openclaw - 💡(How to fix) Fix memory_search tool can fail with CredentialsProviderError while CLI works after switching to local embeddings [1 participants]

Summary

In a real local install, the built-in memory_search tool kept failing with:

Could not load credentials from any providers

while the underlying memory store was present and openclaw memory status --deep reported a configured provider. I was able to repair the CLI/runtime by switching agents.defaults.memorySearch.provider to local and installing node-llama-cpp, but the built-in tool in the live agent session still continued returning the old credential error.

That leaves two likely issues:

1. the live memory_search tool host is holding stale provider state / stale manager state after config changes and gateway restart, and/or
2. the tool path is not honoring the same provider/runtime initialization path that openclaw memory status / openclaw memory search now use successfully.

This looks related to stale-provider reports like #61098, but the symptom here is a credential-provider failure instead of continuing to call the previous remote backend explicitly.

Environment

• OpenClaw: 2026.4.21
• macOS arm64
• Node: v25.6.1
• Global install via npm
• Workspace memory backend: builtin / memory-core

Initial failing state

Config included:

"agents": {
  "defaults": {
    "memorySearch": {
      "enabled": true,
      "fallback": "none",
      "model": "text-embedding-3-small"
    }
  }
}

openclaw memory status --json --deep showed:

"provider": "bedrock",
"requestedProvider": "auto",
"embeddingProbe": {
  "ok": false,
  "error": "Could not load credentials from any providers"
}

and openclaw memory search --query "Mohammad" failed with:

Memory search failed: Could not load credentials from any providers
[memory] sync failed (session-start): CredentialsProviderError: Could not load credentials from any providers
[memory] sync failed (search): CredentialsProviderError: Could not load credentials from any providers

No AWS/Bedrock/OpenAI env vars were set in the shell.

What fixed the CLI/runtime

I changed config to:

"agents": {
  "defaults": {
    "memorySearch": {
      "provider": "local",
      "model": "hf:ggml-org/embeddinggemma-300m-qat-q8_0-GGUF/embeddinggemma-300m-qat-Q8_0.gguf",
      "fallback": "none",
      "enabled": true
    }
  }
}

Then restarted the gateway.

At first local embeddings failed because node-llama-cpp was missing:

Cannot find package 'node-llama-cpp' imported from .../engine-embeddings-*.js

Installing it globally fixed that:

npm i -g node-llama-cpp

After that:

openclaw memory status --json --deep

reported:

"provider": "local",
"requestedProvider": "local",
"embeddingProbe": { "ok": true }

and the model downloaded successfully and initialized.

Remaining bug

After the fix above, the CLI path works, but the built-in memory_search tool in the live agent session still returns:

{
  "disabled": true,
  "unavailable": true,
  "error": "Could not load credentials from any providers"
}

So the tool/runtime still appears stuck on the old credential-failing provider state even though:

• config on disk is updated
• gateway has restarted
• CLI deep status now probes local embeddings successfully

Why this matters

This makes memory recovery confusing:

• users repair the config
• the CLI proves embeddings are healthy
• but the tool the agent must use still reports the old credential error

So the user sees "memory is still broken" even after a valid repair.

Suspected root cause

Likely one of:

1. stale singleton / manager cache in the live tool host
2. the built-in tool captures provider config too early and doesn't re-resolve after restart/config change
3. the tool path is separate from CLI path and still auto-selects a remote provider when it should now be using local

Related issues

• #61098 — memory_search uses stale embedding provider after config.patch + restart
• #55608 — CLI/tool config divergence in other memory paths
• #47251 / #57390 — local memory embeddings break if node-llama-cpp is not present

Suggested fixes

1. Make the built-in memory_search tool re-resolve provider/manager state after restart instead of using stale cached state.
2. Ensure tool and CLI share the same provider resolution path and current config source.
3. If provider changes from remote->local, clear any stale embedding manager cache automatically.
4. Improve the tool error so it distinguishes:
   • stale cached provider state
   • missing local dependency (node-llama-cpp)
   • real remote credential failure