fix(llms): add is_reasoning_model config override for versioned deployments (#5327)

Co-authored-by: kartik-mem0 <kartik.labhshetwar@mem0.ai>

mem0/configs/llms/azure.py

@@ -23,6 +23,7 @@ class AzureOpenAIConfig(BaseLlmConfig):
         vision_details: Optional[str] = "auto",
         reasoning_effort: Optional[str] = None,
         http_client_proxies: Optional[dict] = None,
+        is_reasoning_model: Optional[bool] = None,
         # Azure OpenAI-specific parameters
         azure_kwargs: Optional[Dict[str, Any]] = None,
     ):
@@ -40,6 +41,10 @@ class AzureOpenAIConfig(BaseLlmConfig):
             vision_details: Vision detail level, defaults to "auto"
             reasoning_effort: Effort level for reasoning models ("low", "medium", "high"), defaults to None
             http_client_proxies: HTTP client proxy settings, defaults to None
+            is_reasoning_model: Explicit override for reasoning-model detection.
+                None (default) uses the name-based heuristic. Set True to drop
+                max_tokens/temperature (e.g. for versioned Azure deployments like
+                "gpt-5.4-nano-2026-03-17"), or False to force standard params.
             azure_kwargs: Azure-specific configuration, defaults to None
         """
         # Initialize base parameters
@@ -54,6 +59,7 @@ class AzureOpenAIConfig(BaseLlmConfig):
             vision_details=vision_details,
             reasoning_effort=reasoning_effort,
             http_client_proxies=http_client_proxies,
+            is_reasoning_model=is_reasoning_model,
         )
 
         # Azure OpenAI-specific parameters

mem0/configs/llms/base.py

@@ -25,6 +25,7 @@ class BaseLlmConfig(ABC):
         vision_details: Optional[str] = "auto",
         reasoning_effort: Optional[str] = None,
         http_client_proxies: Optional[Union[Dict, str]] = None,
+        is_reasoning_model: Optional[bool] = None,
     ):
         """
         Initialize a base configuration class instance for the LLM.
@@ -54,6 +55,14 @@ class BaseLlmConfig(ABC):
                 Defaults to None (uses the model's default reasoning effort)
             http_client_proxies: Proxy settings for HTTP client.
                 Can be a dict or string. Defaults to None
+            is_reasoning_model: Explicit override for reasoning-model detection.
+                When None (default), the model is classified automatically from its
+                name (preserving existing behavior). Set to True to force the
+                reasoning-model parameter set (drop max_tokens and temperature),
+                or False to force the standard parameter set. Useful for
+                deployments with custom/versioned model names (e.g. Azure
+                "gpt-5.4-nano-2026-03-17") that the name-based heuristic cannot
+                recognize. Defaults to None
         """
         self.model = model
         self.temperature = temperature
@@ -64,4 +73,5 @@ class BaseLlmConfig(ABC):
         self.enable_vision = enable_vision
         self.vision_details = vision_details
         self.reasoning_effort = reasoning_effort
+        self.is_reasoning_model = is_reasoning_model
         self.http_client = httpx.Client(proxies=http_client_proxies) if http_client_proxies else None

mem0/configs/llms/openai.py

@@ -22,6 +22,7 @@ class OpenAIConfig(BaseLlmConfig):
         vision_details: Optional[str] = "auto",
         reasoning_effort: Optional[str] = None,
         http_client_proxies: Optional[dict] = None,
+        is_reasoning_model: Optional[bool] = None,
         # OpenAI-specific parameters
         openai_base_url: Optional[str] = None,
         models: Optional[List[str]] = None,
@@ -47,6 +48,9 @@ class OpenAIConfig(BaseLlmConfig):
             vision_details: Vision detail level, defaults to "auto"
             reasoning_effort: Effort level for reasoning models ("low", "medium", "high"), defaults to None
             http_client_proxies: HTTP client proxy settings, defaults to None
+            is_reasoning_model: Explicit override for reasoning-model detection.
+                None (default) uses the name-based heuristic. Set True to drop
+                max_tokens/temperature, or False to force standard params.
             openai_base_url: OpenAI API base URL, defaults to None
             models: List of models for OpenRouter, defaults to None
             route: OpenRouter route strategy, defaults to "fallback"
@@ -72,6 +76,7 @@ class OpenAIConfig(BaseLlmConfig):
             vision_details=vision_details,
             reasoning_effort=reasoning_effort,
             http_client_proxies=http_client_proxies,
+            is_reasoning_model=is_reasoning_model,
         )
 
         # OpenAI-specific parameters

mem0/llms/azure_openai.py

@@ -33,6 +33,7 @@ class AzureOpenAILLM(LLMBase):
                 vision_details=config.vision_details,
                 reasoning_effort=getattr(config, 'reasoning_effort', None),
                 http_client_proxies=config.http_client,
+                is_reasoning_model=getattr(config, 'is_reasoning_model', None),
             )
 
         super().__init__(config)

mem0/llms/base.py

@@ -43,13 +43,23 @@ class LLMBase(ABC):
     def _is_reasoning_model(self, model: str) -> bool:
         """
         Check if the model is a reasoning model or GPT-5 series that doesn't support certain parameters.
-        
+
+        An explicit ``is_reasoning_model`` on the config takes precedence over the
+        name-based heuristic. This lets deployments with custom/versioned model
+        names (e.g. Azure ``gpt-5.4-nano-2026-03-17``) opt in or out without
+        relying on string matching. When the config value is ``None`` (default),
+        classification falls back to the name-based heuristic below.
+
         Args:
             model: The model name to check
-            
+
         Returns:
             bool: True if the model is a reasoning model or GPT-5 series
         """
+        explicit = getattr(self.config, "is_reasoning_model", None)
+        if explicit is not None:
+            return explicit
+
         reasoning_models = {
             "o1", "o1-preview", "o3-mini", "o3",
             "gpt-5", "gpt-5o", "gpt-5o-mini", "gpt-5o-micro",

mem0/llms/openai.py

@@ -31,6 +31,7 @@ class OpenAILLM(LLMBase):
                 vision_details=config.vision_details,
                 reasoning_effort=getattr(config, 'reasoning_effort', None),
                 http_client_proxies=config.http_client,
+                is_reasoning_model=getattr(config, 'is_reasoning_model', None),
             )
 
         super().__init__(config)