Give any Vercel AI SDK agent persistent memory in 3 steps (MCP, no SDK required)

Most agents forget everything between sessions. The Vercel AI SDK makes this easy to fix because it speaks MCP natively — and BlueColumn ships an MCP server on npm, so there is nothing to install and no SDK to learn.

1. No package needed

The SDK has first-class MCP client support, and bluecolumn-mcp runs via npx.

2. Connect (app/api/chat/route.ts)

import { streamText, experimental_createMCPClient as createMCPClient } from 'ai';
import { Experimental_StdioMCPTransport as StdioMCPTransport } from 'ai/mcp-stdio';

export async function POST(req: Request) {
  const { messages } = await req.json();

  const mcpClient = await createMCPClient({
    transport: new StdioMCPTransport({
      command: 'npx',
      args: ['-y', 'bluecolumn-mcp@latest'],
      env: { BLUECOLUMN_API_KEY: process.env.BLUECOLUMN_API_KEY! },
    }),
  });

  try {
    const tools = await mcpClient.tools(); // remember, recall, note
    const result = streamText({
      model: yourModel,
      system: 'You have persistent memory. Store durable facts with remember; search with recall before answering questions that depend on past context.',
      messages,
      tools,
    });
    return result.toDataStreamResponse();
  } finally {
    await mcpClient.close();
  }
}

3. Prompt the agent

Add to your system prompt: "Store durable facts immediately with remember. Search with recall before answering questions that depend on past context."

Why this pattern works

• recall returns cited sources — surface them in your UI and every answer is verifiable against the stored memory.
• Writes are idempotent — agents retry, and retries do not triple-store the same fact.
• Namespaces isolate per API key, so one key per user or tenant and nothing leaks across.
• Free tier: 100 writes + 100 reads + 30 audio minutes/mo at https://bluecolumn.ai

One deployment note

Stdio MCP clients spawn a local process, which does not fit edge runtimes. For serverless deployments, keep the route on a long-running Node runtime, or call the memory API directly with two fetch calls:

// write
await fetch('https://api.bluecolumn.ai/remember', { method: 'POST', headers: { Authorization: `Bearer ${key}` }, body: JSON.stringify({ text }) });
// recall
const res = await fetch('https://api.bluecolumn.ai/recall', { method: 'POST', headers: { Authorization: `Bearer ${key}` }, body: JSON.stringify({ q }) });

Same memory, two HTTP calls, works anywhere.

Docs: https://bluecolumn.ai/docs