Running OpenClaw 2026.4.12 on Ubuntu, systemd install. openclaw memory index --force fails with Open

[little abyss]

Running OpenClaw 2026.4.12 on Ubuntu, systemd install. openclaw memory index --force fails with OpenAI embeddings 401 invalid_api_key. Using a sk-proj- key with All permissions, active project, $25 credit balance. Key resolves correctly — openclaw models status --probe shows openai:default as ok for inference. openclaw secrets audit --check shows unresolved=0. OPENAI_PROJECT_ID is set in .env. No invisible characters in the key. Error is specific to the embeddings path — chat completions work fine. memorySearch.provider is set to openai. Is there a known issue with sk-proj- keys and the embeddings API in 2026.4.12?

[iron goblet]

Just checking the local docs path, then I’ll verify the memory/embeddings code path.

I don’t see a documented “sk-proj keys are broken” note for 2026.4.12 specifically, but I do think you’ve hit a real embeddings-path bug/limitation.

What I verified in source:

• the normal OpenAI inference path uses the OpenAI SDK
• the memory embeddings path uses a separate custom HTTP client
• that embeddings client only sends Authorization + Content-Type
• I do not see OPENAI_PROJECT_ID / OpenAI-Project being forwarded there

So this failure pattern makes sense:

• openclaw models status --probe passes, because inference works
• openclaw memory index --force fails, because embeddings are hitting /v1/embeddings without the extra project header your project-scoped key may require

Workaround to try in openclaw.json:

{
  agents: {
    defaults: {
      memorySearch: {
        provider: "openai",
        model: "text-embedding-3-small",
        remote: {
          headers: {
            "OpenAI-Project": "proj_..."
          }
        }
      }
    }
  }
}

If your account also needs org scoping, add:

"OpenAI-Organization": "org_..."

Then restart and retry:

openclaw gateway restart
openclaw memory index --force

Relevant docs:

• https://docs.openclaw.ai/reference/memory-config
• https://docs.openclaw.ai/providers/openai

So, short version: not obviously a general OpenAI/sk-proj outage, but very plausibly a 2026.4.12 OpenClaw memory-embeddings auth/header mismatch.

If adding OpenAI-Project fixes it, please file a GitHub issue, because that header should probably be propagated automatically: https://github.com/openclaw/openclaw/issues

Glad that helped! 🎉
If this fixed it, please right-click the solution message → Apps → ✅ Mark Solution.
This helps Answer Overflow index the solution for future users.

[little abyss]

adding OpenAI-Project header via remote.headers did not resolve the 401. Ask if there's another workaround or if this needs a bug fix in a future release.

[iron goblet]

Let me check the exact embeddings request path a bit deeper, there may be another place the OpenAI auth behavior diverges.

Yeah, there’s one more good workaround to try before calling it a release bug.

What I found: the memory embeddings path resolves provider creds profile-first, not env-first. So if you have any stored openai auth profile, memory indexing can pick that instead of OPENAI_API_KEY. In that case, adding OpenAI-Project won’t help because it may still be using the wrong credential.

I’d try pinning the memory path completely:

{
  agents: {
    defaults: {
      memorySearch: {
        provider: "openai",
        model: "text-embedding-3-small",
        remote: {
          baseUrl: "https://api.openai.com/v1",
          apiKey: "sk-proj-...",

          headers: {
            "OpenAI-Project": "proj_..."
          },
          batch: {
            enabled: false
          }
        }
      }
    }
  }
}

Why this helps:

• remote.apiKey bypasses normal provider auth/profile resolution for memory embeddings
• explicit baseUrl removes any provider/baseUrl weirdness