Where the default embedding is placed by default? Which path on MacOS?

[XQZmeSIR]

Where the default embedding is placed by default? Which path on MacOS?
I want to change it to multilingual one

And is it possible to use ONNX models like

Xenova/multilingual-e5-small ?

[vectornguyen76]

Assuming you mean the local plugin (apps/memos-local-plugin, the one using Transformers.js / Xenova/... models): on current versions the model files are cached in ~/.cache/memos/models on macOS too, unless you set embedding.local.cacheDir. The default model is Xenova/all-MiniLM-L6-v2.

The model is set in config.yaml, which lives in ~/.openclaw/memos-plugin/, ~/.hermes/memos-plugin/ or ~/.dsh/memos-plugin/ depending on the agent (or wherever MEMOS_HOME points). The Viewer's Settings page exposes the same fields. Xenova/multilingual-e5-small should work: the provider loads the model with dtype: "q8", and that repo ships onnx/model_quantized.onnx.

embedding:
  provider: local
  model: Xenova/multilingual-e5-small
  local:
    cacheDir: ""      # empty = ~/.cache/memos/models
    offline: false

Two things to watch:

• e5 models are trained with query: / passage: prefixes. The local provider ignores the query/document role and sends the raw text, so retrieval will work but may be a bit weaker than e5's published numbers.
• Both models are 384-dim, so nothing flags existing memories as mismatched. Their old MiniLM vectors stay as they are until you click Rebuild all vectors in Settings (the "repair" mode only fixes missing or wrong-dimension vectors). Run the rebuild after you switch models.

Code: defaults, cache dir + load, README config, rebuild endpoint.