@tracemem/vercel-ai

TraceMem integration for Vercel AI SDK. Seamlessly add TraceMem's decision memory and MCP capabilities to your Vercel AI agents.

Quickstart

1. Install:

npm install @tracemem/vercel-ai @tracemem/ts-sdk zod

2. Generate Tools:

import { createTraceMemTools } from '@tracemem/vercel-ai';

const tools = createTraceMemTools({
  apiKey: process.env.TRACEMEM_API_KEY,
  // Optional: Add context to every tool call
  context: ({ tool }) => ({
    route: '/api/chat',
    metadata: { source: 'vercel-ai-sdk' }
  })
});

3. Use in Vercel AI SDK (Next.js App Router):

// app/api/chat/route.ts
import { openai } from '@ai-sdk/openai';
import { streamText } from 'ai';
import { createTraceMemTools } from '@tracemem/vercel-ai';

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

  const result = await streamText({
    model: openai('gpt-5.2-codex'),
    messages,
    system: `
      You are a helpful assistant. 
      Use the 'tracememOpen' tool to start a tracked decision context if dealing with complex tasks.
      Always capture the returned 'decisionId' and pass it to subsequent tools (note, read, write, close).
    `,
    tools: createTraceMemTools({
      apiKey: process.env.TRACEMEM_API_KEY,
      defaults: { automationMode: 'propose' }
    }),
  });

  return result.toDataStreamResponse();
}

Tools Provided

The library exports a set of tools mapped to the TraceMem SDK. By default, they are named:

• tracememOpen(action, intent?) -> Returns { decisionId }
• tracememNote(decisionId, message, kind?)
• tracememRead(decisionId, product, purpose, query?)
• tracememEvaluate(decisionId, policy, inputs)
• tracememRequestApproval(decisionId, message)
• tracememWrite(decisionId, product, purpose, mutation)
• tracememClose(decisionId, outcome)
• ...and more.

Context Injection

You can inject metadata into every TraceMem interaction using a context provider:

const tools = createTraceMemTools({
  apiKey: process.env.KEY,
  context: async ({ tool, args }) => {
    // Return context object to be merged into metadata
    return {
      userId: 'user_123',
      requestId: 'req_abc'
    };
  }
});

Security & Redaction

This library automatically attempts to redact sensitive keys (like token, password, secret) from tool inputs before passing them to TraceMem or logging them in context. However, you must ensure that your agent prompts do not encourage leaking secrets into non-redacted fields.

License

Apache License 2.0