Guide · Claude Desktop

Claude's memory stays in Claude. Yours shouldn't.

Claude Desktop remembers things — inside Claude. Tell it your stack on Monday and your Cursor agent still has no idea on Tuesday. Kognite is a memory both can use: a standard MCP server that Claude Desktop connects to like any connector, backed by the same store as every other tool you point at it.
Start freeRead the quickstart
Why

Built-in memory is a silo by design

Product memories are vendor memories: Claude's stays with Anthropic's apps, ChatGPT's with OpenAI's. That's fine until your workflow spans tools — chat in Claude Desktop, code in Cursor or Claude Code, review in VS Code. Each one meets you as a stranger.

An external memory over MCP inverts the ownership: the memory is yours, and the tools connect to it. Switch models or editors next quarter and the accumulated context comes with you.

Setup

Two ways in — pick one

Easiest (paid Claude plans): Settings → Connectors → “Add custom connector”, paste https://mcp.kognite.dev/mcp, then sign in and approve it in the browser tab that opens. No key to create first, and you can revoke it later from the dashboard. Done.

Config file: Settings → Developer → Edit Config, add the block below, then fully quit and reopen Claude Desktop:

claude_desktop_config.json
{
  "mcpServers": {
    "Kognite": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.kognite.dev/mcp"]
    }
  }
}
  • ✓The config-file path needs Node.js installed — mcp-remote runs via npx and bridges Claude Desktop to the remote server.
  • ✓A bare {"url": …} entry is rejected by Claude Desktop — remote servers must go through the bridge shown above. (The Connectors path has no such constraint.)
Heads-up

Two memories, one winner

Claude Desktop's own memory can intercept a bare “remember this” before the MCP tool is considered. If you want Kognite to be the system of record, either turn the built-in capability off under Settings → Capabilities, or name the tool: “use the Kognite memory_add tool to remember…”. For recall, “search my Kognite memory for…” is unambiguous.

What you get

Why route memory outside the app

  • ✓Portability — your memory isn't an app feature; it's a store you own, with a REST API and SDKs behind it.
  • ✓Sharing across surfaces — the same facts serve Claude Code in the terminal and Cursor in the editor.
  • ✓Structure — a knowledge graph with entities, relationships, and temporal edges, not an opaque blob you can't inspect. The dashboard shows exactly what's stored.
  • ✓Control — scopes isolate personal from team memory; the audit log records access; deletion is yours to invoke, entirely.
FAQ

Common questions

Do I need a paid Claude plan?

The Connectors path is a paid-plan feature. The config-file path via mcp-remote works generally and needs Node.js installed.

Should I turn off Claude's built-in memory?

If you want one authoritative memory across tools, yes — otherwise “remember” can land in the app-local memory your other tools can't see. Naming the Kognite tool explicitly also works.

Is this the same memory my coding tools use?

Yes — one URL per tool, one store behind all of them. Eight clients have documented setups.

What does it cost?

Free: 5,000 new memories/mo, searches unmetered. Paid from $19/mo — graph included on every plan.

One memory. Every Claude surface, every tool.

Free tier: 5,000 new memories a month with unmetered search. Two setup paths, both under a minute.

Start freeSee pricing