How to use a custom OpenAI-compatible endpoint to replace the official OpenAI API.

[scientiacoder]

🚀 The feature

Let's say we have an endpoint

curl --location 'https://llm-relay.group.cn/v1/chat/completions'
--header 'Authorization: Bearer sk-xxxxxxx'
--header 'Content-Type: application/json'
--data '{
"model": "gemini-3.1-pro-preview",
"messages": [
{
"role": "system",
"content": "string"
},
{
"role": "user",
"content": "hello"
}
]
}'

and it returns
{
"id": "chatcmpl-24b8a5994d2135c87ef73b5ef60a3d10",
"model": "gemini-3.1-pro-preview",
"object": "chat.completion",
"created": 1773219707,
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": "Hello! How can I help you today?"
},
"finish_reason": "stop"
}
],
"usage": {
"prompt_tokens": 2,
"completion_tokens": 344,
"total_tokens": 346,
"prompt_tokens_details": {
"cached_tokens": 0,
"text_tokens": 2,
"audio_tokens": 0,
"image_tokens": 0
},
"completion_tokens_details": {
"text_tokens": 0,
"audio_tokens": 0,
"reasoning_tokens": 335
},
"input_tokens": 0,
"output_tokens": 0,
"input_tokens_details": null,
"claude_cache_creation_5_m_tokens": 0,
"claude_cache_creation_1_h_tokens": 0
}
}

How to use this customized custom OpenAI-compatible endpoint to REPLACE the OFFICIAL OPEN-AI key since I don't have, thanks.

Motivation, pitch

It is convenient for us to use OpenAI-compatible endpoint instead of official OpenAI

[slideshow-dingo]

You can already do this with the openai provider — just set baseURL and apiKey in the config. Here's a working example using OpenRouter (or any OpenAI-compatible endpoint):

TypeScript SDK (v2.x):

import { Memory } from "mem0ai/oss";

const config = {
  embedder: {
    provider: "ollama",
    config: {
      model: "nomic-embed-text:latest",
      url: "http://localhost:11434",
      embeddingDims: 768,
    }
  },
  llm: {
    provider: "openai",
    config: {
      model: "google/gemini-2.5-flash",
      apiKey: "sk-or-v1-your-openrouter-key",
      baseURL: "https://openrouter.ai/api/v1",
      temperature: 0.2,
      maxTokens: 2000,
    }
  },
  vectorStore: {
    provider: "qdrant",
    config: {
      host: "localhost",
      port: 6333,
      collectionName: "mem0-memories",
    }
  }
};

const memory = new Memory(config);

Python SDK:

from mem0 import Memory

config = {
    "llm": {
        "provider": "openai",
        "config": {
            "model": "google/gemini-2.5-flash",
            "api_key": "sk-or-v1-your-openrouter-key",
            "openai_base_url": "https://openrouter.ai/api/v1",
        }
    }
}

m = Memory.from_config(config)

⚠️ Common gotchas:

1. Key name differences between TypeScript and Python SDKs:

   • TypeScript: apiKey + baseURL (camelCase)
   • Python: api_key + openai_base_url (snake_case with openai_ prefix)

   Mixing these up results in the OpenAI client receiving undefined/null for the key, which silently falls back to OpenAI's default API — causing 401 errors if no OPENAI_API_KEY env var is set.

2. Stale config keys cause silent failures: The OpenAILLM constructor reads config.apiKey and config.baseURL directly. If your config has openaiApiKey, baseUrl, or url instead, they are silently ignored.

3. Embedder is separate from LLM: The embedder provider is independent. For OpenRouter (LLM) + local Ollama (embeddings), you need both sections configured. OpenRouter does not support the embeddings API, so you can't use it as an embedder.

4. Gateway restart required for config changes: If using mem0 via a plugin wrapper (like OpenClaw), config changes to the mem0 plugin often require a full gateway restart — hot reload doesn't pick up the changes.

If your endpoint is OpenAI-compatible, it should work exactly like the above — just replace the baseURL with your custom endpoint URL and use your API key.

Test with:

# Verify it works
node -e "const m = require('mem0ai/oss').Memory; /* your config */"

For more context, the relevant source that confirms this: OpenAILLM in dist/oss/index.js does:

this.openai = new OpenAI({ apiKey: config.apiKey, baseURL: config.baseURL });

[slideshow-dingo]

I've submitted a PR that adds config validation to catch these issues early: #4378

The PR adds validation to both and classes that:

1. Throws a clear error if is missing
2. Warns if users provide , , or instead of the correct / keys

This should prevent the silent failures described in the gotchas section. The fix is in and .

[slideshow-dingo]

I've submitted a PR that adds config validation to catch wrong keys (openaiApiKey, baseUrl, url) that should be (apiKey, baseURL) in the TypeScript SDK. This prevents silent failures and 401 errors.

PR: #4379

The fix validates that 'apiKey' is present and throws clear error messages when wrong keys are detected. It also warns users if they use Python SDK-style keys (snake_case) instead of TypeScript SDK-style keys (camelCase).

[kartik-mem0]

You can already point the openai provider at a custom endpoint by setting baseURL and apiKey in the config. We should make this clearer in the docs though.

[kartik-mem0]

This is already supported. Set the openai provider with openai_base_url and api_key in your config:

from mem0 import Memory

config = {
    "llm": {
        "provider": "openai",
        "config": {
            "model": "gemini-3.1-pro-preview",
            "api_key": "sk-xxxxxxx",
            "openai_base_url": "https://llm-relay.group.cn/v1",
        }
    }
}

m = Memory.from_config(config)

Docs: https://docs.mem0.ai/components/llms/config (see openai_base_url in the config reference)