Best MCP servers for agent memory
11 working options, checked Nov 8, 2026.
Short answer: the top picks for agent memory right now are Memory, MemorySync and ContextStream. We list 11 options here: 2 work without sign-in, 6 need a sign-in and 3 run on your own machine.
A memory MCP server gives an AI assistant something it lacks by default: a record that lasts between conversations. It can save facts, preferences and decisions, then recall them later, even across different AI tools.
Top picks
-
Memory
A persistent knowledge-graph memory your agent can read and write across sessions.
-
MemorySync
One memory, every AI tool
-
ContextStream
Shared project context and hybrid code search for AI coding agents, so decisions carry across Cursor, Claude Code, Codex and more.
-
ClariLayer
Keep decisions and definitions with you across Claude Code, Cursor, and Codex sessions.
-
GreatArrow.ai
Shared business memory and actions for Claude, ChatGPT, Gemini, Cursor and other MCP clients.
All 11 compared
| Server | What it does | Setup | Tools |
|---|---|---|---|
| Memory | A persistent knowledge-graph memory your agent can read and write across sessions. | Runs locally | – |
| MemorySync | One memory, every AI tool | Hosted, sign-in | – |
| ContextStream | Shared project context and hybrid code search for AI coding agents, so decisions carry across Cursor, Claude Code, Codex and more. | Hosted, sign-in | – |
| ClariLayer | Keep decisions and definitions with you across Claude Code, Cursor, and Codex sessions. | Hosted, sign-in | – |
| GreatArrow.ai | Shared business memory and actions for Claude, ChatGPT, Gemini, Cursor and other MCP clients. | Hosted, sign-in | – |
| Neither | Decision memory for Cursor and Claude — cited Decision/Rejected/Constraint by file path | Runs locally | – |
| Avenbridge | Multi-tenant MCP knowledge-base memory with per-tenant hard spend caps | Hosted, no sign-in | 8 |
| Context Nest | Structured, versioned context for AI agents: typed Markdown nodes, selector queries, hash-chained versions and an optional review gate, over MCP or CLI. | Runs locally | – |
| brainattic | The knowledge base your AI reads and writes, under your rules - a governed, hosted MCP server. | Hosted, sign-in | – |
| superagnt_ | Agent toolkit for builders: one MCP server for database, memory, schedules and data tools. | Hosted, sign-in | – |
| MemoryRouter | Persistent memory for any MCP client. Search and save what the user tells you, then carry it into ChatGPT, Claude, Codex or OpenClaw. | Hosted, no sign-in | 10 |
How to choose
- For you or for a team? Personal memory follows one user, while shared memory lets a team and its agents work from the same context.
- Local or hosted? Local memory keeps data on your machine. Hosted memory follows you across devices and tools.
- Check how you edit it. Wrong memories are worse than none, so pick a server that lets you review and delete them.
How to connect one
Here is MemorySync as an example. In Claude Code, run:
claude mcp add --transport http memorysync https://mcp.memorysync.io/mcp
For other apps, follow our guides for Claude, ChatGPT, Cursor and VS Code.
How we picked these
We read each listing's tools and put the closest fit for this job first. Hosted servers must pass a real MCP handshake in our daily check to stay on this page, and local servers are reviewed from their source or package. That shows a server responds, not how well it works for you, and we don't run every tool in production.
Frequently asked questions
Which of these MCP servers work without signing in?
Avenbridge and MemoryRouter answered our latest live check with no account or key.
How do I give Claude persistent memory?
Connect a memory MCP server. The assistant can then store facts and retrieve them in later conversations, either on your machine or in a hosted service.
Can memory be shared across ChatGPT, Claude and Cursor?
Yes. Because MCP is a standard, a hosted memory server can connect to several assistants at once and give them the same context.
How often is this list updated?
Every hosted server here is re-checked daily with a real MCP handshake, and a server drops off this page while it fails that check. Last check: Nov 8, 2026.