Summary

The memory-lancedb plugin previously required an OpenAI API key and only talked to OpenAI’s embeddings API. This PR adds support for Ollama and any other service that exposes the OpenAI-compatible /v1/embeddings API (e.g. local or self-hosted endpoints), so you can run long-term memory without an OpenAI key.

Scope

• Config (config.ts): embedding.apiKey is now optional when embedding.baseUrl is set (e.g. Ollama). Added embedding.baseUrl (default https://api.openai.com/v1) and embedding.dimensions (for models not in the built-in list). Built-in dimensions extended with nomic-embed-text (768) and mxbai-embed-large (1024).
• Embeddings (index.ts): Replaced the hardcoded openai client with a small fetch-based client that uses baseUrl and only sends Authorization when apiKey is set.
• Tests (index.test.ts): Updated “missing apiKey” assertion to the new error message; added test that Ollama-style config (baseUrl + model, no apiKey) parses correctly.
• Docs: Added OLLAMA-SUPPORT.md describing the change and how to use the plugin with Ollama.

User-facing changes

• New config options: plugins.entries.memory-lancedb.config.embedding.baseUrl (e.g. http://localhost:11434/v1 for Ollama), embedding.dimensions (required for unknown models).
• Optional API key: When baseUrl is set to a local/OpenAI-compatible endpoint (e.g. Ollama), embedding.apiKey can be omitted.
• Existing OpenAI usage: Unchanged. Omit baseUrl (or leave default) and set embedding.apiKey as before.

Testing

• pnpm test -- extensions/memory-lancedb/index.test.ts — all tests pass (including new Ollama config test).
• Lint/format: pnpm check (no new issues).

Example (Ollama)

{
  "plugins": {
    "entries": {
      "memory-lancedb": {
        "config": {
          "embedding": {
            "baseUrl": "http://localhost:11434/v1",
            "model": "nomic-embed-text"
          },
          "dbPath": "~/.openclaw/memory/lancedb",
          "autoCapture": true,
          "autoRecall": true
        }
      }
    },
    "slots": { "memory": "memory-lancedb" }
  }
}

No apiKey required; run ollama pull nomic-embed-text and ensure Ollama is serving.

Greptile Summary

Adds support for Ollama and OpenAI-compatible embedding endpoints to the memory-lancedb plugin, allowing users to run long-term memory without an OpenAI API key. The implementation replaces the hardcoded OpenAI client with a generic fetch-based client that conditionally sends authorization headers when an API key is provided.

Changes made:

• config.ts: Made embedding.apiKey optional when embedding.baseUrl is set, added embedding.baseUrl and embedding.dimensions config options, extended built-in dimension mappings for Ollama models (nomic-embed-text, mxbai-embed-large)
• index.ts: Replaced OpenAI SDK with lightweight fetch implementation that works with any OpenAI-compatible /v1/embeddings endpoint
• index.test.ts: Updated test assertions to match new error messages and added validation for Ollama configuration

Issues found:

• The openai package dependency is still present in package.json but is no longer imported or used in the code

Confidence Score: 4/5

• Safe to merge with minor cleanup recommended
• The implementation is solid with proper validation, security considerations (UUID validation, prompt injection filtering), and comprehensive test coverage. The fetch-based approach correctly handles OpenAI-compatible endpoints. One non-critical improvement: removing the unused openai dependency from package.json.
• Check extensions/memory-lancedb/package.json to remove unused dependency

_{Last reviewed commit: 8baa707}