How to Add Persistent Memory to Claude Desktop in 60 Seconds

Every time you open Claude Desktop, it starts blank. No memory of the project you discussed yesterday. No recollection of your preferences, your stack, or the architectural decision you spent 40 minutes explaining last week. Brilliant model. Zero memory.

You can fix that in 60 seconds. Here's how.

Why Claude Forgets (and Why You're Still the One Repeating Yourself)

Anthropic has shipped memory features inside Claude.ai. But Claude Desktop — the version power users actually run, the one with MCP, the one developers configure locally — still starts every session completely fresh.

It's not a bug. It's a boundary. Anthropic doesn't own your memory, and that's actually the right call. It means the memory layer is yours to build and to own. And thanks to MCP, you can add it in a way that works across all the AI tools you use, not just Claude.

MCP stands for Model Context Protocol. It's an open standard that lets AI tools connect to external servers and retrieve information on demand. Think of it as a plugin system for context. Any MCP server can serve up memories, notes, or files — and Claude reads them automatically at the start of a session.

The Hard Way: Docker, SQLite, and DIY Configs

If you've tried to add memory to Claude before, you've probably hit the DIY route. Find a GitHub repo with a local MCP memory server, install dependencies, configure SQLite or Postgres, wire up a JSON config, debug why the server crashed, fix the schema migrations, realize it only works when the local process is running, and eventually close the terminal and open a fresh Claude chat anyway.

These repos exist, and they're not bad. But you're maintaining a local server, keeping it alive in the background, dealing with breakage every time your machine restarts or Node updates. Most people try it once, hit friction, and quietly go back to re-explaining their stack every morning.

There is a faster path.

What You Need Before You Start

Claude Desktop is installed on your machine. A free Kumbukum account. That's it.

Sign up at kumbukum.com — the free plan gives you 500 memories and notes, 1,000 API calls per month, and full semantic search. No credit card required.

Step 1: Get Your Kumbukum MCP URL and API Key

After signing up, your dashboard shows two things you need:

Your MCP server URL: https://mcp.kumbukum.com/mcp

Your personal API key: a string starting with something like kk_...

Copy both. You'll paste them into Claude Desktop's config in the next step.

Step 2: Add Kumbukum to Claude Desktop

Claude Desktop stores its MCP server configuration in a JSON file. Open it with any text editor.

On macOS:

~/Library/Application Support/Claude/claude_desktop_config.json

On Windows:

%APPDATA%\Claude\claude_desktop_config.json

If the file doesn't exist yet, create it. Add the following, replacing YOUR_API_KEY with your actual key:

{ "mcpServers": { "kumbukum": { "url": "https://mcp.kumbukum.com/mcp", "headers": { "Authorization": "Bearer YOUR_API_KEY" } } } }

If you already have other MCP servers configured, add the kumbukum entry inside the existing mcpServers object.

Step 3: Restart Claude Desktop

Save the file and restart Claude Desktop. That's the entire setup.

When Claude loads, it now has access to your Kumbukum memory store. You'll see Kumbukum listed in Claude's available tools in the MCP integrations panel.

Step 4: Store Your First Memory

Start a new Claude conversation and tell it something you want remembered:

"Remember that I'm building a SaaS product called Acme, we use TypeScript and Postgres, and I prefer short answers without bullet points."

Claude calls Kumbukum's store_memory tool automatically and confirms it's saved. From that point on, any Claude session connected to Kumbukum can retrieve that context.

What Claude Can Remember Now

Once Kumbukum is connected, Claude has access to two types of stored information.

Memories are short-form context: your preferences, project decisions, things you've told Claude to keep in mind. "I prefer TypeScript. Always async/await. My primary database is Postgres." Tell Claude once. Done.

Notes are structured documents: API specs, reference materials, onboarding docs — anything you want Claude to pull up accurately rather than hallucinate. You store them once; Claude retrieves the right ones when relevant.

Retrieval is automatic. When you start a session and mention a project, Claude searches your memory store and surfaces the relevant context without you having to ask.

Adding Kumbukum to Cursor

The process is identical for Cursor. Open Cursor's settings, go to Settings > Features > MCP, and add a new server entry with the same URL and API key.

Once added, Claude and Cursor share the same memory store. A decision you log in Claude shows up in Cursor. The architectural choices you made last week travel with you across every tool. That's the actual fix: not "Claude has memory" — your entire AI stack has a shared brain. Check Cursor's MCP documentation if you need the manual config path.

What to Store (and What Not To)

Store: preferences, recurring decisions, project context, client requirements, team conventions. Things that don't change session to session but that your AI needs to work well.

Don't store: things that change constantly, things you'd rather look up fresh, or everything. Kumbukum is not a second brain for every thought you've ever had. Store what makes your AI more useful.

Works Across Your Whole Stack

If your team manages files and digital assets through a platform like Razuna, you can store the context behind those assets in Kumbukum — naming conventions, approval workflows, project briefs — so your AI always has the full picture when working with them.

Teams using a shared inbox like Helpmonks for customer support or internal email can store recurring context in Kumbukum — tone guidelines, customer background, escalation rules — and have their AI pull it in automatically every session.

The Honest Tradeoffs

Kumbukum is a hosted service. Your memories live on their servers, not locally. If that's a hard requirement for your setup, the DIY route is still the right call.

One more thing that matters: Kumbukum is open source. You can inspect the code, self-host it, or contribute at the GitHub repository.

Frequently Asked Questions

Does Kumbukum work with ChatGPT?

Yes, any MCP-compatible client works. ChatGPT supports MCP via its connector settings. Same URL, same API key, same memory store.

Is my data private?

Your memories are private to your account and never shared with other users or used to train AI models. You can export or delete your data at any time.

How is this different from just keeping a CLAUDE.md file?

A markdown file is local, manual, and visible only to the tool that reads it. Kumbukum is client-server: your memory store is accessible from every tool you connect, on any device, without copying files or remembering to update them.

Claude is one of the best AI models available. Brilliant, but "amnesiac" is still a frustrating way to use it.

Adding memory is a 60-second config edit and one free account. Try Kumbukum free and stop re-explaining your stack every morning.