Vercel AI SDK | FishMem Docs

Install the framework and the universal FishMem client:

npm install ai @fishmem/sdk

Create one server-side client:

import { FishMem } from "@fishmem/sdk";

export const fishmem = new FishMem({
  apiKey: process.env.FISHMEM_API_KEY!,
});

Before calling the model, search with the current user scope and format a small context block:

const { results } = await fishmem.memories.search({
  query: latestUserMessage,
  user_id: userId,
  top_k: 5,
});

const memoryContext = results
  .map((item) => `- ${item.memory}`)
  .join("\n");

Add that block to your system context. After the turn, store only a durable preference, constraint, decision, or fact:

await fishmem.memories.add({
  content: "Alex wants TypeScript examples with no framework fallback.",
  infer: false,
  user_id: userId,
  metadata: { source: "assistant-turn" },
});

Do not send the entire transcript automatically. Your application or agent should decide what is durable, then submit a self-contained memory.