Go ADK runtime: memory embedding fails with 401 invalid_issuer (OpenAI provider) — works on Python runtime with identical config

[djannot]

Summary

When a declarative Agent runs on the Go ADK runtime (runtime: go) with long-term memory enabled (OpenAI embedding provider), save_memory fails at embedding generation with HTTP 401 invalid_issuer. Nothing is persisted, so cross-session load_memory returns memories: null.

The exact same agent config on the Python runtime works — the embedding call to api.openai.com/v1/embeddings returns 200 and the memory is saved. Since the default declarative runtime is now go, memory-enabled agents silently stop persisting memory.

Environment

• kagent OSS v0.10.0-beta1
• Go runtime image: cr.kagent.dev/kagent-dev/kagent/golang-adk@sha256:0be1b7a0fe9560b097351cd97c2f71accf7f6f288fcdc0d5d6aae57e66d81b01
• Python runtime image: cr.kagent.dev/kagent-dev/kagent/app@sha256:ed96ed74b55eaceb92aec97dd84a0f523645c7372e501db9a019b637c7f67ee5
• Embedding ModelConfig: provider: OpenAI, model: text-embedding-3-small; key is a valid OpenAI service-account key (sk-svcacct-…)

Error (Go agent pod)

msg:   Tool execution failed   tool: save_memory   caller: agent/agent.go:408
error: failed to generate embedding: API returned status 401: {
  "error": {
    "message": "Your authentication token is not from a valid issuer.",
    "type": "invalid_request_error",
    "code": "invalid_issuer",
    "param": null
  },
  "status": 401
}

Root cause (narrowed)

The Go embedding path is go/adk/pkg/embedding/embedding.go → openAIProvider.generate(), which POSTs to <base_url | https://api.openai.com/v1>/embeddings with Authorization: Bearer os.Getenv("OPENAI_API_KEY") and dimensions: 768.

Everything that request depends on is valid — only the Go runtime fails:

Test	Result
Configured key + {"model":"text-embedding-3-small","dimensions":768} → OpenAI embeddings, from host	200
Same request from a pod in the agent's namespace (same egress)	200
Python runtime agent, identical config → embeddings	200, memory saved (httpx log: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK")
Chat completions with the same key (Go agent)	works
Go runtime agent → embeddings	401 invalid_issuer

The Go and Python pods are identical for the embedding auth path — same config.json (model: openai/…, base_url ""; memory.embedding: {provider: openai, model: text-embedding-3-small}), same OPENAI_API_KEY sourced from the same secret, and no base_url/proxy/OBO env in either. The only difference is the runtime image.

The non-OpenAI error body shape ("status": 401 alongside the error object; code: invalid_issuer) suggests the Go runtime is sending a different Authorization token (or hitting a different endpoint) than the configured OpenAI key against api.openai.com, rather than the raw OPENAI_API_KEY that works everywhere else.

Reproduction

1. Deploy a declarative Agent with runtime: go and memory enabled, using an OpenAI embedding ModelConfig (text-embedding-3-small).
2. Send a message that triggers save_memory, e.g. "My favorite color is blue. Please remember that."
3. Pod log shows save_memory … failed to generate embedding: API returned status 401 … invalid_issuer; the memory table stays empty.
4. Switch the same Agent to runtime: python and repeat → the embedding call returns 200 and a row is persisted.

Expected

The Go ADK memory embedding path should use the same working OpenAI credential/endpoint as chat completions (and as the Python runtime), so save_memory persists memories.

Workaround

Pin memory-enabled agents to runtime: python.

---

🤖 Filed with Claude Code