@memfork/vercel-ai — Vercel AI SDK Middleware

Drop-in middleware for the Vercel AI SDK that gives any model branch-aware, on-chain memory via MemForks.

Before generating: recalls relevant facts from the current branch and injects them as system context.
After generating: commits key decisions on-chain with full provenance.

Install

npm install @memfork/vercel-ai @memfork/core ai

# First-run — provision credentials (zero copy-paste on testnet):
memfork init --quick

Usage

Option A — explicit config

import { withMemForks } from "@memfork/vercel-ai";
import { generateText } from "ai";
import { openai } from "@ai-sdk/openai";

const model = withMemForks(openai("gpt-4o"), {
  treeId:  process.env.MEMFORK_TREE_ID!,
  signer:  process.env.MEMFORK_PRIVATE_KEY!,
  memwal: {
    accountId:   process.env.MEMFORK_MEMWAL_ACCOUNT!,
    delegateKey: process.env.MEMFORK_MEMWAL_KEY!,
  },
  branch: "feature/my-feature",
});

const { text } = await generateText({
  model,
  messages: [{ role: "user", content: "What did we decide about the auth system?" }],
});
// The model sees recalled facts about the auth system from prior sessions.
// After generating, the response is committed on-chain.

Option B — auto-resolve from ~/.memfork/credentials.json

import { createMemForksModel } from "@memfork/vercel-ai";
import { openai } from "@ai-sdk/openai";

const model = await createMemForksModel(openai("gpt-4o"), {
  branch: "feature/my-feature",
});

Branch per user session

import { withMemForks } from "@memfork/vercel-ai";

const model = withMemForks(openai("gpt-4o"), {
  ...config,
  branchFromContext: ({ messages }) => {
    // Map thread/session ID to a branch name
    const threadId = extractThreadId(messages);
    return `session/${threadId}`;
  },
});

How it works

| Phase | What happens | |-------|-------------| | transformParams (before generate) | Recalls top-N facts from the branch, prepends them to the system prompt | | wrapGenerate (after generate) | Fire-and-forget memfork commit anchors the response on-chain | | wrapStream (after stream) | Same commit, triggered when the stream closes |

Recall failures are silent — they never break a generation.
Commit failures are fire-and-forget — they never delay a response.

Options

| Option | Type | Default | Description | |--------|------|---------|-------------| | branch | string | "main" | Git branch to scope memory to | | recallLimit | number | 5 | Max facts to inject. Set 0 to disable recall | | autoCommit | boolean | true | Commit after generating. Set false for read-only mode | | recallThreshold | number | 0.4 | Semantic distance cutoff (lower = stricter) | | branchFromContext | fn | — | Dynamic branch from request messages |

Works with any Vercel AI SDK function

import { generateText, streamText, generateObject } from "ai";

// All three work identically — middleware applies to both generate and stream paths.
const { text }   = await generateText({ model, messages });
const stream     = await streamText({ model, messages });
const { object } = await generateObject({ model, messages, schema });