Summary

The memory-lancedb plugin currently restricts embedding configuration to OpenAI models only (text-embedding-3-small / text-embedding-3-large). The configSchema in openclaw.plugin.json blocks both baseUrl and arbitrary model names.

This prevents using local embedding providers like Ollama (which serves an OpenAI-compatible /v1/embeddings endpoint) or other self-hosted solutions.

Related

Issue #17650 requests adding gemini-embedding-001 and baseUrl to the schema. This request extends that to support any OpenAI-compatible embedding endpoint, including local ones.

Use Case

Many users run Ollama locally with embedding models like nomic-embed-text. Ollama exposes an OpenAI-compatible API at http://localhost:11434/v1. The memory-lancedb plugin's Embeddings class already uses the openai npm package, which accepts a baseURL constructor parameter — so the runtime code is nearly ready.

The memorySearch config in agents.defaults already supports custom baseUrl + arbitrary models for the core memory search. It would be consistent for memory-lancedb to support the same.

Proposed Changes

1. Add baseUrl to the embedding config schema in openclaw.plugin.json
2. Remove the enum restriction on model — or add an escape hatch for arbitrary model names. Different providers use different model IDs.
3. Pass baseUrl to the OpenAI client constructor in index.ts: new OpenAI({ apiKey, baseURL: baseUrl })
4. Make vector dimensions configurable (or auto-detect) since local models have varying dimensions (e.g., nomic-embed-text = 768).

Example Config (Ollama)

{
  "plugins": {
    "entries": {
      "memory-lancedb": {
        "enabled": true,
        "config": {
          "embedding": {
            "apiKey": "ollama",
            "model": "nomic-embed-text",
            "baseUrl": "http://localhost:11434/v1"
          },
          "autoCapture": true,
          "autoRecall": true
        }
      }
    },
    "slots": { "memory": "memory-lancedb" }
  }
}

Environment

• OpenClaw: 2026.2.17
• OS: Windows 11
• Ollama: nomic-embed-text on local network

This would make memory-lancedb provider-agnostic — consistent with how the rest of OpenClaw handles OpenAI-compatible endpoints.