@usecarry/vercel-ai

Drop-in proof-carrying memory for the Vercel AI SDK.

Wrap any model in one line. Carry gates the user's memory before generation — the model only ever sees authorized memory — and emits an Answer Receipt (what was used, what was blocked) on every call.

import { withCarryMemory, createMemoryStore } from "@usecarry/vercel-ai";
import { generateText } from "ai";
import { openai } from "@ai-sdk/openai";

const store = createMemoryStore({
  memories: [
    { id: "m1", namespace: "health", content: "Allergic to penicillin" },
    { id: "m2", namespace: "billing", content: "Card ending 4242" },
  ],
  policy: { aria: { billing: false } }, // revoke billing for this agent
});

const model = withCarryMemory(openai("gpt-4o"), {
  store,
  agent: "aria",
  onReceipt: (r) => console.log(r.used, r.blockedNamespaces), // proof of what it used
});

const { text } = await generateText({ model, prompt: "Am I allergic to anything?" });
// → recalls only `health`, injects it, answers. `billing` is never fetched.

That's the whole integration. Works with generateText, streamText, generateObject — anything that takes a model.

Why

A raw vector-DB memory layer hands your model everything it retrieves and hopes for the best. Carry enforces an agent × namespace policy at retrieval, so a revoked namespace is never injected into the prompt — and every generation carries a receipt proving exactly what memory reached the model.

API

• withCarryMemory(model, opts) — wraps a model with the Carry middleware. Returns the same model type.
• carryMiddleware(opts) — the raw LanguageModelMiddleware if you compose your own wrapLanguageModel.
• createMemoryStore({ memories, policy }) — an in-memory Carry vault (default-allow; policy[agent][ns] = false revokes).
• CarryStore — implement recall(agent, query) to back it with your own store (Walrus, a DB, the carry CLI vault, …).

Options: { store, agent?, onReceipt? }. onReceipt(receipt) fires on every call with { agent, query, used, blockedNamespaces, createdAt }.

Part of Carry — proof-carrying memory for AI agents, stored on Walrus, enforced on Sui.

MIT © Carry