openclaw - 💡(How to fix) Fix Bug: memory-core sync fails with CredentialsProviderError despite OLLAMA_API_KEY set (distinct from #62282) [1 participants]

Bug Description

On every gateway startup and periodically during runtime, the memory-core plugin logs:

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
memory sync failed (watch): CredentialsProviderError: Could not load credentials from any providers

This is distinct from the more commonly reported Unknown memory embedding provider: ollama error (see #62282, #63681). In our case the ollama provider IS registered and visible (active-memory works using the same endpoint), but the memory-core sync path fails to obtain credentials.

Environment

• OpenClaw version: 2026.4.21 (npm), gateway running 2026.4.14+ (see uptime)
• Ollama: running at http://winhost:11434 (no auth, local)
• Memory embedding model: qwen3-embedding:0.6b (available in Ollama)
• memorySearch.provider: ollama
• OLLAMA_API_KEY: set to ollama-local as env var (satisfies any non-empty key check)
• OS: WSL2 / Linux

Steps to Reproduce

1. Configure memorySearch.provider = "ollama" and OLLAMA_API_KEY=ollama-local in environment
2. Start gateway or observe after restart
3. Error appears within seconds on session-start sync and periodically on search/bootstrap

Expected Behavior

Memory sync completes without credential errors. Ollama local instances require no API key.

Actual Behavior

CredentialsProviderError: Could not load credentials from any providers — the memory embedding provider is found but cannot produce credentials.

Root Cause Analysis

1. The ollama memory embedding provider IS registered and appears in active-memory logs as working
2. The failure occurs specifically in MemoryIndexManager.loadProviderResult() → createEmbeddingProvider() → getAdapter(provider) path used by memory-core sync
3. The Ollama memory embedding adapter's create() function calls buildApiKeyCredential(OLLAMA_PROVIDER_ID, result.credential, ...) where result.credential comes from resolveOllamaDiscoveryResult()
4. resolveOllamaDiscoveryResult() attempts to get credentials via the auth profile resolution chain
5. When the auth chain fails to resolve (even with OLLAMA_API_KEY set), it throws CredentialsProviderError

The key finding: OLLAMA_API_KEY=ollama-local is set in the process environment and active-memory works fine with the same Ollama endpoint, but memory-core sync fails. This suggests two different credential resolution paths:

• active-memory → uses ollama provider directly via api.runtime.chat() — succeeds
• memory-core sync → uses MemoryIndexManager which goes through createEmbeddingProvider() → ollamaMemoryEmbeddingProviderAdapter.create() → auth chain — fails

The difference may be that active-memory uses the OpenClaw inference plugin's Ollama integration (which bypasses auth profiles), while memory-core uses the dedicated memory embedding provider adapter which requires proper auth profile resolution.

Code Locations

• dist/memory-embedding-provider-runtime-hFl2REG1.js — getMemoryEmbeddingProvider() and listMemoryEmbeddingProviders()
• dist/manager-ChtEZeo0.js:168 — createEmbeddingProvider()
• dist/extensions/ollama/index.js:13-75 — ollamaMemoryEmbeddingProviderAdapter.create() with buildApiKeyCredential()
• dist/provider-auth-helpers-_LvigNEb.js:34 — resolveApiKeySecretInput()
• dist/dreaming-DA2EhSBZ.js:1465-1478 — reconcileManagedDreamingCron() where error surfaces

Existing Related Issues

• #62282 — Ollama embedding broke in 2026.4.5 (Unknown memory embedding provider: ollama)
• #63681 — CLI memory index fails with Unknown memory embedding provider: ollama (CLI vs gateway difference)
• #19943 — openclaw doctor false positive for memory embeddings
• #21894 — Memory search silently disabled with no user notification
• #66452 — Memory embedding provider normalization (mentioned in 2026.4.20 changelog)

Additional Context

The OLLAMA_API_KEY=ollama-local workaround (setting any non-empty value) was attempted but did not resolve the issue — the problem is not the API key value itself but the credential resolution chain failing to find/produce a valid credential object for the memory embedding provider adapter path.

Suggested Fix

1. Investigate why ollamaMemoryEmbeddingProviderAdapter.create() fails to resolve credentials when OLLAMA_API_KEY is set, while active-memory's Ollama calls succeed
2. Consider using the same Ollama invocation path for memory embeddings as active-memory uses
3. Alternatively: document that Ollama local instances require explicit auth profile setup for memory-core (separate from the OLLAMA_API_KEY env var)