@memhq/vercel-ai

MemHQ middleware for the Vercel AI SDK. Wraps streamText / generateText to prepend user memories before the model call and persist the assistant turn after — no changes to your prompt code.

Install

npm install @memhq/vercel-ai @memhq/sdk ai

Quickstart

import { withMemHQ } from "@memhq/vercel-ai";
import { streamText } from "ai";
import { openai } from "@ai-sdk/openai";

const result = await withMemHQ(
  { apiKey: process.env.MEMHQ_API_KEY!, userId: "user_42" },
  (args) =>
    streamText({
      model: openai("gpt-4o-mini"),
      ...args,
    }),
  {
    messages: [{ role: "user", content: "What's my favorite coffee?" }],
  },
);

for await (const chunk of result.textStream) {
  process.stdout.write(chunk);
}

withMemHQ:

1. Pulls the latest user message
2. Calls MemHQ search and injects the top results as a system message
3. Hooks onFinish to persist the user + assistant turn into MemHQ

Non-streaming

import { generateText } from "ai";

const result = await withMemHQ(
  { apiKey: process.env.MEMHQ_API_KEY!, userId: "user_42" },
  (args) => generateText({ model: openai("gpt-4o-mini"), ...args }),
  { messages: [{ role: "user", content: "Suggest a gift for me" }] },
);

console.log(result.text);

Configuration

| Option | Description | | --- | --- | | apiKey | MemHQ project API key. Falls back to MEMHQ_API_KEY. | | userId | External user id. Required. | | recallLimit | Memories prepended per call. Default 5. | | searchMode | "hybrid" (default), "vector", or "lexical". | | sessionId | Tagged onto persisted messages as metadata. | | baseUrl | Override the API base. |

Reference

Full reference: https://docs.memhq.ai/sdks/vercel-ai