@onemem/vercel-ai-provider

Wrap any Vercel AI SDK model so every model call is recorded as a verifiable on-chain OneMem TraceSession (Sui + Walrus + Seal) — Merkle-chained ActionCalls with content stored on Walrus and Seal-encrypted.

Publication note, 2026-06-18: @onemem/[email protected] is current on npm after pnpm registry:status --strict and includes createOneMemMemory(...). Re-run that command before making a fresh public install claim.

Usage

import { createOneMemMemory, withOneMem } from "@onemem/vercel-ai-provider";
import { openai } from "@ai-sdk/openai";
import { generateText, streamText } from "ai";

const model = withOneMem(openai("gpt-4o"), {
  agentId: "my-app",          // optional, stamped on the trace
  // target / signer / network are all optional — auto-provisioned on first use
});

// Use it like any AI SDK model — generate and stream both record a trace.
const { text } = await generateText({ model, prompt: "What did we decide about auth?" });

// Optional explicit memory helper — recall before a call, capture after.
const memory = createOneMemMemory();
const prompt = await memory.recallInto("What did we decide about auth?");
const result = await generateText({ model, prompt });
await memory.capture(`User asked: ${prompt}\nAssistant answered: ${result.text}`);

How it works

withOneMem wraps the model with AI SDK middleware (wrapLanguageModel → wrapGenerate / wrapStream). After each call it records a 1-call TraceSession via @onemem/sdk-ts/runtime's recordSession. Recording is fire-and-forget: it never adds latency to, or breaks, your model call — a OneMem failure is logged and swallowed.

Zero-config: the namespace, ReadWrite cap, and signer are auto-provisioned on first use and persisted under ~/.onemem/ (same engine as the OpenClaw/Hermes plugins). Override via options or env (ONEMEM_NAMESPACE_ID + ONEMEM_RW_CAP_ID, ONEMEM_PRIVATE_KEY, SUI_NETWORK).

Options

| Option | Default | Notes | |---|---|---| | agentId / environment | "vercel-ai" | stamped on the trace | | network | $SUI_NETWORK or testnet | | | privateKey | sui keystore → generated wallet | signer | | target | env → auto-provisioned | { namespaceId, rwCapId } | | enableTrace | true | set false to disable | | onTrace | — | callback with the on-chain session id |

Memory helper

createOneMemMemory() exposes the Mem0-style side of OneMem for Vercel AI apps:

• recall(query, topK?) searches configured MemWal memories.
• recallInto(input, topK?) prepends recalled memories as a context block and returns input unchanged when memory is disabled or no memories match.
• capture(text) stores a durable memory through MemWal and returns true on success, false when disabled or failed.

Memory is explicit so it is easy to see what prompt text changes. Configure it with the same credentials used by the SDK memory layer: delegate key, MemWalAccount id, embedding API key, MemWal package id, and relayer URL. Local onemem login credentials and env vars are both supported by the shared runtime resolver.

Scope (v0.1)

Trace capture is automatic for model.generate / model.stream. In repo-local source, memory recall/capture is shipped as the explicit createOneMemMemory() helper. Automatic memory extraction and per-tool-call interception remain tracked follow-ups.

Requires ai v6 (peer dependency). Spec: docs/05-our-architecture/04-frameworks/vercel-ai-provider.md.