notes your preferences and corrections. Anthropic's docs say auto memory is machine-local and is not shared across machines or cloud environments. In the Claude apps, memory is on by default for Free, Pro and Max plans, and on Team and Enterprise plans an owner decides whether members get it (Claude Help Center).

That memory stays inside Claude. An MCP memory server earns its place when you want one memory behind several clients, memory you can open and back up, or a store that several agents or people share. Memory also differs from the context window, which empties when a session ends; the AI agent memory guide covers the theory.

MCP memory servers compared at a glance

Local MCP memory servers for one developer

1. Knowledge Graph Memory Server (the reference memory MCP)

The Memory server in the official modelcontextprotocol/servers repository is the protocol's own reference implementation, and its README calls it the Knowledge Graph Memory Server. It stores entities, directed relations and observations (one fact each) in a local JSONL file, with nine tools to create, search, read and delete them.

claude mcp add memory -- npx -y @modelcontextprotocol/server-memory

Claude Desktop takes the same npx command in claude_desktop_config.json; MEMORY_FILE_PATH sets the file's location.

Strength: no API keys, nothing to configure, and one readable file.

Honest limitation: search_nodes is a case-insensitive substring match with no embeddings, so "TypeScript" finds your TypeScript preference but "the language I prefer" does not. The repository calls its servers reference implementations and educational examples. License: MIT per its README. GitHub: 91,044 stars for the whole servers repository.

2. Basic Memory MCP

Basic Memory keeps memory as Markdown files in ~/basic-memory, indexed for full-text and vector search with local FastEmbed embeddings. Observations and wikilinks in the notes form its knowledge graph, and the same folder opens in Obsidian.

claude mcp add basic-memory -- uvx --prerelease=allow basic-memory mcp

Per the README, the --prerelease=allow flag is required; without it uv silently installs an older release.

Strength: memory you can read and fix by hand, and Basic Memory Teams adds a shared cloud workspace for groups.

Honest limitation: it needs Python through uv, and sync across devices is manual unless you pay for Basic Memory Cloud, from $15 per seat a month after a 7-day trial. License: AGPL-3.0. GitHub: 4,106 stars.

3. mcp-memory-service

mcp-memory-service is a community project that runs as one self-hosted service with MCP, a REST API and a web dashboard. It embeds memories locally with ONNX models, links them with typed edges such as causes and contradicts, and consolidates old memories, storing them in SQLite, Cloudflare, a hybrid of the two, or Milvus. After pip install mcp-memory-service:

claude mcp add memory -- memory server

Strength: built for many agents on one store, with X-Agent-ID tagging, a REST API for LangGraph, CrewAI and AutoGen, and remote MCP over OAuth for claude.ai and ChatGPT.

Honest limitation: more moving parts than the reference server, and it binds to localhost by default, so exposing it to a network safely is your job. License: Apache-2.0. GitHub: 1,987 stars.

Hosted memory MCP servers with nothing to run

4. Mem0 MCP

Mem0's official MCP server is a hosted service at mcp.mem0.ai: nothing runs on your machine, and your memories live in your Mem0 account (Mem0 docs). An LLM extracts facts, decisions and preferences from what you add, and search combines semantic and keyword scoring with a built-in graph that links memories mentioning the same entities. Its 11 tools scope memories to a user, agent, app or run.

npx mcp-add --name mem0-mcp --type http --url "https://mcp.mem0.ai/mcp" --clients "claude code,cursor"

Claude Desktop rejects that helper, so add the URL under Settings, Connectors, Add custom connector.

Strength: zero infrastructure, and its Claude Code plugin keeps a shared memory pool per repository.

Honest limitation: your memories sit in Mem0's cloud. The free Hobby plan allows 10,000 add requests and 1,000 retrieval requests a month, and Starter is $19 a month (Mem0 pricing). The graph records no typed relationships. SDK: Apache-2.0, 66,671 GitHub stars.

5. Supermemory MCP

Supermemory's MCP server is a remote endpoint at mcp.supermemory.ai that signs you in with OAuth instead of an API key. Memories, documents and a profile of stable and recent facts live in spaces, and you choose which spaces each client can reach.

claude mcp add --transport http supermemory https://mcp.supermemory.ai/mcp

Strength: one account behind many clients, team spaces, and an interactive memory graph in clients that support MCP Apps. A local mode (npx supermemory local) runs its Memory API on your machine.

Honest limitation: the documented MCP endpoint is the hosted one. The free plan includes $5 of credits a month, and Pro is $19 a month with three team seats (Supermemory pricing). License: MIT. GitHub: 31,121 stars.

Knowledge graph memory MCP servers you run yourself

6. Graphiti MCP server

Graphiti is Zep's open-source framework for temporal knowledge graphs, and its repository includes an MCP server that the README labels experimental. Every fact is an edge with a validity window: when something changes, the old fact is invalidated rather than deleted, so an agent can ask what was true last quarter. Docker Compose starts FalkorDB and the server at localhost:8000/mcp/.

claude mcp add --transport http graphiti http://localhost:8000/mcp/

Strength: explicit history for facts that change, with provenance back to each episode.

Honest limitation: each episode triggers several LLM calls (OpenAI by default), so you pay for tokens and may need to lower SEMAPHORE_LIMIT to avoid 429 errors. For the Docker server, Claude Desktop needs a gateway such as mcp-remote (or run Graphiti over stdio), and anonymous telemetry stays on until you set GRAPHITI_TELEMETRY_ENABLED=false. License: Apache-2.0. GitHub: 31,488 stars.

7. Cognee MCP

Cognee turns documents, code and conversations into a knowledge graph plus vectors (SQLite, Kuzu and LanceDB in its default template), and cognee-mcp centers on three memory tools: remember, recall and forget. Since version 1.6.0 (September 2026) its core library can run on local models with no LLM key, though the MCP quick start uses an OpenAI key. With its Docker image in HTTP mode:

claude mcp add cognee-http -t http http://localhost:8000/mcp

Strength: in API mode, several MCP instances share one Cognee backend, so a team queries the same graph (Cognee docs).

Honest limitation: more setup than the file-based servers, and using the client's own model instead of an API key needs MCP sampling, which the README says Claude Code did not grant as of early 2026. License: Apache-2.0. GitHub: 31,477 stars.

Memory MCP servers built for a whole company

8. Zep Context MCP Server

Zep's Context MCP Server is the managed counterpart to Graphiti, built for organizations. Each person signs in through Google Workspace or, on Enterprise, your own OIDC identity provider, and reaches their own memory graph from Claude, ChatGPT, Claude Code, Codex or Cursor. The sign-in identity fixes which graph a token can reach, and admins can share Context Graphs project-wide or per user group. Zep builds the graphs from data you provide through its SDK, its zep-ingest tool or its Batch API (Zep docs).

Strength: temporal facts, per-user isolation, a bring-your-own-cloud option, and SOC 2 Type II plus a HIPAA BAA on Enterprise plans. Best if your engineers already build agents on Zep.

Honest limitation: someone has to build and run the ingestion. Flex is $125 a month with 5 Memory MCP Server seats, and Flex Plus is $375 a month with 15 (Zep pricing).

9. Coworker MCP

Coworker MCP exposes your company's data and workflows to any MCP-compatible AI tool, and its memory is OM2. You can think of OM2 as a context graph: it connects to 50+ work tools such as Slack, Jira, Salesforce, Google Drive, Notion and GitHub, discovers the people, companies, projects and deals in them, and breaks documents, messages and tickets into atomic facts linked by entities and relationships. It reads those tools continuously and invalidates outdated information.

The difference is where the memory comes from. If a renewal date sits in Salesforce and a Slack thread, nobody has to tell an agent to save it. Access policies from the source tools travel with every fact, so people and agents only recall what they could already see, with provenance back to the source. It works in Claude, Claude Code, ChatGPT, Cursor and Windsurf, with your existing Coworker login and SSO.

Strength: company knowledge that nobody has to curate, shared across agents and people within their existing permissions. Coworker is SOC 2 Type II, GDPR and CASA Tier 2 compliant and does not train models on customer data.

Honest limitation: it is not a personal memory tool. It runs on isolated single-tenant infrastructure in Coworker's cloud or via VPC, not on your laptop, and you start by booking a demo rather than running a command. To have Claude remember that you prefer pnpm, use an open-source server above.

Why aren't Letta, OpenMemory, ByteRover and Context7 on this list?

Letta has a first-party hosted MCP server, but it delegates work rather than storing memory: its six tools list, create and message stateful Letta agents, which keep their own memory and run tools on Letta Cloud without approval prompts. See Mem0 vs Zep vs Letta for Letta as a framework.

OpenMemory, Mem0's local MCP server, was announced as sunset in April 2026 and removed from Mem0's repository on July 29, 2026 (pull request).

ByteRover's V3 CLI, formerly Cipher, still documents an MCP connector, the default for Claude Desktop, but its GitHub repository is archived and it has had no npm release since May 27, 2026. The newer V4 runs through a desktop app and a skill, and none of its 17 doc pages mention MCP.

Context7 is a context MCP server, not a memory server: it pulls current library documentation into the prompt and remembers nothing about you (GitHub).

How do you choose a memory MCP server?

Before installing one, ask:

1. Where does the memory live, and can you export it?
2. Does an LLM extraction pass run, and cost tokens, on every write?
3. What happens when a fact changes?
4. Does it respect the permissions of the system each fact came from?

For the security side, see MCP security and, for company rollouts, enterprise MCP.

How do you add an MCP memory server to Claude Desktop, Claude Code and Cursor?

In Claude Code, add a local server with claude mcp add, a name, two dashes and its start command, or a remote one with --transport http, a name and its URL (Claude Code docs). Claude Desktop reads local servers from claude_desktop_config.json and takes remote ones under Settings, Connectors, Add custom connector; Cursor reads .cursor/mcp.json. Then test it: tell it something, then ask in a new chat. The Claude MCP guide has more.

The short version

For your own memory, an open-source server is free to run and does the job. For one permission-aware memory that every agent and person at your company shares, built from the tools you already use, book a demo to see Coworker MCP on your own stack. If you want numbers first, ask for a Context Impact Report, where Coworker runs its benchmark on your company's data.

Frequently asked questions

What is an MCP memory server?

An MCP memory server is a Model Context Protocol server that gives an AI client a persistent store it can write to and search across sessions. MCP clients such as Claude, Cursor or ChatGPT call its tools, so one memory can serve several clients.

What is the best MCP memory server for Claude Code?

Start with Claude Code's built-in CLAUDE.md files and auto memory, which keep your preferences per repository on your machine. For the same memory in other clients, add Basic Memory for readable Markdown notes, Mem0 MCP or Supermemory MCP for hosted memory, or the reference server for the simplest local setup.

What happened to OpenMemory MCP?

Mem0 announced the sunset of OpenMemory, its local MCP memory server, in April 2026 and removed it from its repository on July 29, 2026. Mem0's supported MCP option is now its hosted server at mcp.mem0.ai.

What is the difference between a knowledge graph memory MCP and a vector memory MCP?

A vector memory server stores text as embeddings and returns whatever is semantically similar to a query. A knowledge graph memory server stores entities and their relationships, so it can follow connections such as who owns a project. Basic Memory, Mem0, Cognee and Graphiti combine both; the enterprise knowledge graph explainer goes deeper.

Can a team share one MCP memory server?

Yes. Basic Memory Teams and Supermemory spaces give a team a shared hosted workspace, and Cognee's API mode lets several MCP instances share one graph. For memory tied to company identity, Zep's Context MCP Server maps each person through single sign-on, and Coworker MCP builds shared memory from the company's own tools while keeping their permissions.

Is it safe to put company data in an MCP memory server?

It depends on where the server runs and whether it respects permissions. A local server knows nothing about who else should see the data, and a hosted one stores memories in the vendor's cloud. Check where memories live, whether access follows your identity provider, and whether each fact keeps its source system's permissions.

What is the difference between a memory MCP and a context MCP?

A memory MCP server stores what an AI learns and returns it in later sessions. A context MCP server supplies information the AI does not need to remember, such as Context7, which pulls current library documentation into the prompt.

Related reading

• What is MCP? A complete guide
• 12 Claude Code MCP servers every AI team needs
• How to choose MCP servers for your team
• Best MCP servers for databases
• What is AI agent memory?
• Mem0 vs Zep vs Letta
• What is an enterprise knowledge graph?
• MCP security: risks and mitigations
• Organizational memory (OM2)
• Coworker MCP