@moxxy/plugin-embeddings-transformers is the local embedding option. It runs @huggingface/transformers on the CPU — no API call, no key, no network — useful when you want vector recall without sending your memory contents to OpenAI.

Install

pnpm add @moxxy/plugin-embeddings-transformers @huggingface/transformers

The model weights download on first use (cached under ~/.cache/huggingface/).

Use

import { createTransformersEmbedder } from '@moxxy/plugin-embeddings-transformers';
import { buildMemoryPlugin } from '@moxxy/plugin-memory';
import { CachedEmbeddingProvider } from '@moxxy/sdk';

const embedder = new CachedEmbeddingProvider({
  inner: createTransformersEmbedder({
    model: 'Xenova/all-MiniLM-L6-v2', // default
  }),
  cacheDir: '~/.moxxy/memory/.embeddings',
});

const { plugin } = buildMemoryPlugin({ embedder });
session.pluginHost.registerStatic(plugin);

Exports

• TransformersEmbedder, createTransformersEmbedder
• TransformersEmbedderOptions
• PipelineFactory — escape hatch for tests / custom model loaders.

Trade-offs

• Pros: no API key, no network, no per-call cost, no data leaves your machine.
• Cons: first call is slow (downloads weights, warms a worker); embedding quality on the default MiniLM is below text-embedding-3-large.

For memory of ~hundreds of entries, MiniLM is fine. For thousands + nuanced semantic recall, OpenAI's models do better.