@dmemo/sdk-wrappers

dMemo's inference leg for raw OpenAI/Anthropic SDK users: a preconfigured 0G Compute Router client factory, plus memory-injection wrappers built on each SDK's own native fetch/middleware extension points — no subclassing, no monkey-patching.

This package is dependency-light on purpose: it does not depend on @dmemo/core. It types against a small structural DmemoMemorySession interface, satisfied by @dmemo/core's DmemoSession public surface.

Install

npm install @dmemo/sdk-wrappers

Router client preset (T2.1)

import { createRouterClients, listPrivateModels } from '@dmemo/sdk-wrappers';

const { openai, anthropic } = createRouterClients({
  network: 'testnet',
  apiKey: process.env.ZEROG_API_KEY!, // a Router sk-... key, minted at pc.0g.ai
});

// Filters GET /v1/models for verifiability === "TeeML" (private inference).
const models = await listPrivateModels({ network: 'testnet' });

Every request carries X-0G-Provider-Trust-Mode: private (TeeML-only routing) by default — this is what makes "private inference" true. As of 2026-07-25 there is no TeeML chat model live on testnet (catalog drift — see docs/disclosure.md); listPrivateModels() will reflect whatever the Router currently advertises.

SDK memory wrappers (T2.2)

import OpenAI from 'openai';
import { createOpenAIMemoryFetch } from '@dmemo/sdk-wrappers';

const client = new OpenAI({
  baseURL: '<router-url>/v1',
  apiKey: 'sk-...',
  fetch: createOpenAIMemoryFetch(dmemoSession, { topK: 5 }),
});

The custom fetch injects session.search() results into the request body before the call, and writes the assistant's reply back via session.add()

• session.flush() after — handling both non-streaming responses and streamed responses (Stream.tee()) transparently.

For Anthropic/Claude traffic, createAnthropicMemoryMiddleware (built on the SDK's exported middleware extension point, v0.101.0+) is the preferred alternative to custom-fetch — it runs inside the retry loop and gives an independent stream copy via ctx.parse(response).

What this package does not do

• No retry/failover/provider pooling (the Router does this).
• No support for the deprecated @0glabs/0g-serving-broker.
• No /v1/responses support (Codex inference through 0G is out of scope — the Router has no such endpoint).