MemorySync

Model Context Protocol

One memory, every AI tool

Connect MemorySync to the assistants you already use. Context follows you between them, and a second server hands coding agents the current documentation.

Read the MCP docsConfigure a client

Works with Claude, Cursor, VS Code, Devin Desktop, Codex, OpenCode and any client that speaks remote MCP.

Run this, then restart your client.

bash
npx -y memorysync-mcp-install

The first memory tool call opens your browser to authorize access. There is no key to paste and nothing to register.

15
Memory tools
4
Resources
2
Prompts
7
Clients documented
0
Keys to paste

What you connect to

Two servers, two jobs

Both are hosted and speak Streamable HTTP, so nothing runs on your machine. Connect to either or both.

Memory server

Fifteen tools over your memory, four resources and two prompts. Sign in through the browser with OAuth 2.1, or send an API key for headless and CI use.

  • Search and store durable facts, preferences and decisions
  • Follow typed relationships, including contradictions
  • Read the standing profile derived from stored memory
text
https://mcp.memorysync.io/mcp

Documentation server

Search and read every MemorySync documentation page from inside your coding agent. No account, no key, no consent screen.

  • Stops an agent inventing method names and parameters
  • Returns the plain Markdown of any page
  • Ships a prompt that makes the agent read before it writes
text
https://docs.memorysync.io/mcp

Tool surface

What the assistant can call

Every tool below exists on the memory server. Which of them a given client is offered depends on the permission you approve when you connect it: grant mcp:read and it gets the read tools only; grant mcp:write and the write and delete tools appear too. You choose per connection.

Read

mcp:read

Search, list, follow relationships, inspect status and read the profile.

  • search_memory
  • list_memories
  • get_memory
  • get_related_memories
  • list_entities
  • list_memory_events
  • get_event_status
  • get_memory_profile
  • list_connected_sources
  • whoami

Write

mcp:write

Store a new fact, or change the labels on one that already exists.

  • add_memory
  • update_memory

Delete

mcp:write

Remove memory. The two bulk tools preview first and need an explicit confirmation.

  • delete_memory
  • delete_entity
  • delete_all_memories

A tool the connection cannot call is withheld rather than advertised and then refused, so a read-only connection looks like a coherent read-only server instead of one whose tools fail. Call whoami at any time to see which permissions were granted.

Setup

Client-specific setup

Each client keeps its MCP configuration somewhere different, in a shape of its own. Open yours for the exact file and the exact block — a config written in another client's format usually parses without error and then never connects.

Run this in your terminal.

bash
claude mcp add memorysync --transport http https://mcp.memorysync.io/mcp

Or add it manually to ~/.claude.json:

json
{
  "mcpServers": {
    "memorysync": {
      "type": "http",
      "url": "https://mcp.memorysync.io/mcp"
    }
  }
}

Claude Code speaks Streamable HTTP directly and runs the browser sign-in itself on the first tool call.

Why it matters

What changes once it is connected

Without memory every conversation starts from nothing, in every separate tool. These are the differences you notice first.
01

Context carries between tools

A preference recorded in Claude is available in Cursor. Both read the same memory instead of their own local history, so you stop re-explaining your stack to every new session.

02

The assistant decides when to remember

You do not call an API. The model searches memory before answering and stores a fact when you state one, because each tool describes when it applies.

03

Contradictions surface instead of averaging

MemorySync records typed, confidence-weighted relationships between memories, including contradictions. When two facts disagree the assistant follows the edge and asks which one still holds.

04

Destructive calls need a real confirmation

delete_entity and delete_all_memories return a count first and act only on an exact confirmation string, so an ambiguous instruction cannot clear a scope.

05

Coding agents read the current docs

The second server exposes every documentation page, so an agent writes integration code from the live reference rather than from training data that has drifted.

06

Connected sources are visible

GitHub, Slack, Notion, Google Drive, OneDrive, Granola, Amazon S3 and crawled sites arrive as memory the assistant can cite and explain the origin of.

Documentation

Keep reading

MCP overviewWhat each server exposes, and how the two differ.Connect a clientPer-client configuration, API-key auth for CI, and what to check when a client will not connect.Memory toolsEvery tool, what it takes and returns, plus the resources, prompts and scopes.IntegrationsThe knowledge sources that feed the memory your assistant reads.

Model Context Protocol

Give your assistant a memory that lasts

Add one URL to the tools you already use. Sign in from the client and your context carries between them.

Get startedMCP docs

Free to start. No credit card, and no API key needed for MCP sign-in.