Summary

When memorySearch.provider: "local" is configured and the CLI (openclaw memory index, openclaw memory status) correctly uses local embeddings, the memory_search tool still uses Gemini embeddings silently.

Environment

• OpenClaw: 2026.2.26
• OS: macOS 15.3 (Sequoia) arm64
• Gateway: local launchctl service

Config

agents: {
  defaults: {
    memorySearch: {
      provider: "local",
      fallback: "none",
      local: {
        modelPath: "/Users/brt-bot/.node-llama-cpp/models/embeddinggemma-300m.gguf"
      },
      query: { hybrid: { enabled: true } }
    }
  }
}

What works (CLI)

• openclaw memory status shows: Provider: local (requested: local)
• openclaw memory index --verbose uses local embeddings successfully (multiple "batch start" messages)
• Manual node-llama-cpp test produces 768-dim embeddings with Metal GPU

What fails (tool)

Calling memory_search returns results with provider: "gemini" and model: "gemini-embedding-001":

{
  "results": [...],
  "provider": "gemini",
  "model": "gemini-embedding-001",
  "mode": "hybrid"
}

This happens even with fallback: "none" set.

Expected behavior

memory_search tool should use local embeddings when memorySearch.provider: "local" is configured.

Actual behavior

The tool appears to bypass the local provider and use Gemini (likely auto-selected from resolved API key).

Difference from #8131

Issue #8131 was about auth errors when using local. That issue is closed. This is different: the tool works but silently uses the wrong provider.

Related

• Bug: memory_search/memory_get tools require OpenAI/Google API key even when memorySearch.provider=local (CLI index works) #8131 (closed) - auth errors with local provider
• memory_search tool ignores memory.backend=qmd, tries cloud embedding providers #11278 (closed) - qmd backend ignored