openclaw - 💡(How to fix) Fix Memory embeddings: provider=ollama ignored, calls Gemini instead [1 comments, 2 participants]

Bug Description

When agents.defaults.memorySearch.provider is set to ollama, the gateway ignores this setting and calls Gemini embedding API instead, using stale auth credentials from auth-profiles.json.

Reproduction

1. Config: agents.defaults.memorySearch.provider = "ollama" (valid adapter, registered in builtinMemoryEmbeddingProviderAdapters)
2. Session starts or search query runs
3. Gateway calls createEmbeddingProvider with correct provider ID
4. resolveMemorySearchConfig correctly merges config and returns provider: "ollama"
5. BUT: createEmbeddingProvider somehow routes to Gemini instead of Ollama adapter
6. Auth resolution pulls google:default profile from auth-profiles.json (stale key)
7. Error: API key not valid (Gemini endpoint)

Root Cause

Transport routing issue: when provider: "ollama" and transport: "remote":

• includeRemote = true (because primaryAdapter?.transport !== "local")
• Auth resolution diverges and calls Gemini auth ID instead of Ollama
• The auth-profiles lookup for ollama should map to Ollama credentials, not Google

Likely in: manager-*.js or memory-search-*.js where provider → auth resolution happens.

Environment

• OpenClaw v2026.3.28
• Node 22.22.0
• Ollama local (http://localhost:11434)
• Config: {"provider": "ollama", "remote": false, "local": {"modelPath": "nomic-embed-text"}}

Expected Behavior

When provider=ollama is set, use Ollama adapter. Auth resolution should not touch Gemini.

Actual Behavior

Gemini embedding provider is called with stale credentials, returning "API key not valid" error on every session-start or search query.

Workaround

Disable memory search entirely: agents.defaults.memorySearch.enabled = false