How long does MCP setup for AI memory actually take?

About five minutes for any of the four major MCP clients (Claude Desktop, Cursor, Windsurf, Claude Code). The slowest step is signing up and generating an API key. The config edit itself is a single JSON block.

Does the MCP server work on any Alma plan?

Yes. The MCP server is free to install (npm package) and works on every paid Alma plan.

What if my MCP client is not Claude Desktop, Cursor, Windsurf or Claude Code?

The Alma MCP server speaks vanilla MCP over stdio, so any compliant client works. The pattern is always the same: command + args + env block. Set ALMA_DEBUG=1 to inspect the handshake if you get stuck.

How to Use MCP for AI Memory: 5-Minute Setup

Q: Can I use the same Alma memory across multiple MCP clients?

Yes. The same API key in different clients reads and writes the same memory store. Save a memory in Claude Desktop in the morning, reference it from Cursor in the afternoon, query from Claude Code on the terminal — same brain, every surface.

May 2026 · 7 min read · Fran Olivares, Founder of OlivaresAI

To use MCP for AI memory, install @olivaresai/alma-mcp as the server, paste a single block of JSON into your MCP client config (Claude Desktop, Cursor, Windsurf, Claude Code or any other compatible client) with your Alma API key, and restart. The server exposes 35 memory tools the AI can call autonomously: save a memory, search memories, assemble context, extract facts, configure Soul blocks. Total setup time is about five minutes and works on the free Alma tier.

The Model Context Protocol (MCP) is the open standard that lets AI clients talk to external tools and data sources. Once your client speaks MCP, you can plug in a memory server and the AI immediately gains tools to read from and write to a persistent memory store — across every conversation, every project, every day. This guide walks through the five-minute setup for the four most common clients in 2026 and answers the questions that come up afterwards.

What does MCP for AI memory actually give you?

When the Alma MCP server is connected, your AI client gets 35 new tools that it can call without you typing anything. The most important ones: alma_remember (save a fact or preference), alma_recall (search memories by natural-language query), alma_assemble_context (build a structured context for the current question from all memory layers), alma_extract (mine facts from a block of text), alma_update_block (configure Soul Engine blocks for AI identity), plus tools for episodes, procedures, environments, conversations, web search, file generation and more.

In practice, the AI calls these tools when the user asks something memory-relevant. "Remember that I prefer ESM modules" → the AI calls alma_remember. "What stack did we agree on for the new service?" → the AI calls alma_recall. The tool round-trip is sub-second; the user just sees a slightly more thoughtful answer.

How do I set up MCP for AI memory in Claude Desktop?

Step 1. Sign up at alma.olivares.ai. Start with the Starter plan ($14/month) for the entry tier.

Step 2. In Settings → API Keys, generate a key and copy it.

Step 3. Open Claude Desktop's config file. On macOS it lives at ~/Library/Application Support/Claude/claude_desktop_config.json; on Windows at %APPDATA%\Claude\claude_desktop_config.json.

Step 4. Add the Alma server inside the mcpServers object: { "alma": { "command": "npx", "args": ["-y", "@olivaresai/alma-mcp"], "env": { "ALMA_API_KEY": "your-key" } } }. If mcpServers doesn't exist yet, create it at the top level of the JSON.

Step 5. Restart Claude Desktop. The Alma tools appear in the slash-menu and Claude can call them autonomously. Try: "remember that my favourite editor is Helix". A turn later: "what's my favourite editor?" — Claude recalls.

How do I set up MCP for AI memory in Cursor?

Cursor's MCP UI lives in Settings → Features → MCP. Add a new server with name alma, command npx, args -y @olivaresai/alma-mcp, and add the env var ALMA_API_KEY with your key. Save and restart Cursor. The Alma tools appear in the chat tool-use indicator; Composer and the inline chat both have access. Combined with Cursor's project context, this gives the AI both your codebase awareness and persistent cross-project memory.

How do I set up MCP for AI memory in Windsurf?

Windsurf reads MCP configs from ~/.windsurf/mcp_config.json. The structure mirrors Claude Desktop: an mcpServers object with the same Alma block. Restart Windsurf after editing. The Alma tools then appear in Cascade and Windsurf can use them during agentic flows — particularly useful for multi-file refactors where you want the AI to recall conventions from prior sessions.

How do I set up MCP for AI memory in Claude Code?

Claude Code (the terminal-native Claude client) has its own MCP config at ~/.config/claude-code/mcp.json. Drop in the same Alma block. Run claude mcp list to verify the server is registered. Once connected, Claude Code can use Alma's tools during coding sessions; particularly powerful when combined with the Alma environments feature so each repo has its own memory namespace. See the Claude Code integration docs for the exact JSON.

What if my MCP client isn't on the list?

Alma's MCP server speaks vanilla MCP over stdio, so any compliant client works. The pattern is always the same: the client expects a command + args + env config block, runs the server as a child process, and discovers tools via the MCP handshake. If you get stuck on a non-listed client, the server logs (set ALMA_DEBUG=1 in the env) tell you whether the handshake succeeded.

Common workflows after setup

Stack capture. Tell the AI "I'm using Bun + Hono + D1 for the new service" — it calls alma_remember. Two days later in a different conversation: "scaffold a route handler for /users" — the AI recalls the stack and produces idiomatic Bun + Hono code, no re-explanation.

Decision logging. "We decided to skip Redis for the MVP and use D1 directly." The agent saves the decision with category decision and high importance. Weeks later when reviewing the architecture, "why no Redis?" returns the original decision with rationale.

Cross-client continuity. Save a memory in Claude Desktop in the morning, reference it from Cursor in the afternoon, query it from Claude Code on the terminal at night. Same Alma account, same memory store, every client sees everything.

Does MCP work with conversations, not just tools?

Yes. Alma's MCP server also exposes resources (read-only data the client can subscribe to): soul, memories, environments, conversations, budget, blocks, episodes, procedures and more. Clients that surface MCP resources (Claude Desktop in particular) show these in a sidebar and can stream updates. You can browse the entire memory store from inside Claude Desktop without opening a browser.

What about safety, cost and privacy?

The MCP server only does what the AI tells it to do, and only with your API key — there is no implicit broadcasting. All memory writes are scoped to your Alma account, encrypted at rest, and exportable on demand. The server runs locally on your machine; it doesn't proxy through any third party other than the Alma API itself. Cost-wise: the server is free to install; the Starter plan ($14/mo) is Alma's entry tier and includes everything needed to run the MCP server.

What's the next step after the 5-minute setup?

Once the server is running, the highest-leverage move is to populate Soul Engine blocks (personality, expertise, communication style, hard rules) so the AI has a consistent identity in addition to factual memory. The Alma web app at alma.olivares.ai exposes the Soul editor; see Soul Engine: AI Personality That Sticks for the full pattern.

See plans