mdb-engine 0.1.6__py3-none-any.whl → 0.4.12__py3-none-any.whl

mdb_engine/memory/service.py

@@ -1,43 +1,29 @@
 """
 Mem0 Memory Service Implementation
-
-This module provides a wrapper around Mem0.ai for intelligent memory management.
-It integrates seamlessly with mdb-engine's MongoDB connection.
-mem0 handles embeddings and LLM via environment variables (.env).
+Production-ready wrapper for Mem0.ai with strict metadata schema for MongoDB.
 """
 
 import logging
 import os
 import tempfile
-from typing import Any, Dict, List, Optional, Union
+from typing import Any
 
 # Set MEM0_DIR environment variable early to avoid permission issues
-# mem0 tries to create .mem0 directory at import time, so we set this before any import
 if "MEM0_DIR" not in os.environ:
-    # Use /tmp/.mem0 which should be writable in most environments
     mem0_dir = os.path.join(tempfile.gettempdir(), ".mem0")
     try:
         os.makedirs(mem0_dir, exist_ok=True)
         os.environ["MEM0_DIR"] = mem0_dir
     except OSError:
-        # Fallback: try user's home directory
-        try:
-            home_dir = os.path.expanduser("~")
-            mem0_dir = os.path.join(home_dir, ".mem0")
-            os.makedirs(mem0_dir, exist_ok=True)
-            os.environ["MEM0_DIR"] = mem0_dir
-        except OSError:
-            # Last resort: current directory (may fail but won't crash import)
-            os.environ["MEM0_DIR"] = os.path.join(os.getcwd(), ".mem0")
+        # Fallback: current directory
+        os.environ["MEM0_DIR"] = os.path.join(os.getcwd(), ".mem0")
 
-# Try to import mem0 (optional dependency)
-# Import is lazy to avoid permission issues at module load time
+# Lazy Import
 MEM0_AVAILABLE = None
 Memory = None
 
 
 def _check_mem0_available():
-    """Lazy check if mem0 is available."""
     global MEM0_AVAILABLE, Memory
     if MEM0_AVAILABLE is None:
         try:
@@ -47,1239 +33,448 @@ def _check_mem0_available():
         except ImportError:
             MEM0_AVAILABLE = False
             Memory = None
-        except OSError as e:
-            logger.warning(
-                f"Failed to set up mem0 directory: {e}. Memory features may be limited."
-            )
-            MEM0_AVAILABLE = False
-            Memory = None
-
     return MEM0_AVAILABLE
 
 
 logger = logging.getLogger(__name__)
 
 
-def _detect_provider_from_env() -> str:
-    """
-    Detect provider from environment variables.
-
-    Returns:
-        "azure" if Azure OpenAI credentials are present, otherwise "openai"
-    """
-    if os.getenv("AZURE_OPENAI_API_KEY") and os.getenv("AZURE_OPENAI_ENDPOINT"):
-        return "azure"
-    elif os.getenv("OPENAI_API_KEY"):
-        return "openai"
-    else:
-        # Default to openai if nothing is configured
-        return "openai"
-
-
-def _detect_embedding_dimensions(model_name: str) -> Optional[int]:
-    """
-    Auto-detect embedding dimensions from model name.
-
-    Args:
-        model_name: Embedding model name (e.g., "text-embedding-3-small")
-
-    Returns:
-        Number of dimensions, or None if unknown (should use config/default)
-
-    Examples:
-        >>> _detect_embedding_dimensions("text-embedding-3-small")
-        1536
-    """
-    # Normalize model name (remove provider prefix)
-    normalized = model_name.lower()
-    if "/" in normalized:
-        normalized = normalized.split("/", 1)[1]
-
-    # OpenAI models
-    if "text-embedding-3-small" in normalized:
-        return 1536
-    elif "text-embedding-3-large" in normalized:
-        return 3072
-    elif "text-embedding-ada-002" in normalized or "ada-002" in normalized:
-        return 1536
-    elif "text-embedding-ada" in normalized:
-        return 1536
-
-    # Cohere models (common ones)
-    if "embed-english-v3" in normalized:
-        return 1024
-    elif "embed-multilingual-v3" in normalized:
-        return 1024
-
-    # Unknown model - return None to use config/default
-    return None
-
-
 class Mem0MemoryServiceError(Exception):
-    """
-    Base exception for all Mem0 Memory Service failures.
-    """
-
     pass
 
 
-def _build_vector_store_config(
-    db_name: str, collection_name: str, mongo_uri: str, embedding_model_dims: int
-) -> Dict[str, Any]:
-    """Build vector store configuration for mem0."""
-    return {
-        "vector_store": {
-            "provider": "mongodb",
-            "config": {
-                "db_name": db_name,
-                "collection_name": collection_name,
-                "mongo_uri": mongo_uri,
-                "embedding_model_dims": embedding_model_dims,
-            },
-        }
-    }
-
-
-def _build_embedder_config(
-    provider: str, embedding_model: str, app_slug: str
-) -> Dict[str, Any]:
-    """Build embedder configuration for mem0."""
-    clean_embedding_model = embedding_model.replace("azure/", "").replace("openai/", "")
-    if provider == "azure":
-        azure_endpoint = os.getenv("AZURE_OPENAI_ENDPOINT")
-        azure_api_key = os.getenv("AZURE_OPENAI_API_KEY")
-        azure_api_version = os.getenv(
-            "AZURE_OPENAI_API_VERSION",
-            os.getenv("OPENAI_API_VERSION", "2024-02-15-preview"),
-        )
-
-        if not azure_endpoint or not azure_api_key:
-            raise Mem0MemoryServiceError(
-                "Azure OpenAI requires AZURE_OPENAI_ENDPOINT and "
-                "AZURE_OPENAI_API_KEY environment variables"
-            )
-
-        config = {
-            "provider": "azure_openai",
-            "config": {
-                "model": clean_embedding_model,
-                "azure_kwargs": {
-                    "azure_deployment": clean_embedding_model,
-                    "api_version": azure_api_version,
-                    "azure_endpoint": azure_endpoint,
-                    "api_key": azure_api_key,
-                },
-            },
-        }
-    else:
-        config = {
-            "provider": "openai",
-            "config": {"model": clean_embedding_model},
-        }
-
-    provider_name = "Azure OpenAI" if provider == "azure" else "OpenAI"
-    logger.info(
-        f"Configuring mem0 embedder ({provider_name}): "
-        f"provider='{config['provider']}', "
-        f"model='{clean_embedding_model}'",
-        extra={
-            "app_slug": app_slug,
-            "embedding_model": embedding_model,
-            "embedder_provider": config["provider"],
-            "provider": provider,
-        },
-    )
-    return config
-
-
-def _build_llm_config(
-    provider: str, chat_model: str, temperature: float, app_slug: str
-) -> Dict[str, Any]:
-    """Build LLM configuration for mem0."""
-    clean_chat_model = chat_model.replace("azure/", "").replace("openai/", "")
-    if provider == "azure":
-        deployment_name = os.getenv("AZURE_OPENAI_DEPLOYMENT_NAME") or clean_chat_model
-        clean_chat_model = deployment_name
-
-        azure_endpoint = os.getenv("AZURE_OPENAI_ENDPOINT")
-        azure_api_key = os.getenv("AZURE_OPENAI_API_KEY")
-        azure_api_version = os.getenv(
-            "AZURE_OPENAI_API_VERSION",
-            os.getenv("OPENAI_API_VERSION", "2024-02-15-preview"),
-        )
-
-        if not azure_endpoint or not azure_api_key:
-            raise Mem0MemoryServiceError(
-                "Azure OpenAI LLM requires AZURE_OPENAI_ENDPOINT and "
-                "AZURE_OPENAI_API_KEY environment variables"
-            )
-
-        config = {
-            "provider": "azure_openai",
-            "config": {
-                "model": clean_chat_model,
-                "temperature": temperature,
-                "azure_kwargs": {
-                    "azure_deployment": clean_chat_model,
-                    "api_version": azure_api_version,
-                    "azure_endpoint": azure_endpoint,
-                    "api_key": azure_api_key,
-                },
-            },
-        }
-    else:
-        config = {
-            "provider": "openai",
-            "config": {"model": clean_chat_model, "temperature": temperature},
-        }
-
-    llm_provider_name = "Azure OpenAI" if provider == "azure" else "OpenAI"
-    logger.info(
-        f"Configuring mem0 LLM ({llm_provider_name}): "
-        f"provider='{config['provider']}', "
-        f"model='{clean_chat_model}'",
-        extra={
-            "app_slug": app_slug,
-            "original_model": chat_model,
-            "llm_provider": config["provider"],
-            "llm_provider_type": provider,
-            "temperature": temperature,
-        },
-    )
-    return config
-
-
-def _initialize_memory_instance(mem0_config: Dict[str, Any], app_slug: str) -> tuple:
-    """Initialize Mem0 Memory instance and return (instance, init_method)."""
-    logger.debug(
-        "Initializing Mem0 Memory with config structure",
-        extra={
-            "app_slug": app_slug,
-            "config_keys": list(mem0_config.keys()),
-            "vector_store_provider": mem0_config.get("vector_store", {}).get(
-                "provider"
-            ),
-            "embedder_provider": mem0_config.get("embedder", {}).get("provider"),
-            "llm_provider": (
-                mem0_config.get("llm", {}).get("provider")
-                if mem0_config.get("llm")
-                else None
-            ),
-            "full_config": mem0_config,
-        },
-    )
-
-    init_method = None
-    try:
-        if hasattr(Memory, "from_config"):
-            memory_instance = Memory.from_config(mem0_config)
-            init_method = "Memory.from_config()"
-        else:
-            try:
-                from mem0.config import Config
-
-                config_obj = Config(**mem0_config)
-                memory_instance = Memory(config_obj)
-                init_method = "Memory(Config())"
-            except (ImportError, TypeError) as config_error:
-                logger.warning(
-                    f"Could not create Config object, trying dict: {config_error}",
-                    extra={"app_slug": app_slug},
-                )
-                memory_instance = Memory(mem0_config)
-                init_method = "Memory(dict)"
-    except (
-        ImportError,
-        AttributeError,
-        TypeError,
-        ValueError,
-        RuntimeError,
-        KeyError,
-    ) as init_error:
-        error_msg = str(init_error)
-        logger.error(
-            f"Failed to initialize Memory instance: {error_msg}",
-            exc_info=True,
-            extra={
-                "app_slug": app_slug,
-                "error": error_msg,
-                "error_type": type(init_error).__name__,
-                "config_keys": (
-                    list(mem0_config.keys())
-                    if isinstance(mem0_config, dict)
-                    else "not_dict"
-                ),
-            },
-        )
-        raise Mem0MemoryServiceError(
-            f"Failed to initialize Memory instance: {error_msg}. "
-            f"Ensure mem0ai is installed and Azure OpenAI environment "
-            f"variables are set correctly."
-        ) from init_error
-
-    return memory_instance, init_method
-
-
 class Mem0MemoryService:
-    """
-    Service for managing user memories using Mem0.ai.
-
-    This service provides intelligent memory management that:
-    - Stores and retrieves memories in MongoDB (using mdb-engine's connection)
-    - Uses mem0's embedder for embeddings (configured via environment variables)
-    - Optionally extracts memories from conversations (requires LLM if infer: true)
-    - Retrieves relevant memories for context-aware responses
-    - Optionally builds knowledge graphs for entity relationships
-
-    Embeddings and LLM are configured via environment variables (.env) and mem0 handles
-    provider routing automatically.
-    """
-
     def __init__(
         self,
         mongo_uri: str,
         db_name: str,
         app_slug: str,
-        config: Optional[Dict[str, Any]] = None,
+        config: dict[str, Any] | None = None,
     ):
-        """
-        Initialize Mem0 Memory Service.
-
-        Args:
-            mongo_uri: MongoDB connection URI
-            db_name: Database name
-            app_slug: App slug (used for collection naming)
-            config: Optional memory configuration dict (from manifest.json
-                   memory_config)
-                   Can include: collection_name, enable_graph, infer,
-                   embedding_model, chat_model, temperature, etc.
-                   Note: embedding_model_dims is auto-detected by embedding a
-                   test string - no need to specify!
-                   Embeddings and LLM are configured via environment variables
-                   (.env).
-
-        Raises:
-            Mem0MemoryServiceError: If mem0 is not available or initialization fails
-        """
-        # Lazy check for mem0 availability
         if not _check_mem0_available():
-            raise Mem0MemoryServiceError(
-                "Mem0 dependencies not available. Install with: pip install mem0ai"
-            )
+            raise Mem0MemoryServiceError("Mem0 not installed. pip install mem0ai")
 
         self.mongo_uri = mongo_uri
         self.db_name = db_name
         self.app_slug = app_slug
-
-        # Extract config with defaults
-        self.collection_name = (config or {}).get(
-            "collection_name", f"{app_slug}_memories"
-        )
-        config_embedding_dims = (config or {}).get(
-            "embedding_model_dims"
-        )  # Optional - will be auto-detected
-        self.enable_graph = (config or {}).get("enable_graph", False)
+        self.collection_name = (config or {}).get("collection_name", f"{app_slug}_memories")
         self.infer = (config or {}).get("infer", True)
-        self.async_mode = (config or {}).get("async_mode", True)
 
-        # Get model names from config or environment
-        # Default embedding model from config or env, fallback to common default
+        # Ensure GOOGLE_API_KEY is set for mem0 compatibility
+        # (mem0 expects GOOGLE_API_KEY, not GEMINI_API_KEY)
+        # This ensures we use the DIRECT Gemini API
+        # (generativelanguage.googleapis.com), NOT Vertex AI
+        if os.getenv("GEMINI_API_KEY") and not os.getenv("GOOGLE_API_KEY"):
+            os.environ["GOOGLE_API_KEY"] = os.getenv("GEMINI_API_KEY")
+            logger.info(
+                "Set GOOGLE_API_KEY from GEMINI_API_KEY for mem0 compatibility (direct Gemini API)"
+            )
+
+        # Verify we're NOT using Vertex AI (which would use GOOGLE_APPLICATION_CREDENTIALS)
+        if os.getenv("GOOGLE_APPLICATION_CREDENTIALS"):
+            logger.warning(
+                "GOOGLE_APPLICATION_CREDENTIALS is set - this would use Vertex AI, "
+                "not direct Gemini API"
+            )
+
+        # 1. Models & Config
         embedding_model = (config or {}).get("embedding_model") or os.getenv(
             "EMBEDDING_MODEL", "text-embedding-3-small"
         )
-        chat_model = (
-            (config or {}).get("chat_model")
-            or os.getenv("CHAT_MODEL")
-            or os.getenv("AZURE_OPENAI_DEPLOYMENT_NAME", "gpt-4o")
-        )
-        temperature = (config or {}).get(
-            "temperature", float(os.getenv("LLM_TEMPERATURE", "0.0"))
-        )
-
-        # Detect provider from environment variables
-        provider = _detect_provider_from_env()
+        chat_model = (config or {}).get("chat_model") or os.getenv("CHAT_MODEL", "gpt-4o")
 
-        # Verify required environment variables are set
-        if provider == "azure":
-            if not os.getenv("AZURE_OPENAI_API_KEY") or not os.getenv(
-                "AZURE_OPENAI_ENDPOINT"
-            ):
-                raise Mem0MemoryServiceError(
-                    "Azure OpenAI provider requires AZURE_OPENAI_API_KEY and "
-                    "AZURE_OPENAI_ENDPOINT environment variables to be set."
-                )
-        else:
-            if not os.getenv("OPENAI_API_KEY"):
-                raise Mem0MemoryServiceError(
-                    "OpenAI provider requires OPENAI_API_KEY environment variable to be set."
-                )
+        # 2. Build Mem0 Configuration
+        embedding_dims = (config or {}).get(
+            "embedding_model_dims"
+        ) or 1536  # Default for text-embedding-3-small
+        mem0_config = {
+            "vector_store": {
+                "provider": "mongodb",
+                "config": {
+                    "db_name": db_name,
+                    "collection_name": self.collection_name,
+                    "mongo_uri": mongo_uri,
+                    "embedding_model_dims": embedding_dims,
+                },
+            },
+            "embedder": self._build_provider_config("embedder", embedding_model),
+            "llm": self._build_provider_config("llm", chat_model) if self.infer else None,
+        }
 
+        # Add custom prompts to make fact extraction less restrictive (for document processing)
+        # The default mem0 prompts are too restrictive and filter out general facts
+        if self.infer:
+            # Long prompt string - using concatenation to avoid line length issues
+            fact_extraction_prompt = (
+                "You are a helpful assistant that extracts key facts, insights, "
+                "and information from documents and conversations.\n\n"
+                "Your task is to extract factual information, insights, and important details "
+                "from the provided content. Extract facts that would be useful for future "
+                "reference, including:\n"
+                "- Key concepts, definitions, and explanations\n"
+                "- Important dates, names, and entities\n"
+                "- Processes, procedures, and methodologies\n"
+                "- Insights, conclusions, and recommendations\n"
+                "- Relationships between concepts\n"
+                "- Any other factual information that would be valuable to remember\n\n"
+                'Return your response as a JSON object with a "facts" array. '
+                "Each fact should be a clear, standalone statement.\n\n"
+                "Example:\n"
+                'Input: "The Innovation Hub was established on August 14, 2024 by '
+                "David Vainchenker and Todd O'Brien. It focuses on experimental AI projects." + "\n"
+                'Output: {{"facts": ["The Innovation Hub was established on August 14, 2024", '
+                '"The Innovation Hub was founded by David Vainchenker and Todd O\'Brien", '
+                '"The Innovation Hub focuses on experimental AI projects"]}}' + "\n\n"
+                "Now extract facts from the following content:"
+            )
+            mem0_config["prompts"] = {"fact_extraction": fact_extraction_prompt}
+
+        # Filter None
+        mem0_config = {k: v for k, v in mem0_config.items() if v is not None}
+
+        # 3. Initialize
         try:
-            # Detect embedding dimensions using model name (fallback method)
-            detected_dims = _detect_embedding_dimensions(embedding_model)
-            self.embedding_model_dims = (
-                detected_dims
-                if detected_dims is not None
-                else (config_embedding_dims or 1536)
-            )
+            if hasattr(Memory, "from_config"):
+                self.memory = Memory.from_config(mem0_config)
+            else:
+                self.memory = Memory(mem0_config)
+            logger.info(f"✅ Mem0 Service active: {self.collection_name}")
+        except (
+            ValueError,
+            TypeError,
+            ConnectionError,
+            OSError,
+            AttributeError,
+            RuntimeError,
+        ) as e:
+            raise Mem0MemoryServiceError(f"Failed to init Mem0: {e}") from e
 
-            # Build mem0 config with MongoDB as vector store
-            mem0_config = _build_vector_store_config(
-                self.db_name,
-                self.collection_name,
-                self.mongo_uri,
-                self.embedding_model_dims,
-            )
+    def _build_provider_config(self, component, model_name):
+        """
+        Build provider configuration for embeddings or LLM.
 
-            # Configure mem0 embedder
-            mem0_config["embedder"] = _build_embedder_config(
-                provider, embedding_model, app_slug
-            )
+        For embeddings: Always use Azure OpenAI if available, otherwise OpenAI
+        For LLM: Detect provider from model name (gemini/google -> google_ai, else Azure/OpenAI)
+        """
+        clean_model = (
+            model_name.replace("azure/", "")
+            .replace("openai/", "")
+            .replace("google/", "")
+            .replace("gemini/", "")
+        )
 
-            # Configure LLM for inference (if infer: true)
-            if self.infer:
-                mem0_config["llm"] = _build_llm_config(
-                    provider, chat_model, temperature, app_slug
+        # For embeddings, always prefer Azure if available
+        if component == "embedder":
+            provider = "azure_openai" if os.getenv("AZURE_OPENAI_API_KEY") else "openai"
+            cfg = {"provider": provider, "config": {"model": clean_model}}
+
+            if provider == "azure_openai":
+                # Support both AZURE_OPENAI_EMBEDDING_DEPLOYMENT_NAME and AZURE_EMBEDDING_DEPLOYMENT
+                deployment_name = (
+                    os.getenv("AZURE_OPENAI_EMBEDDING_DEPLOYMENT_NAME")
+                    or os.getenv("AZURE_EMBEDDING_DEPLOYMENT")
+                    or clean_model
                 )
-        except (ValueError, TypeError, KeyError, AttributeError, ImportError) as e:
-            logger.error(
-                f"Failed to configure mem0: {e}",
-                extra={"app_slug": app_slug, "error": str(e)},
-            )
-            raise Mem0MemoryServiceError(f"Failed to configure mem0: {e}") from e
-
-        # Add graph store configuration if enabled
-        if self.enable_graph:
-            # Note: Graph store requires separate configuration (neo4j, memgraph, etc.)
-            # For now, we just enable it - actual graph store config should come from manifest
-            graph_config = (config or {}).get("graph_store")
-            if graph_config:
-                mem0_config["graph_store"] = graph_config
-            else:
-                logger.warning(
-                    "Graph memory enabled but no graph_store config provided. "
-                    "Graph features will not work. Configure graph_store in manifest.json",
-                    extra={"app_slug": app_slug},
-                )
-
-        try:
-            # Initialize Mem0 Memory instance
-            self.memory, init_method = _initialize_memory_instance(
-                mem0_config, app_slug
-            )
-
-            # Verify the memory instance has required methods
-            if not hasattr(self.memory, "get_all"):
-                logger.warning(
-                    f"Memory instance missing 'get_all' method for app '{app_slug}'",
-                    extra={"app_slug": app_slug, "init_method": init_method},
+                # Use API version from env or default
+                api_version = os.getenv("AZURE_OPENAI_API_VERSION", "2024-02-01")
+                cfg["config"]["azure_kwargs"] = {
+                    "api_version": api_version,
+                    "azure_deployment": deployment_name,
+                    "azure_endpoint": os.getenv("AZURE_OPENAI_ENDPOINT"),
+                    "api_key": os.getenv("AZURE_OPENAI_API_KEY"),
+                }
+                logger.info(
+                    f"Using Azure OpenAI embedding provider with deployment: "
+                    f"{deployment_name}, API version: {api_version}"
                 )
-            if not hasattr(self.memory, "add"):
-                logger.warning(
-                    f"Memory instance missing 'add' method for app '{app_slug}'",
-                    extra={"app_slug": app_slug, "init_method": init_method},
+            return cfg
+
+        # For LLM, detect provider from model name or env vars
+        model_lower = model_name.lower()
+        # Mem0 uses "gemini" as provider name (not "google_ai" or "vertexai")
+        # GOOGLE_API_KEY should already be set in __init__ if GEMINI_API_KEY was provided
+        has_gemini_key = os.getenv("GOOGLE_API_KEY") or os.getenv("GEMINI_API_KEY")
+        if "gemini" in model_lower or "google" in model_lower or has_gemini_key:
+            # Use Gemini provider for Mem0 (direct Gemini API, NOT Vertex AI)
+            provider = "gemini"
+            # Explicitly set API key in config to ensure direct Gemini API usage
+            api_key = os.getenv("GOOGLE_API_KEY") or os.getenv("GEMINI_API_KEY")
+            cfg = {
+                "provider": provider,
+                "config": {
+                    "model": clean_model,
+                    "api_key": api_key,  # Explicitly set to ensure direct API usage
+                },
+            }
+            logger.info(f"Using Gemini LLM provider (direct API) with model: {clean_model}")
+            return cfg
+        else:
+            # Use Azure OpenAI if available, otherwise OpenAI
+            provider = "azure_openai" if os.getenv("AZURE_OPENAI_API_KEY") else "openai"
+            cfg = {"provider": provider, "config": {"model": clean_model}}
+
+            if provider == "azure_openai":
+                deployment_name = os.getenv("AZURE_OPENAI_DEPLOYMENT_NAME", clean_model)
+                # Use API version from env or default (match .env default)
+                api_version = os.getenv("AZURE_OPENAI_API_VERSION", "2024-02-01")
+                cfg["config"]["azure_kwargs"] = {
+                    "api_version": api_version,
+                    "azure_deployment": deployment_name,
+                    "azure_endpoint": os.getenv("AZURE_OPENAI_ENDPOINT"),
+                    "api_key": os.getenv("AZURE_OPENAI_API_KEY"),
+                }
+                logger.info(
+                    f"Using Azure OpenAI LLM provider with deployment: "
+                    f"{deployment_name}, API version: {api_version}"
                 )
+            else:
+                logger.info(f"Using OpenAI LLM provider with model: {clean_model}")
+            return cfg
 
-            logger.info(
-                f"Mem0 Memory Service initialized using {init_method} for app '{app_slug}'",
-                extra={
-                    "app_slug": app_slug,
-                    "init_method": init_method,
-                    "collection_name": self.collection_name,
-                    "db_name": self.db_name,
-                    "enable_graph": self.enable_graph,
-                    "infer": self.infer,
-                    "has_get_all": hasattr(self.memory, "get_all"),
-                    "has_add": hasattr(self.memory, "add"),
-                    "embedder_provider": mem0_config.get("embedder", {}).get(
-                        "provider"
-                    ),
-                    "embedder_model": mem0_config.get("embedder", {})
-                    .get("config", {})
-                    .get("model"),
-                    "llm_provider": (
-                        mem0_config.get("llm", {}).get("provider")
-                        if self.infer
-                        else None
-                    ),
-                    "llm_model": (
-                        mem0_config.get("llm", {}).get("config", {}).get("model")
-                        if self.infer
-                        else None
-                    ),
-                },
-            )
-        except (
-            ImportError,
-            AttributeError,
-            TypeError,
-            ValueError,
-            RuntimeError,
-            KeyError,
-        ) as e:
-            logger.error(
-                f"Failed to initialize Mem0 Memory Service for app '{app_slug}': {e}",
-                exc_info=True,
-                extra={"app_slug": app_slug, "error": str(e)},
-            )
-            raise Mem0MemoryServiceError(
-                f"Failed to initialize Mem0 Memory Service: {e}"
-            ) from e
+    # --- Core Operations ---
 
     def add(
         self,
-        messages: Union[str, List[Dict[str, str]]],
-        user_id: Optional[str] = None,
-        metadata: Optional[Dict[str, Any]] = None,
+        messages: str | list[dict[str, str]],
+        user_id: str | None = None,
+        metadata: dict[str, Any] | None = None,
+        bucket_id: str | None = None,
+        bucket_type: str | None = None,
+        raw_content: str | None = None,
         **kwargs,
-    ) -> List[Dict[str, Any]]:
+    ) -> list[dict[str, Any]]:
         """
-        Add memories from messages or text.
-
-        This method intelligently extracts memories from conversations
-        and stores them in MongoDB. Memories are processed asynchronously
-        by default for better performance.
+        Add memories with user scoping and metadata convenience.
+        All operations are scoped per user_id for safety.
+        bucket_id and bucket_type are stored in metadata for filtering convenience.
+        """
+        if isinstance(messages, str):
+            messages = [{"role": "user", "content": messages}]
 
-        Args:
-            messages: Either a string or list of message dicts with 'role' and 'content'
-            user_id: Optional user ID to associate memories with
-            metadata: Optional metadata dict (e.g., {"category": "preferences"})
-            **kwargs: Additional mem0.add() parameters:
-                    - infer: Whether to infer memories (default: True)
-                    Note: async_mode is not a valid parameter for Mem0's add()
-                          method.
-                          Mem0 processes memories asynchronously by default.
-                          Graph features are configured at initialization via
-                          enable_graph in config, not per-add call.
+        # Merge metadata
+        final_metadata = dict(metadata) if metadata else {}
 
-        Returns:
-            List of memory events (each with 'id', 'event', 'data')
+        # CRITICAL: Database indexing relies on these fields being in metadata
+        if bucket_id:
+            final_metadata["bucket_id"] = bucket_id
+            final_metadata["context_id"] = bucket_id  # Backwards compatibility
 
-        Example:
-            ```python
-            memories = memory_service.add(
-                messages=[
-                    {"role": "user", "content": "I love sci-fi movies"},
-                    {"role": "assistant", "content": "Noted! I'll remember that."}
-                ],
-                user_id="alice",
-                metadata={"category": "preferences"}
-            )
-            ```
-        """
-        try:
-            # Normalize messages format
-            if isinstance(messages, str):
-                messages = [{"role": "user", "content": messages}]
+        if bucket_type:
+            final_metadata["bucket_type"] = bucket_type
 
-            # Prepare kwargs with defaults from config
-            # async_mode is not a valid parameter for Mem0's add() method
-            add_kwargs = {"infer": kwargs.pop("infer", self.infer), **kwargs}
-            add_kwargs.pop("async_mode", None)
+        # Store raw_content in metadata if provided (metadata convenience)
+        if raw_content:
+            final_metadata["raw_content"] = raw_content
 
-            # enable_graph is configured at initialization, not per-add call
-            # Mem0 processes asynchronously by default
-            # Log message content preview for debugging
-            message_preview = []
-            for i, msg in enumerate(messages[:5]):  # Show first 5 messages
-                if isinstance(msg, dict):
-                    role = msg.get("role", "unknown")
-                    content = msg.get("content", "")
-                    preview = content[:150] + "..." if len(content) > 150 else content
-                    message_preview.append(f"{i+1}. {role}: {preview}")
+        # Infer defaults to configured value unless overridden
+        infer = kwargs.pop("infer", self.infer)
 
-            logger.info(
-                f"🔵 CALLING mem0.add() - app_slug='{self.app_slug}', "
-                f"user_id='{user_id}', messages={len(messages)}, "
-                f"infer={add_kwargs.get('infer', 'N/A')}",
-                extra={
-                    "app_slug": self.app_slug,
-                    "user_id": user_id,
-                    "collection_name": self.collection_name,
-                    "message_count": len(messages),
-                    "message_preview": "\n".join(message_preview),
-                    "infer": add_kwargs.get("infer"),
-                    "metadata": metadata or {},
-                    "add_kwargs": add_kwargs,
-                },
+        try:
+            logger.debug(
+                f"Calling mem0.add() with infer={infer}, user_id={user_id}, bucket_id={bucket_id}"
             )
-
             result = self.memory.add(
                 messages=messages,
-                user_id=str(user_id),  # Ensure string - mem0 might be strict about this
-                metadata=metadata or {},
-                **add_kwargs,
+                user_id=str(user_id) if user_id else None,
+                metadata=final_metadata,
+                infer=infer,
+                **kwargs,
             )
-
-            # Normalize result format - mem0.add() may return different formats
-            if isinstance(result, dict):
-                # Some versions return {"results": [...]} or {"data": [...]}
-                if "results" in result:
-                    result = result["results"]
-                elif "data" in result:
-                    result = result["data"] if isinstance(result["data"], list) else []
-                elif "memory" in result:
-                    # Single memory object
-                    result = [result]
-
-            # Ensure result is always a list
-            if not isinstance(result, list):
-                result = [result] if result else []
-
-            result_length = len(result) if isinstance(result, list) else 0
+            # Log raw result before normalization
             logger.debug(
-                f"Raw result from mem0.add(): type={type(result)}, "
-                f"length={result_length}",
-                extra={
-                    "app_slug": self.app_slug,
-                    "user_id": user_id,
-                    "result_type": str(type(result)),
-                    "is_list": isinstance(result, list),
-                    "result_length": len(result) if isinstance(result, list) else 0,
-                    "result_sample": (
-                        result[0]
-                        if result and isinstance(result, list) and len(result) > 0
-                        else None
-                    ),
-                },
+                f"mem0.add() raw result: type={type(result)}, "
+                f"value={str(result)[:500] if result else 'None'}"
             )
-
+            normalized = self._normalize_result(result)
             logger.info(
-                f"Added {len(result)} memories for user '{user_id}'",
-                extra={
-                    "app_slug": self.app_slug,
-                    "user_id": user_id,
-                    "message_count": len(messages),
-                    "memory_count": len(result) if isinstance(result, list) else 0,
-                    "memory_ids": (
-                        [
-                            m.get("id") or m.get("_id")
-                            for m in result
-                            if isinstance(m, dict)
-                        ]
-                        if result
-                        else []
-                    ),
-                    "infer_enabled": add_kwargs.get("infer", False),
-                    "has_llm": (
-                        hasattr(self.memory, "llm") and self.memory.llm is not None
-                        if hasattr(self.memory, "llm")
-                        else False
-                    ),
-                },
+                f"mem0.add() normalized to {len(normalized)} memories "
+                f"(raw result type: {type(result)})"
             )
-
-            # If 0 memories and infer is enabled, log helpful info
-            if len(result) == 0 and add_kwargs.get("infer", False):
-                # Extract conversation content for analysis
-                conversation_text = "\n".join(
-                    [
-                        f"{msg.get('role', 'unknown')}: {msg.get('content', '')[:100]}"
-                        for msg in messages[:5]
-                    ]
+            if not normalized and infer:
+                logger.warning(
+                    f"⚠️ mem0.add() with infer=True returned empty result. Raw result: {result}"
                 )
-
-                logger.info(
-                    "ℹ️ mem0.add() returned 0 memories. This is normal if the "
-                    "conversation doesn't contain extractable facts. "
-                    "mem0 extracts personal preferences, facts, and details - "
-                    "not generic greetings or small talk. "
-                    "Try conversations like 'I love pizza' or 'I work as a "
-                    "software engineer' to see memories extracted.",
-                    extra={
-                        "app_slug": self.app_slug,
-                        "user_id": user_id,
-                        "message_count": len(messages),
-                        "infer": True,
-                        "has_llm": (
-                            hasattr(self.memory, "llm") and self.memory.llm is not None
-                            if hasattr(self.memory, "llm")
-                            else False
-                        ),
-                        "conversation_preview": conversation_text,
-                    },
+                # Try to understand why - check if it's a dict with empty results
+                if isinstance(result, dict):
+                    logger.warning(f"   Result dict keys: {list(result.keys())}")
+                    if "results" in result:
+                        logger.warning(f"   result['results']: {result['results']}")
+                    if "data" in result:
+                        logger.warning(f"   result['data']: {result['data']}")
+            return normalized
+        except (
+            ValueError,
+            TypeError,
+            ConnectionError,
+            OSError,
+            AttributeError,
+            RuntimeError,
+            KeyError,
+        ) as e:
+            error_msg = str(e)
+            # Handle rate limit errors gracefully - try storing without inference
+            if (
+                "429" in error_msg
+                or "RESOURCE_EXHAUSTED" in error_msg
+                or "rate limit" in error_msg.lower()
+            ):
+                logger.warning(
+                    f"Rate limit hit during memory inference, storing without inference: "
+                    f"{error_msg}"
                 )
-
-            return result
-
-        except (AttributeError, TypeError, ValueError, RuntimeError, KeyError) as e:
-            logger.error(
-                f"Failed to add memories: {e}",
-                exc_info=True,
-                extra={"app_slug": self.app_slug, "user_id": user_id, "error": str(e)},
-            )
-            raise Mem0MemoryServiceError(f"Failed to add memories: {e}") from e
+                # Retry without inference to at least store the raw content
+                try:
+                    result = self.memory.add(
+                        messages=messages,
+                        user_id=str(user_id) if user_id else None,
+                        metadata=final_metadata,
+                        infer=False,  # Disable inference to avoid rate limits
+                        **kwargs,
+                    )
+                    logger.info("Successfully stored memory without inference due to rate limit")
+                    return self._normalize_result(result)
+                except (
+                    ValueError,
+                    TypeError,
+                    ConnectionError,
+                    OSError,
+                    AttributeError,
+                    RuntimeError,
+                    KeyError,
+                ) as retry_error:
+                    logger.exception("Failed to store memory even without inference")
+                    raise Mem0MemoryServiceError(
+                        f"Add failed (rate limited, retry also failed): {retry_error}"
+                    ) from retry_error
+            else:
+                logger.exception("Mem0 Add Failed")
+                raise Mem0MemoryServiceError(f"Add failed: {e}") from e
 
     def get_all(
         self,
-        user_id: Optional[str] = None,
-        limit: Optional[int] = None,
-        retry_on_empty: bool = True,
-        max_retries: int = 2,
-        retry_delay: float = 0.5,
+        user_id: str | None = None,
+        limit: int = 100,
+        filters: dict[str, Any] | None = None,
         **kwargs,
-    ) -> List[Dict[str, Any]]:
+    ) -> list[dict[str, Any]]:
         """
-        Get all memories for a user.
-
-        Args:
-            user_id: User ID to retrieve memories for
-            limit: Optional limit on number of memories to return
-            retry_on_empty: If True, retry if result is empty (handles async processing delay)
-            max_retries: Maximum number of retries if result is empty
-            retry_delay: Delay in seconds between retries
-            **kwargs: Additional mem0.get_all() parameters
-
-        Returns:
-            List of memory dictionaries
+        Get all memories with direct database filtering.
         """
-        import time
-
         try:
-            # Verify memory instance is valid before calling
-            if not hasattr(self, "memory") or self.memory is None:
-                logger.error(
-                    f"Memory instance is None or missing for app '{self.app_slug}'",
-                    extra={"app_slug": self.app_slug, "user_id": user_id},
-                )
-                return []
-
-            logger.info(
-                f"🟢 CALLING mem0.get_all() - app_slug='{self.app_slug}', "
-                f"user_id='{user_id}' (type: {type(user_id).__name__}), "
-                f"collection='{self.collection_name}'",
-                extra={
-                    "app_slug": self.app_slug,
-                    "user_id": user_id,
-                    "user_id_type": type(user_id).__name__,
-                    "user_id_repr": repr(user_id),
-                    "collection_name": self.collection_name,
-                    "limit": limit,
-                    "kwargs": kwargs,
-                },
-            )
-
-            result = None
-            attempt = 0
-
-            while attempt <= max_retries:
-                if attempt > 0:
-                    # Wait before retry to allow async processing to complete
-                    time.sleep(retry_delay * attempt)  # Exponential backoff
-                    logger.debug(
-                        f"Retrying mem0.get_all (attempt {attempt + 1}/{max_retries + 1})",
-                        extra={
-                            "app_slug": self.app_slug,
-                            "user_id": user_id,
-                            "attempt": attempt + 1,
-                        },
-                    )
-
-                # Call with safety - catch any exceptions from mem0
-                try:
-                    logger.debug(
-                        f"🟢 EXECUTING: memory.get_all(user_id='{user_id}', "
-                        f"limit={limit}, kwargs={kwargs})",
-                        extra={
-                            "app_slug": self.app_slug,
-                            "user_id": user_id,
-                            "collection_name": self.collection_name,
-                            "attempt": attempt + 1,
-                        },
-                    )
-                    result = self.memory.get_all(
-                        user_id=str(user_id), limit=limit, **kwargs
-                    )  # Ensure string
-                    result_length = (
-                        len(result) if isinstance(result, (list, dict)) else "N/A"
-                    )
-                    logger.debug(
-                        f"🟢 RESULT RECEIVED: type={type(result).__name__}, "
-                        f"length={result_length}",
-                        extra={
-                            "app_slug": self.app_slug,
-                            "user_id": user_id,
-                            "result_type": type(result).__name__,
-                            "result_length": (
-                                len(result) if isinstance(result, (list, dict)) else 0
-                            ),
-                            "attempt": attempt + 1,
-                        },
-                    )
-                except AttributeError as attr_error:
-                    logger.error(
-                        f"Memory.get_all method not available: {attr_error}",
-                        extra={
-                            "app_slug": self.app_slug,
-                            "user_id": user_id,
-                            "error": str(attr_error),
-                            "attempt": attempt + 1,
-                        },
-                    )
-                    return []  # Return empty list instead of retrying
-                # Type 4: Let other exceptions bubble up to framework handler
-
-                logger.debug(
-                    f"Raw result from mem0.get_all (attempt {attempt + 1}): type={type(result)}",
-                    extra={
-                        "app_slug": self.app_slug,
-                        "user_id": user_id,
-                        "attempt": attempt + 1,
-                        "result_type": str(type(result)),
-                        "is_dict": isinstance(result, dict),
-                        "is_list": isinstance(result, list),
-                        "result_length": (
-                            len(result) if isinstance(result, (list, dict)) else 0
-                        ),
-                    },
-                )
-
-                # Handle Mem0 v2 API response format: {"results": [...], "total": N}
-                if isinstance(result, dict):
-                    if "results" in result:
-                        result = result["results"]  # Extract results array
-                        logger.debug(
-                            "Extracted results from dict response",
-                            extra={
-                                "app_slug": self.app_slug,
-                                "user_id": user_id,
-                                "result_count": (
-                                    len(result) if isinstance(result, list) else 0
-                                ),
-                            },
-                        )
-                    elif "data" in result:
-                        # Alternative format: {"data": [...]}
-                        result = (
-                            result["data"] if isinstance(result["data"], list) else []
-                        )
-
-                # Ensure result is always a list for backward compatibility
-                if not isinstance(result, list):
-                    result = [result] if result else []
-
-                # If we got results or retries are disabled, break
-                if not retry_on_empty or len(result) > 0 or attempt >= max_retries:
-                    break
-
-                attempt += 1
-
-            logger.info(
-                f"Retrieved {len(result)} memories for user '{user_id}' "
-                f"(after {attempt + 1} attempt(s))",
-                extra={
-                    "app_slug": self.app_slug,
-                    "user_id": user_id,
-                    "memory_count": len(result) if isinstance(result, list) else 0,
-                    "attempts": attempt + 1,
-                    "sample_memory": (
-                        result[0]
-                        if result and isinstance(result, list) and len(result) > 0
-                        else None
-                    ),
-                },
-            )
+            call_kwargs = {"limit": limit}
+            if user_id:
+                call_kwargs["user_id"] = str(user_id)
+            if filters:
+                call_kwargs["filters"] = filters  # Passed to MongoDB $match
 
-            return result
+            call_kwargs.update(kwargs)
 
-        except (AttributeError, TypeError, ValueError, RuntimeError, KeyError) as e:
-            attempt_num = (
-                attempt + 1 if "attempt" in locals() and attempt is not None else 1
-            )
-            logger.error(
-                f"Failed to get memories: {e}",
-                exc_info=True,
-                extra={
-                    "app_slug": self.app_slug,
-                    "user_id": user_id,
-                    "error": str(e),
-                    "error_type": type(e).__name__,
-                    "attempt": attempt_num,
-                },
-            )
-            raise Mem0MemoryServiceError(f"Failed to get memories: {e}") from e
+            return self._normalize_result(self.memory.get_all(**call_kwargs))
+        except (
+            ValueError,
+            TypeError,
+            ConnectionError,
+            OSError,
+            AttributeError,
+            RuntimeError,
+            KeyError,
+        ):
+            logger.exception("Mem0 get_all failed")
+            return []
 
     def search(
         self,
         query: str,
-        user_id: Optional[str] = None,
-        limit: Optional[int] = None,
-        metadata: Optional[Dict[str, Any]] = None,
-        filters: Optional[Dict[str, Any]] = None,
+        user_id: str | None = None,
+        limit: int = 5,
+        filters: dict[str, Any] | None = None,
         **kwargs,
-    ) -> List[Dict[str, Any]]:
+    ) -> list[dict[str, Any]]:
         """
-        Search for relevant memories using semantic search.
-
-        Args:
-            query: Search query string
-            user_id: Optional user ID to scope search to
-            limit: Optional limit on number of results
-            metadata: Optional metadata dict to filter results
-                     (e.g., {"category": "travel"})
-                     Deprecated in favor of 'filters' parameter for Mem0 1.0.0+
-            filters: Optional enhanced filters dict (Mem0 1.0.0+) with operators
-                     like {"category": {"eq": "travel"}}
-            **kwargs: Additional mem0.search() parameters
-
-        Returns:
-            List of relevant memory dictionaries
-
-        Example:
-            ```python
-            # Simple metadata filter (backward compatible)
-            results = memory_service.search(
-                query="What are my travel plans?",
-                user_id="alice",
-                metadata={"category": "travel"}
-            )
-
-            # Enhanced filters (Mem0 1.0.0+)
-            results = memory_service.search(
-                query="high priority tasks",
-                user_id="alice",
-                filters={
-                    "AND": [
-                        {"category": "work"},
-                        {"priority": {"gte": 5}}
-                    ]
-                }
-            )
-            ```
-        """
-        try:
-            # Build search kwargs
-            search_kwargs = {"limit": limit, **kwargs}
-
-            # Prefer 'filters' parameter (Mem0 1.0.0+) over 'metadata' (legacy)
-            if filters is not None:
-                search_kwargs["filters"] = filters
-            elif metadata:
-                # Backward compatibility: convert simple metadata to filters format
-                # Try 'filters' first, fallback to 'metadata' if it fails
-                search_kwargs["filters"] = metadata
-
-            # Call search - try with filters first, fallback to metadata if needed
-            try:
-                result = self.memory.search(
-                    query=query, user_id=user_id, **search_kwargs
-                )
-            except (TypeError, ValueError) as e:
-                # If filters parameter doesn't work, try with metadata (backward compatibility)
-                if "filters" in search_kwargs and metadata:
-                    logger.debug(
-                        f"Filters parameter failed, trying metadata parameter: {e}",
-                        extra={"app_slug": self.app_slug, "user_id": user_id},
-                    )
-                    search_kwargs.pop("filters", None)
-                    search_kwargs["metadata"] = metadata
-                    result = self.memory.search(
-                        query=query, user_id=user_id, **search_kwargs
-                    )
-                else:
-                    raise
-
-            # Handle response format - search may return dict with "results" key
-            if isinstance(result, dict):
-                if "results" in result:
-                    result = result["results"]
-                elif "data" in result:
-                    result = result["data"] if isinstance(result["data"], list) else []
-
-            # Ensure result is always a list
-            if not isinstance(result, list):
-                result = [result] if result else []
-
-            logger.debug(
-                f"Searched memories for user '{user_id}'",
-                extra={
-                    "app_slug": self.app_slug,
-                    "user_id": user_id,
-                    "query": query,
-                    "metadata_filter": metadata,
-                    "filters": filters,
-                    "result_count": len(result) if isinstance(result, list) else 0,
-                },
-            )
-
-            return result
-
-        except (AttributeError, TypeError, ValueError, RuntimeError, KeyError) as e:
-            logger.error(
-                f"Failed to search memories: {e}",
-                exc_info=True,
-                extra={
-                    "app_slug": self.app_slug,
-                    "user_id": user_id,
-                    "query": query,
-                    "metadata": metadata,
-                    "filters": filters,
-                    "error": str(e),
-                },
-            )
-            raise Mem0MemoryServiceError(f"Failed to search memories: {e}") from e
-
-    def get(
-        self, memory_id: str, user_id: Optional[str] = None, **kwargs
-    ) -> Dict[str, Any]:
-        """
-        Get a single memory by ID.
-
-        Args:
-            memory_id: Memory ID to retrieve
-            user_id: Optional user ID for scoping
-            **kwargs: Additional mem0.get() parameters
-
-        Returns:
-            Memory dictionary
-
-        Example:
-            ```python
-            memory = memory_service.get(memory_id="mem_123", user_id="alice")
-            ```
-        """
-        try:
-            # Mem0's get() method doesn't accept user_id as a parameter
-            # User scoping should be handled via metadata or filters if needed
-            # For now, we just get by memory_id
-            result = self.memory.get(memory_id=memory_id, **kwargs)
-
-            # If user_id is provided, verify the memory belongs to that user
-            # by checking metadata or user_id field in the result
-            if user_id and isinstance(result, dict):
-                result_user_id = result.get("user_id") or result.get(
-                    "metadata", {}
-                ).get("user_id")
-                if result_user_id and result_user_id != user_id:
-                    logger.warning(
-                        f"Memory {memory_id} does not belong to user {user_id}",
-                        extra={
-                            "memory_id": memory_id,
-                            "user_id": user_id,
-                            "result_user_id": result_user_id,
-                        },
-                    )
-                    raise Mem0MemoryServiceError(
-                        f"Memory {memory_id} does not belong to user {user_id}"
-                    )
-
-            logger.debug(
-                f"Retrieved memory '{memory_id}' for user '{user_id}'",
-                extra={
-                    "app_slug": self.app_slug,
-                    "user_id": user_id,
-                    "memory_id": memory_id,
-                },
-            )
-
-            return result
-
-        except (AttributeError, TypeError, ValueError, RuntimeError, KeyError) as e:
-            logger.error(
-                f"Failed to get memory: {e}",
-                exc_info=True,
-                extra={
-                    "app_slug": self.app_slug,
-                    "user_id": user_id,
-                    "memory_id": memory_id,
-                    "error": str(e),
-                },
-            )
-            raise Mem0MemoryServiceError(f"Failed to get memory: {e}") from e
-
-    def update(
-        self,
-        memory_id: str,
-        data: Union[str, List[Dict[str, str]]],
-        user_id: Optional[str] = None,
-        metadata: Optional[Dict[str, Any]] = None,
-        **kwargs,
-    ) -> Dict[str, Any]:
+        Semantic search with metadata filters, scoped per user.
         """
-        Update a memory by ID with new data.
-
-        Args:
-            memory_id: Memory ID to update
-            data: New data (string or list of message dicts with 'role' and 'content')
-            user_id: Optional user ID for scoping
-            metadata: Optional metadata dict to update
-            **kwargs: Additional mem0.update() parameters
+        final_filters = filters or {}
 
-        Returns:
-            Updated memory dictionary
-
-        Example:
-            ```python
-            updated = memory_service.update(
-                memory_id="mem_123",
-                data="I am a software engineer using Python and FastAPI.",
-                user_id="bob"
-            )
-            ```
-        """
         try:
-            # Normalize data format
-            if isinstance(data, str):
-                data = [{"role": "user", "content": data}]
-
-            # Mem0's update() may not accept user_id directly
-            # Pass it in metadata if user_id is provided
-            update_metadata = metadata or {}
-            if user_id:
-                update_metadata["user_id"] = user_id
+            call_kwargs = {"limit": limit}
+            if final_filters:
+                call_kwargs["filters"] = final_filters
 
-            # Try with user_id first, fall back without it if it fails
-            try:
-                result = self.memory.update(
-                    memory_id=memory_id,
-                    data=data,
-                    user_id=user_id,
-                    metadata=update_metadata,
-                    **kwargs,
+            return self._normalize_result(
+                self.memory.search(
+                    query=query, user_id=str(user_id) if user_id else None, **call_kwargs, **kwargs
                 )
-            except TypeError as e:
-                if "unexpected keyword argument 'user_id'" in str(e):
-                    # Mem0 doesn't accept user_id, try without it
-                    result = self.memory.update(
-                        memory_id=memory_id,
-                        data=data,
-                        metadata=update_metadata,
-                        **kwargs,
-                    )
-                else:
-                    raise
-
-            logger.info(
-                f"Updated memory '{memory_id}' for user '{user_id}'",
-                extra={
-                    "app_slug": self.app_slug,
-                    "user_id": user_id,
-                    "memory_id": memory_id,
-                },
-            )
-
-            return result
-
-        except (AttributeError, TypeError, ValueError, RuntimeError, KeyError) as e:
-            logger.error(
-                f"Failed to update memory: {e}",
-                exc_info=True,
-                extra={
-                    "app_slug": self.app_slug,
-                    "user_id": user_id,
-                    "memory_id": memory_id,
-                    "error": str(e),
-                },
             )
-            raise Mem0MemoryServiceError(f"Failed to update memory: {e}") from e
-
-    def delete(self, memory_id: str, user_id: Optional[str] = None, **kwargs) -> bool:
-        """
-        Delete a memory by ID.
-
-        Args:
-            memory_id: Memory ID to delete
-            user_id: Optional user ID for scoping
-            **kwargs: Additional mem0.delete() parameters
+        except (
+            ValueError,
+            TypeError,
+            ConnectionError,
+            OSError,
+            AttributeError,
+            RuntimeError,
+            KeyError,
+        ):
+            logger.exception("Mem0 search failed")
+            return []
 
-        Returns:
-            True if deletion was successful
-        """
+    def get(self, memory_id: str, user_id: str | None = None, **kwargs) -> dict[str, Any]:
         try:
-            # Mem0's delete() may not accept user_id directly
-            # Try with user_id first, fall back without it if it fails
-            try:
-                result = self.memory.delete(
-                    memory_id=memory_id, user_id=user_id, **kwargs
-                )
-            except TypeError as e:
-                if "unexpected keyword argument 'user_id'" in str(e):
-                    # Mem0 doesn't accept user_id, try without it
-                    # User scoping should be handled via metadata or filters
-                    result = self.memory.delete(memory_id=memory_id, **kwargs)
-                else:
-                    raise
-
-            logger.info(
-                f"Deleted memory '{memory_id}' for user '{user_id}'",
-                extra={
-                    "app_slug": self.app_slug,
-                    "user_id": user_id,
-                    "memory_id": memory_id,
-                },
-            )
-
-            return result
-
-        except (AttributeError, TypeError, ValueError, RuntimeError, KeyError) as e:
-            logger.error(
-                f"Failed to delete memory: {e}",
-                exc_info=True,
-                extra={
-                    "app_slug": self.app_slug,
-                    "user_id": user_id,
-                    "memory_id": memory_id,
-                    "error": str(e),
-                },
-            )
-            raise Mem0MemoryServiceError(f"Failed to delete memory: {e}") from e
-
-    def delete_all(self, user_id: Optional[str] = None, **kwargs) -> bool:
-        """
-        Delete all memories for a user.
-
-        Args:
-            user_id: User ID to delete all memories for
-            **kwargs: Additional mem0.delete_all() parameters
-
-        Returns:
-            True if deletion was successful
+            return self.memory.get(memory_id, **kwargs)
+        except (
+            ValueError,
+            TypeError,
+            ConnectionError,
+            OSError,
+            AttributeError,
+            RuntimeError,
+            KeyError,
+        ):
+            return None
 
-        Example:
-            ```python
-            success = memory_service.delete_all(user_id="alice")
-            ```
-        """
+    def delete(self, memory_id: str, user_id: str | None = None, **kwargs) -> bool:
         try:
-            result = self.memory.delete_all(user_id=user_id, **kwargs)
-
-            logger.info(
-                f"Deleted all memories for user '{user_id}'",
-                extra={"app_slug": self.app_slug, "user_id": user_id},
-            )
+            self.memory.delete(memory_id, **kwargs)
+            return True
+        except (
+            AttributeError,
+            ValueError,
+            RuntimeError,
+            KeyError,
+            TypeError,
+            ConnectionError,
+            OSError,
+        ):
+            return False
 
+    def delete_all(self, user_id: str | None = None, **kwargs) -> bool:
+        try:
+            self.memory.delete_all(user_id=user_id, **kwargs)
+            return True
+        except (
+            AttributeError,
+            ValueError,
+            RuntimeError,
+            KeyError,
+            TypeError,
+            ConnectionError,
+            OSError,
+        ):
+            return False
+
+    def _normalize_result(self, result: Any) -> list[dict[str, Any]]:
+        """Normalize Mem0's return type (dict vs list)."""
+        if result is None:
+            return []
+        if isinstance(result, dict):
+            if "results" in result:
+                return result["results"]
+            if "data" in result:
+                return result["data"]
+            return [result]
+        if isinstance(result, list):
             return result
+        return []
 
-        except (AttributeError, TypeError, ValueError, RuntimeError, KeyError) as e:
-            logger.error(
-                f"Failed to delete all memories: {e}",
-                exc_info=True,
-                extra={"app_slug": self.app_slug, "user_id": user_id, "error": str(e)},
-            )
-            raise Mem0MemoryServiceError(f"Failed to delete all memories: {e}") from e
-
-
-def get_memory_service(
-    mongo_uri: str, db_name: str, app_slug: str, config: Optional[Dict[str, Any]] = None
-) -> Mem0MemoryService:
-    """
-    Get or create a Mem0MemoryService instance (cached).
-
-    Args:
-        mongo_uri: MongoDB connection URI
-        db_name: Database name
-        app_slug: App slug
-        config: Optional memory configuration dict
-
-    Returns:
-        Mem0MemoryService instance
-    """
-    # Lazy check for mem0 availability
-    if not _check_mem0_available():
-        raise Mem0MemoryServiceError(
-            "Mem0 dependencies not available. Install with: pip install mem0ai"
-        )
 
-    return Mem0MemoryService(
-        mongo_uri=mongo_uri, db_name=db_name, app_slug=app_slug, config=config
-    )
+def get_memory_service(mongo_uri, db_name, app_slug, config=None):
+    return Mem0MemoryService(mongo_uri, db_name, app_slug, config)