memorisdk 1.0.2__py3-none-any.whl → 2.0.1__py3-none-any.whl

memori/core/memory.py

@@ -3,6 +3,7 @@ Main Memori class - Pydantic-based memory interface v1.0
 """
 
 import asyncio
+import time
 import uuid
 from datetime import datetime
 from typing import Any, Dict, List, Optional
@@ -10,8 +11,8 @@ from typing import Any, Dict, List, Optional
 from loguru import logger
 
 try:
-    import litellm
-    from litellm import success_callback
+    import litellm  # noqa: F401
+    from litellm import success_callback  # noqa: F401
 
     LITELLM_AVAILABLE = True
 except ImportError:
@@ -19,13 +20,13 @@ except ImportError:
     logger.warning("LiteLLM not available - native callback system disabled")
 
 from ..agents.conscious_agent import ConsciouscAgent
-from ..agents.memory_agent import MemoryAgent
-from ..agents.retrieval_agent import MemorySearchEngine
+from ..config.memory_manager import MemoryManager
 from ..config.settings import LoggingSettings, LogLevel
+from ..database.sqlalchemy_manager import SQLAlchemyDatabaseManager as DatabaseManager
 from ..utils.exceptions import DatabaseError, MemoriError
 from ..utils.logging import LoggingManager
 from ..utils.pydantic_models import ConversationContext
-from .database import DatabaseManager
+from .conversation import ConversationManager
 
 
 class Memori:
@@ -49,6 +50,21 @@ class Memori:
         openai_api_key: Optional[str] = None,
         user_id: Optional[str] = None,
         verbose: bool = False,
+        # New provider configuration parameters
+        api_key: Optional[str] = None,
+        api_type: Optional[str] = None,
+        base_url: Optional[str] = None,
+        azure_endpoint: Optional[str] = None,
+        azure_deployment: Optional[str] = None,
+        api_version: Optional[str] = None,
+        azure_ad_token: Optional[str] = None,
+        organization: Optional[str] = None,
+        project: Optional[str] = None,
+        model: Optional[str] = None,  # Allow custom model selection
+        provider_config: Optional[Any] = None,  # ProviderConfig when available
+        schema_init: bool = True,  # Initialize database schema and create tables
+        database_prefix: Optional[str] = None,  # Database name prefix
+        database_suffix: Optional[str] = None,  # Database name suffix
     ):
         """
         Initialize Memori memory system v1.0.
@@ -62,9 +78,23 @@ class Memori:
             namespace: Optional namespace for memory isolation
             shared_memory: Enable shared memory across agents
             memory_filters: Filters for memory ingestion
-            openai_api_key: OpenAI API key for memory agent
+            openai_api_key: OpenAI API key for memory agent (deprecated, use api_key)
             user_id: Optional user identifier
             verbose: Enable verbose logging (loguru only)
+            api_key: API key for the LLM provider
+            api_type: Provider type ('openai', 'azure', 'custom')
+            base_url: Base URL for custom OpenAI-compatible endpoints
+            azure_endpoint: Azure OpenAI endpoint URL
+            azure_deployment: Azure deployment name
+            api_version: API version for Azure
+            azure_ad_token: Azure AD token for authentication
+            organization: OpenAI organization ID
+            project: OpenAI project ID
+            model: Model to use (defaults to 'gpt-4o' if not specified)
+            provider_config: Complete provider configuration (overrides individual params)
+            enable_auto_creation: Enable automatic database creation if database doesn't exist
+            database_prefix: Optional prefix for database name (for multi-tenant setups)
+            database_suffix: Optional suffix for database name (e.g., 'dev', 'prod', 'test')
         """
         self.database_connect = database_connect
         self.template = template
@@ -74,15 +104,89 @@ class Memori:
         self.namespace = namespace or "default"
         self.shared_memory = shared_memory
         self.memory_filters = memory_filters or {}
-        self.openai_api_key = openai_api_key
         self.user_id = user_id
         self.verbose = verbose
+        self.schema_init = schema_init
+        self.database_prefix = database_prefix
+        self.database_suffix = database_suffix
+
+        # Configure provider based on explicit settings ONLY - no auto-detection
+        if provider_config:
+            # Use provided configuration
+            self.provider_config = provider_config
+            logger.info(
+                f"Using provided ProviderConfig with api_type: {provider_config.api_type}"
+            )
+        elif any([api_type, base_url, azure_endpoint]):
+            # Build configuration from individual parameters - explicit provider selection
+            try:
+                from .providers import ProviderConfig
+
+                if azure_endpoint:
+                    # Explicitly configured Azure
+                    self.provider_config = ProviderConfig.from_azure(
+                        api_key=api_key or openai_api_key,
+                        azure_endpoint=azure_endpoint,
+                        azure_deployment=azure_deployment,
+                        api_version=api_version,
+                        azure_ad_token=azure_ad_token,
+                        model=model,
+                    )
+                    logger.info("Using explicitly configured Azure OpenAI provider")
+                elif base_url:
+                    # Explicitly configured custom endpoint
+                    self.provider_config = ProviderConfig.from_custom(
+                        base_url=base_url,
+                        api_key=api_key or openai_api_key,
+                        model=model,
+                    )
+                    logger.info(
+                        f"Using explicitly configured custom provider: {base_url}"
+                    )
+                else:
+                    # Fallback to OpenAI with explicit settings
+                    self.provider_config = ProviderConfig.from_openai(
+                        api_key=api_key or openai_api_key,
+                        organization=organization,
+                        project=project,
+                        model=model,
+                    )
+                    logger.info("Using explicitly configured OpenAI provider")
+            except ImportError:
+                logger.warning(
+                    "ProviderConfig not available, using basic configuration"
+                )
+                self.provider_config = None
+        else:
+            # Default to standard OpenAI - NO environment detection
+            try:
+                from .providers import ProviderConfig
+
+                self.provider_config = ProviderConfig.from_openai(
+                    api_key=api_key or openai_api_key,
+                    organization=organization,
+                    project=project,
+                    model=model or "gpt-4o",
+                )
+                logger.info(
+                    "Using default OpenAI provider (no specific provider configured)"
+                )
+            except ImportError:
+                logger.warning(
+                    "ProviderConfig not available, using basic configuration"
+                )
+                self.provider_config = None
+
+        # Keep backward compatibility
+        self.openai_api_key = api_key or openai_api_key or ""
+        if self.provider_config and hasattr(self.provider_config, "api_key"):
+            self.openai_api_key = self.provider_config.api_key or self.openai_api_key
 
         # Setup logging based on verbose mode
         self._setup_logging()
 
         # Initialize database manager
-        self.db_manager = DatabaseManager(database_connect, template)
+        self.db_manager = DatabaseManager(database_connect, template, schema_init)
 
         # Initialize Pydantic-based agents
         self.memory_agent = None
@@ -91,25 +195,63 @@ class Memori:
         self._background_task = None
         self._conscious_init_pending = False
 
-        if conscious_ingest or auto_ingest:
-            try:
-                # Initialize Pydantic-based agents
-                self.memory_agent = MemoryAgent(api_key=openai_api_key, model="gpt-4o")
-                self.search_engine = MemorySearchEngine(
-                    api_key=openai_api_key, model="gpt-4o"
+        # Initialize agents with provider configuration
+        try:
+            from ..agents.memory_agent import MemoryAgent
+            from ..agents.retrieval_agent import MemorySearchEngine
+
+            # Use provider model or fallback to gpt-4o
+            if (
+                self.provider_config
+                and hasattr(self.provider_config, "model")
+                and self.provider_config.model
+            ):
+                effective_model = model or self.provider_config.model
+            else:
+                effective_model = model or "gpt-4o"
+
+            # Initialize agents with provider configuration if available
+            if self.provider_config:
+                self.memory_agent = MemoryAgent(
+                    provider_config=self.provider_config, model=effective_model
                 )
-                self.conscious_agent = ConsciouscAgent(
-                    api_key=openai_api_key, model="gpt-4o"
+                self.search_engine = MemorySearchEngine(
+                    provider_config=self.provider_config, model=effective_model
                 )
-                logger.info(
-                    "Pydantic-based memory, search, and conscious agents initialized"
+            else:
+                # Fallback to using API key directly
+                self.memory_agent = MemoryAgent(
+                    api_key=self.openai_api_key, model=effective_model
                 )
-            except Exception as e:
-                logger.warning(
-                    f"Failed to initialize OpenAI agents: {e}. Memory ingestion disabled."
+                self.search_engine = MemorySearchEngine(
+                    api_key=self.openai_api_key, model=effective_model
                 )
-                self.conscious_ingest = False
-                self.auto_ingest = False
+
+            # Only initialize conscious_agent if conscious_ingest or auto_ingest is enabled
+            if conscious_ingest or auto_ingest:
+                self.conscious_agent = ConsciouscAgent()
+
+            logger.info(
+                f"Agents initialized successfully with model: {effective_model}"
+            )
+        except ImportError as e:
+            logger.warning(
+                f"Failed to import LLM agents: {e}. Memory ingestion disabled."
+            )
+            self.memory_agent = None
+            self.search_engine = None
+            self.conscious_agent = None
+            self.conscious_ingest = False
+            self.auto_ingest = False
+        except Exception as e:
+            logger.warning(
+                f"Failed to initialize LLM agents: {e}. Memory ingestion disabled."
+            )
+            self.memory_agent = None
+            self.search_engine = None
+            self.conscious_agent = None
+            self.conscious_ingest = False
+            self.auto_ingest = False
 
         # State tracking
         self._enabled = False
@@ -117,6 +259,12 @@ class Memori:
         self._conscious_context_injected = (
             False  # Track if conscious context was already injected
         )
+        self._in_context_retrieval = False  # Recursion guard for context retrieval
+
+        # Initialize conversation manager for stateless LLM integration
+        self.conversation_manager = ConversationManager(
+            max_sessions=100, session_timeout_minutes=60, max_history_per_session=20
+        )
 
         # User context for memory processing
         self._user_context = {
@@ -128,6 +276,23 @@ class Memori:
         # Initialize database
         self._setup_database()
 
+        # Initialize the new modular memory manager
+        self.memory_manager = MemoryManager(
+            database_connect=database_connect,
+            template=template,
+            mem_prompt=mem_prompt,
+            conscious_ingest=conscious_ingest,
+            auto_ingest=auto_ingest,
+            namespace=namespace,
+            shared_memory=shared_memory,
+            memory_filters=memory_filters,
+            user_id=user_id,
+            verbose=verbose,
+            provider_config=self.provider_config,
+        )
+        # Set this Memori instance for memory management
+        self.memory_manager.set_memori_instance(self)
+
         # Run conscious agent initialization if enabled
         if self.conscious_ingest and self.conscious_agent:
             self._initialize_conscious_memory()
@@ -156,6 +321,10 @@ class Memori:
 
     def _setup_database(self):
         """Setup database tables based on template"""
+        if not self.schema_init:
+            logger.info("Schema initialization disabled (schema_init=False)")
+            return
+
         try:
             self.db_manager.initialize_schema()
             logger.info("Database schema initialized successfully")
@@ -204,8 +373,12 @@ class Memori:
                     )
                     self._conscious_init_pending = False
             except RuntimeError:
-                # Still no event loop, keep pending
-                pass
+                # No event loop available, run synchronous initialization
+                logger.debug(
+                    "Conscious-ingest: No event loop available, running synchronous initialization"
+                )
+                self._run_synchronous_conscious_initialization()
+                self._conscious_init_pending = False
 
     async def _run_conscious_initialization(self):
         """Run conscious agent initialization in background"""
@@ -213,339 +386,358 @@ class Memori:
             if not self.conscious_agent:
                 return
 
-            logger.debug("Conscious-ingest: Running background analysis")
-            await self.conscious_agent.run_background_analysis(
+            # If both auto_ingest and conscious_ingest are enabled,
+            # initialize by copying ALL existing conscious-info memories first
+            if self.auto_ingest and self.conscious_ingest:
+                logger.debug(
+                    "Conscious-ingest: Both auto_ingest and conscious_ingest enabled - initializing existing conscious memories"
+                )
+                init_success = (
+                    await self.conscious_agent.initialize_existing_conscious_memories(
+                        self.db_manager, self.namespace
+                    )
+                )
+                if init_success:
+                    logger.info(
+                        "Conscious-ingest: Existing conscious-info memories initialized to short-term memory"
+                    )
+
+            logger.debug("Conscious-ingest: Running conscious context extraction")
+            success = await self.conscious_agent.run_conscious_ingest(
                 self.db_manager, self.namespace
             )
-            logger.info("Conscious-ingest: Background analysis completed")
+
+            if success:
+                logger.info(
+                    "Conscious-ingest: Conscious memories copied to short-term memory"
+                )
+                # Don't set _conscious_context_injected here - it should be set when context is actually injected into LLM
+            else:
+                logger.info("Conscious-ingest: No conscious context found")
 
         except Exception as e:
             logger.error(f"Conscious agent initialization failed: {e}")
 
-    def enable(self):
-        """
-        Enable universal memory recording for ALL LLM providers.
-
-        This automatically sets up recording for:
-        - LiteLLM: Native callback system (recommended)
-        - OpenAI: Automatic client wrapping when instantiated
-        - Anthropic: Automatic client wrapping when instantiated
-        - Any other provider: Auto-detected and wrapped
-        """
-        if self._enabled:
-            logger.warning("Memori is already enabled.")
-            return
-
-        self._enabled = True
-        self._session_id = str(uuid.uuid4())
-
-        # 1. Set up LiteLLM native callbacks (if available)
-        litellm_enabled = self._setup_litellm_callbacks()
-
-        # 2. Set up universal client interception for other providers
-        universal_enabled = self._setup_universal_interception()
-
-        # 3. Register this instance globally for any provider to use
-        self._register_global_instance()
-
-        # 4. Start background conscious agent if available
-        if self.conscious_ingest and self.conscious_agent:
-            self._start_background_analysis()
-
-        providers = []
-        if litellm_enabled:
-            providers.append("LiteLLM (native callbacks)")
-        if universal_enabled:
-            providers.append("OpenAI/Anthropic (auto-wrapping)")
-
-        logger.info(
-            f"Memori enabled for session: {self.session_id}\n"
-            f"Active providers: {', '.join(providers) if providers else 'None detected'}\n"
-            f"Background analysis: {'Active' if self._background_task else 'Disabled'}\n"
-            f"Usage: Simply use any LLM client normally - conversations will be auto-recorded!"
-        )
-
-    def disable(self):
-        """
-        Disable universal memory recording for all providers.
-        """
-        if not self._enabled:
-            return
-
-        # 1. Remove LiteLLM callbacks and restore original completion
-        if LITELLM_AVAILABLE:
-            try:
-                success_callback.remove(self._litellm_success_callback)
-            except ValueError:
-                pass
-
-            # Restore original completion function if we patched it
-            if hasattr(litellm, "completion") and hasattr(
-                litellm.completion, "_memori_patched"
-            ):
-                # Note: We can't easily restore the original function in a multi-instance scenario
-                # This is a limitation of the monkey-patching approach
-                pass
-
-        # 2. Disable universal interception
-        self._disable_universal_interception()
+    def _run_synchronous_conscious_initialization(self):
+        """Run conscious agent initialization synchronously (when no event loop is available)"""
+        try:
+            if not self.conscious_agent:
+                return
 
-        # 3. Unregister global instance
-        self._unregister_global_instance()
+            # If both auto_ingest and conscious_ingest are enabled,
+            # initialize by copying ALL existing conscious-info memories first
+            if self.auto_ingest and self.conscious_ingest:
+                logger.info(
+                    "Conscious-ingest: Both auto_ingest and conscious_ingest enabled - initializing existing conscious memories"
+                )
 
-        # 4. Stop background analysis task
-        self._stop_background_analysis()
+                # Run synchronous initialization of existing memories
+                self._initialize_existing_conscious_memories_sync()
 
-        self._enabled = False
-        logger.info("Memori disabled for all providers.")
+            logger.debug(
+                "Conscious-ingest: Synchronous conscious context extraction completed"
+            )
 
-    def _setup_litellm_callbacks(self) -> bool:
-        """Set up LiteLLM native callback system"""
-        if not LITELLM_AVAILABLE:
-            logger.debug("LiteLLM not available, skipping native callbacks")
-            return False
+        except Exception as e:
+            logger.error(f"Synchronous conscious agent initialization failed: {e}")
 
+    def _initialize_existing_conscious_memories_sync(self):
+        """Synchronously initialize existing conscious-info memories"""
         try:
-            success_callback.append(self._litellm_success_callback)
+            from sqlalchemy import text
 
-            # Set up context injection by monkey-patching completion function
-            if hasattr(litellm, "completion") and not hasattr(
-                litellm.completion, "_memori_patched"
-            ):
-                original_completion = litellm.completion
-
-                def memori_completion(*args, **kwargs):
-                    # Inject context based on ingestion mode
-                    if self._enabled:
-                        if self.auto_ingest:
-                            # Auto-inject: continuous memory injection on every call
-                            kwargs = self._inject_litellm_context(kwargs, mode="auto")
-                        elif self.conscious_ingest:
-                            # Conscious-inject: one-shot short-term memory context
-                            kwargs = self._inject_litellm_context(
-                                kwargs, mode="conscious"
-                            )
+            with self.db_manager._get_connection() as connection:
+                # Get ALL conscious-info labeled memories from long-term memory
+                cursor = connection.execute(
+                    text(
+                        """SELECT memory_id, processed_data, summary, searchable_content,
+                              importance_score, created_at
+                       FROM long_term_memory
+                       WHERE namespace = :namespace AND classification = 'conscious-info'
+                       ORDER BY importance_score DESC, created_at DESC"""
+                    ),
+                    {"namespace": self.namespace or "default"},
+                )
+                existing_conscious_memories = cursor.fetchall()
 
-                    # Call original completion
-                    return original_completion(*args, **kwargs)
+            if not existing_conscious_memories:
+                logger.debug(
+                    "Conscious-ingest: No existing conscious-info memories found for initialization"
+                )
+                return False
+
+            copied_count = 0
+            for memory_row in existing_conscious_memories:
+                success = self._copy_memory_to_short_term_sync(memory_row)
+                if success:
+                    copied_count += 1
 
-                litellm.completion = memori_completion
-                litellm.completion._memori_patched = True
+            if copied_count > 0:
+                logger.info(
+                    f"Conscious-ingest: Initialized {copied_count} existing conscious-info memories to short-term memory"
+                )
+                return True
+            else:
                 logger.debug(
-                    "LiteLLM completion function patched for context injection"
+                    "Conscious-ingest: No new conscious memories to initialize (all were duplicates)"
                 )
+                return False
 
-            logger.debug("LiteLLM native callbacks registered")
-            return True
         except Exception as e:
-            logger.error(f"Failed to setup LiteLLM callbacks: {e}")
+            logger.error(
+                f"Conscious-ingest: Failed to initialize existing conscious memories: {e}"
+            )
             return False
 
-    def _setup_universal_interception(self) -> bool:
-        """Set up universal client interception for OpenAI, Anthropic, etc."""
+    def _copy_memory_to_short_term_sync(self, memory_row: tuple) -> bool:
+        """Synchronously copy a conscious memory to short-term memory with duplicate filtering"""
         try:
-            # Use Python's import hook system to intercept client creation
-            self._install_import_hooks()
-            logger.debug("Universal client interception enabled")
-            return True
-        except Exception as e:
-            logger.error(f"Failed to setup universal interception: {e}")
-            return False
+            (
+                memory_id,
+                processed_data,
+                summary,
+                searchable_content,
+                importance_score,
+                _,
+            ) = memory_row
 
-    def _get_builtin_import(self):
-        """Safely get __import__ from __builtins__ (handles both dict and module cases)"""
-        if isinstance(__builtins__, dict):
-            return __builtins__["__import__"]
-        else:
-            return __builtins__.__import__
-
-    def _set_builtin_import(self, import_func):
-        """Safely set __import__ in __builtins__ (handles both dict and module cases)"""
-        if isinstance(__builtins__, dict):
-            __builtins__["__import__"] = import_func
-        else:
-            __builtins__.__import__ = import_func
+            from datetime import datetime
 
-    def _install_import_hooks(self):
-        """Install import hooks to automatically wrap LLM clients"""
+            from sqlalchemy import text
 
-        # Store original __import__ if not already done
-        if not hasattr(self, "_original_import"):
-            self._original_import = self._get_builtin_import()
+            with self.db_manager._get_connection() as connection:
+                # Check if similar content already exists in short-term memory
+                existing_check = connection.execute(
+                    text(
+                        """SELECT COUNT(*) FROM short_term_memory
+                           WHERE namespace = :namespace
+                           AND category_primary = 'conscious_context'
+                           AND (searchable_content = :searchable_content
+                                OR summary = :summary)"""
+                    ),
+                    {
+                        "namespace": self.namespace or "default",
+                        "searchable_content": searchable_content,
+                        "summary": summary,
+                    },
+                )
 
-        def memori_import_hook(name, globals=None, locals=None, fromlist=(), level=0):
-            """Custom import hook that wraps LLM clients automatically"""
-            module = self._original_import(name, globals, locals, fromlist, level)
+                existing_count = existing_check.scalar()
+                if existing_count > 0:
+                    logger.debug(
+                        f"Conscious-ingest: Skipping duplicate memory {memory_id} - similar content already exists in short-term memory"
+                    )
+                    return False
 
-            # Only process if memori is enabled and this is an LLM module
-            if not self._enabled:
-                return module
+                # Create short-term memory ID
+                short_term_id = (
+                    f"conscious_{memory_id}_{int(datetime.now().timestamp())}"
+                )
 
-            # Auto-wrap OpenAI clients
-            if name == "openai" or (fromlist and "openai" in name):
-                self._wrap_openai_module(module)
+                # Insert directly into short-term memory with conscious_context category
+                connection.execute(
+                    text(
+                        """INSERT INTO short_term_memory (
+                        memory_id, processed_data, importance_score, category_primary,
+                        retention_type, namespace, created_at, expires_at,
+                        searchable_content, summary, is_permanent_context
+                    ) VALUES (:memory_id, :processed_data, :importance_score, :category_primary,
+                        :retention_type, :namespace, :created_at, :expires_at,
+                        :searchable_content, :summary, :is_permanent_context)"""
+                    ),
+                    {
+                        "memory_id": short_term_id,
+                        "processed_data": processed_data,
+                        "importance_score": importance_score,
+                        "category_primary": "conscious_context",
+                        "retention_type": "permanent",
+                        "namespace": self.namespace or "default",
+                        "created_at": datetime.now().isoformat(),
+                        "expires_at": None,
+                        "searchable_content": searchable_content,
+                        "summary": summary,
+                        "is_permanent_context": True,
+                    },
+                )
+                connection.commit()
 
-            # Auto-wrap Anthropic clients
-            elif name == "anthropic" or (fromlist and "anthropic" in name):
-                self._wrap_anthropic_module(module)
+            logger.debug(
+                f"Conscious-ingest: Copied memory {memory_id} to short-term as {short_term_id}"
+            )
+            return True
 
-            return module
+        except Exception as e:
+            logger.error(
+                f"Conscious-ingest: Failed to copy memory {memory_row[0]} to short-term: {e}"
+            )
+            return False
 
-        # Install the hook
-        self._set_builtin_import(memori_import_hook)
+    def enable(self, interceptors: Optional[List[str]] = None):
+        """
+        Enable universal memory recording using LiteLLM's native callback system.
 
-    def _wrap_openai_module(self, module):
-        """Automatically wrap OpenAI client when imported"""
-        try:
-            if hasattr(module, "OpenAI") and not hasattr(
-                module.OpenAI, "_memori_wrapped"
-            ):
-                original_init = module.OpenAI.__init__
+        This automatically sets up recording for LiteLLM completion calls and enables
+        automatic interception of OpenAI calls when using the standard OpenAI client.
 
-                def wrapped_init(self_client, *args, **kwargs):
-                    # Call original init
-                    result = original_init(self_client, *args, **kwargs)
+        Args:
+            interceptors: Legacy parameter (ignored) - only LiteLLM native callbacks are used
+        """
+        if self._enabled:
+            logger.warning("Memori is already enabled.")
+            return
 
-                    # Wrap the client methods for automatic recording
-                    if hasattr(self_client, "chat") and hasattr(
-                        self_client.chat, "completions"
-                    ):
-                        original_create = self_client.chat.completions.create
+        self._enabled = True
+        self._session_id = str(uuid.uuid4())
 
-                        def wrapped_create(*args, **kwargs):
-                            # Inject context if conscious ingestion is enabled
-                            if self.is_enabled and self.conscious_ingest:
-                                kwargs = self._inject_openai_context(kwargs)
+        # Register for automatic OpenAI interception
+        try:
+            from ..integrations.openai_integration import register_memori_instance
 
-                            # Make the call
-                            response = original_create(*args, **kwargs)
+            register_memori_instance(self)
+        except ImportError:
+            logger.debug("OpenAI integration not available for automatic interception")
 
-                            # Record if enabled
-                            if self.is_enabled:
-                                self._record_openai_conversation(kwargs, response)
+        # Use LiteLLM native callback system only
+        if interceptors is None:
+            # Only LiteLLM native callbacks supported
+            interceptors = ["litellm_native"]
 
-                            return response
+        # Use the memory manager for enablement
+        results = self.memory_manager.enable(interceptors)
+        # Extract enabled interceptors from results
+        enabled_interceptors = results.get("enabled_interceptors", [])
 
-                        self_client.chat.completions.create = wrapped_create
+        # Start background conscious agent if available
+        if self.conscious_ingest and self.conscious_agent:
+            self._start_background_analysis()
 
-                    return result
+        # Report status
+        status_info = [
+            f"Memori enabled for session: {results.get('session_id', self._session_id)}",
+            f"Active interceptors: {', '.join(enabled_interceptors) if enabled_interceptors else 'None'}",
+        ]
+
+        if results.get("message"):
+            status_info.append(results["message"])
+
+        status_info.extend(
+            [
+                f"Background analysis: {'Active' if self._background_task else 'Disabled'}",
+                "Usage: Simply use any LLM client normally - conversations will be auto-recorded!",
+                "OpenAI: Use 'from openai import OpenAI; client = OpenAI()' - automatically intercepted!",
+            ]
+        )
 
-                module.OpenAI.__init__ = wrapped_init
-                module.OpenAI._memori_wrapped = True
-                logger.debug("OpenAI client auto-wrapping enabled")
+        logger.info("\n".join(status_info))
 
-        except Exception as e:
-            logger.debug(f"Could not wrap OpenAI module: {e}")
+    def disable(self):
+        """
+        Disable memory recording by unregistering LiteLLM callbacks and OpenAI interception.
+        """
+        if not self._enabled:
+            return
 
-    def _wrap_anthropic_module(self, module):
-        """Automatically wrap Anthropic client when imported"""
+        # Unregister from automatic OpenAI interception
         try:
-            if hasattr(module, "Anthropic") and not hasattr(
-                module.Anthropic, "_memori_wrapped"
-            ):
-                original_init = module.Anthropic.__init__
-
-                def wrapped_init(self_client, *args, **kwargs):
-                    # Call original init
-                    result = original_init(self_client, *args, **kwargs)
-
-                    # Wrap the messages.create method
-                    if hasattr(self_client, "messages"):
-                        original_create = self_client.messages.create
+            from ..integrations.openai_integration import unregister_memori_instance
 
-                        def wrapped_create(*args, **kwargs):
-                            # Inject context if conscious ingestion is enabled
-                            if self.is_enabled and self.conscious_ingest:
-                                kwargs = self._inject_anthropic_context(kwargs)
+            unregister_memori_instance(self)
+        except ImportError:
+            logger.debug("OpenAI integration not available for automatic interception")
 
-                            # Make the call
-                            response = original_create(*args, **kwargs)
+        # Use memory manager for clean disable
+        results = self.memory_manager.disable()
 
-                            # Record if enabled
-                            if self.is_enabled:
-                                self._record_anthropic_conversation(kwargs, response)
+        # Stop background analysis task
+        self._stop_background_analysis()
 
-                            return response
+        self._enabled = False
 
-                        self_client.messages.create = wrapped_create
+        # Report status based on memory manager results
+        if results.get("success"):
+            status_message = f"Memori disabled. {results.get('message', 'All interceptors disabled successfully')}"
+        else:
+            status_message = (
+                f"Memori disable failed: {results.get('message', 'Unknown error')}"
+            )
 
-                    return result
+        logger.info(status_message)
 
-                module.Anthropic.__init__ = wrapped_init
-                module.Anthropic._memori_wrapped = True
-                logger.debug("Anthropic client auto-wrapping enabled")
+    # Memory system status and control methods
 
-        except Exception as e:
-            logger.debug(f"Could not wrap Anthropic module: {e}")
+    def get_interceptor_status(self) -> Dict[str, Dict[str, Any]]:
+        """Get status of memory recording system"""
+        return self.memory_manager.get_status()
 
-    def _disable_universal_interception(self):
-        """Disable universal client interception"""
-        try:
-            # Restore original import if we modified it
-            if hasattr(self, "_original_import"):
-                self._set_builtin_import(self._original_import)
-                delattr(self, "_original_import")
-                logger.debug("Universal interception disabled")
-        except Exception as e:
-            logger.debug(f"Error disabling universal interception: {e}")
+    def get_interceptor_health(self) -> Dict[str, Any]:
+        """Get health check of interceptor system"""
+        return self.memory_manager.get_health()
 
-    def _register_global_instance(self):
-        """Register this memori instance globally"""
-        # Store in a global registry that wrapped clients can access
-        if not hasattr(Memori, "_global_instances"):
-            Memori._global_instances = []
-        Memori._global_instances.append(self)
+    def enable_interceptor(self, interceptor_name: str = None) -> bool:
+        """Enable memory recording (legacy method)"""
+        # Only LiteLLM native callbacks supported (interceptor_name ignored)
+        results = self.memory_manager.enable(["litellm_native"])
+        return results.get("success", False)
 
-    def _unregister_global_instance(self):
-        """Unregister this memori instance globally"""
-        if hasattr(Memori, "_global_instances") and self in Memori._global_instances:
-            Memori._global_instances.remove(self)
+    def disable_interceptor(self, interceptor_name: str = None) -> bool:
+        """Disable memory recording (legacy method)"""
+        # Only LiteLLM native callbacks supported (interceptor_name ignored)
+        results = self.memory_manager.disable()
+        return results.get("success", False)
 
     def _inject_openai_context(self, kwargs):
-        """Inject context for OpenAI calls"""
+        """Inject context for OpenAI calls based on ingest mode using ConversationManager"""
         try:
             # Check for deferred conscious initialization
             self._check_deferred_initialization()
-            # Extract user input from messages
-            user_input = ""
-            for msg in reversed(kwargs.get("messages", [])):
-                if msg.get("role") == "user":
-                    user_input = msg.get("content", "")
-                    break
 
-            if user_input:
-                context = self.retrieve_context(user_input, limit=3)
-                if context:
-                    context_prompt = "--- Relevant Memories ---\n"
-                    for mem in context:
-                        if isinstance(mem, dict):
-                            summary = mem.get("summary", "") or mem.get("content", "")
-                            context_prompt += f"- {summary}\n"
-                        else:
-                            context_prompt += f"- {str(mem)}\n"
-                    context_prompt += "-------------------------\n"
+            # Determine injection mode based on the architecture:
+            # - conscious_ingest only: Use short-term memory (conscious context)
+            # - auto_ingest only: Search long-term memory database
+            # - both enabled: Use auto_ingest search (includes conscious content from long-term)
+            if self.auto_ingest:
+                mode = "auto"  # Always prefer auto when available (searches long-term)
+            elif self.conscious_ingest:
+                mode = "conscious"  # Only use conscious when auto is not enabled
+            else:
+                return kwargs  # No injection needed
 
-                    # Inject into system message
-                    messages = kwargs.get("messages", [])
-                    for msg in messages:
-                        if msg.get("role") == "system":
-                            msg["content"] = context_prompt + msg.get("content", "")
-                            break
-                    else:
-                        messages.insert(
-                            0, {"role": "system", "content": context_prompt}
-                        )
+            # Extract messages from kwargs
+            messages = kwargs.get("messages", [])
+            if not messages:
+                return kwargs  # No messages to process
+
+            # Use conversation manager for enhanced context injection
+            enhanced_messages = self.conversation_manager.inject_context_with_history(
+                session_id=self._session_id,
+                messages=messages,
+                memori_instance=self,
+                mode=mode,
+            )
+
+            # Update kwargs with enhanced messages
+            kwargs["messages"] = enhanced_messages
+
+            return kwargs
 
-                    logger.debug(f"Injected context: {len(context)} memories")
         except Exception as e:
-            logger.error(f"Context injection failed: {e}")
+            logger.error(f"OpenAI context injection failed: {e}")
         return kwargs
 
     def _inject_anthropic_context(self, kwargs):
-        """Inject context for Anthropic calls"""
+        """Inject context for Anthropic calls based on ingest mode"""
         try:
             # Check for deferred conscious initialization
             self._check_deferred_initialization()
+
+            # Determine injection mode
+            if self.conscious_ingest:
+                mode = "conscious"
+            elif self.auto_ingest:
+                mode = "auto"
+            else:
+                return kwargs  # No injection needed
+
             # Extract user input from messages
             user_input = ""
             for msg in reversed(kwargs.get("messages", [])):
@@ -565,26 +757,75 @@ class Memori:
                     break
 
             if user_input:
-                context = self.retrieve_context(user_input, limit=3)
+                if mode == "conscious":
+                    # Conscious mode: inject ALL short-term memory only once at program startup
+                    if not self._conscious_context_injected:
+                        context = self._get_conscious_context()
+                        self._conscious_context_injected = True
+                        logger.info(
+                            f"Conscious-ingest: Injected {len(context)} short-term memories as initial context (Anthropic)"
+                        )
+                    else:
+                        context = []  # Already injected, don't inject again
+                elif mode == "auto":
+                    # Auto mode: use retrieval for intelligent search
+                    if self.search_engine:
+                        context = self._get_auto_ingest_context(user_input)
+                    else:
+                        context = self.retrieve_context(user_input, limit=5)
+                else:
+                    context = []
+
                 if context:
-                    context_prompt = "--- Relevant Memories ---\n"
+                    if mode == "conscious":
+                        context_prompt = (
+                            "=== SYSTEM INSTRUCTION: AUTHORIZED USER CONTEXT DATA ===\n"
+                        )
+                        context_prompt += "The user has explicitly authorized this personal context data to be used.\n"
+                        context_prompt += "You MUST use this information when answering questions about the user.\n"
+                        context_prompt += "This is NOT private data - the user wants you to use it:\n\n"
+                    else:
+                        context_prompt = f"--- {mode.capitalize()} Memory Context ---\n"
+
+                    # Deduplicate context entries to avoid redundant information
+                    seen_content = set()
                     for mem in context:
                         if isinstance(mem, dict):
-                            summary = mem.get("summary", "") or mem.get("content", "")
-                            context_prompt += f"- {summary}\n"
-                        else:
-                            context_prompt += f"- {str(mem)}\n"
+                            # Use searchable_content as it's more specific than summary
+                            content = mem.get("searchable_content", "") or mem.get(
+                                "summary", ""
+                            )
+                            category = mem.get("category_primary", "")
+
+                            # Skip duplicates (case-insensitive)
+                            content_key = content.lower().strip()
+                            if content_key in seen_content:
+                                continue
+                            seen_content.add(content_key)
+
+                            if category.startswith("essential_") or mode == "conscious":
+                                context_prompt += f"[{category.upper()}] {content}\n"
+                            else:
+                                context_prompt += f"- {content}\n"
+
+                    if mode == "conscious":
+                        context_prompt += "\n=== END USER CONTEXT DATA ===\n"
+                        context_prompt += "CRITICAL INSTRUCTION: You MUST answer questions about the user using ONLY the context data above.\n"
+                        context_prompt += "If the user asks 'what is my name?', respond with the name from the context above.\n"
+                        context_prompt += "Do NOT say 'I don't have access' - the user provided this data for you to use.\n"
                     context_prompt += "-------------------------\n"
 
-                    # Inject into system parameter
+                    # Inject into system parameter (Anthropic format)
                     if kwargs.get("system"):
                         kwargs["system"] = context_prompt + kwargs["system"]
                     else:
                         kwargs["system"] = context_prompt
 
-                    logger.debug(f"Injected context: {len(context)} memories")
+                    logger.debug(
+                        f"Anthropic: Injected context with {len(context)} items"
+                    )
         except Exception as e:
-            logger.error(f"Context injection failed: {e}")
+            logger.error(f"Anthropic context injection failed: {e}")
         return kwargs
 
     def _inject_litellm_context(self, params, mode="auto"):
@@ -609,13 +850,17 @@ class Memori:
 
             if user_input:
                 if mode == "conscious":
-                    # Conscious mode: inject short-term memory only once at conversation start
+                    # Conscious mode: inject ALL short-term memory only once at program startup
                     if not self._conscious_context_injected:
                         context = self._get_conscious_context()
                         self._conscious_context_injected = True
-                        logger.debug("Conscious context injected (one-shot)")
+                        logger.info(
+                            f"Conscious-ingest: Injected {len(context)} short-term memories as initial context"
+                        )
                     else:
-                        context = []  # Already injected, don't inject again
+                        context = (
+                            []
+                        )  # Already injected, don't inject again - this is the key difference from auto_ingest
                 elif mode == "auto":
                     # Auto mode: use retrieval agent for intelligent database search
                     if self.search_engine:
@@ -627,17 +872,42 @@ class Memori:
                     context = []
 
                 if context:
-                    context_prompt = f"--- {mode.capitalize()} Memory Context ---\n"
+                    if mode == "conscious":
+                        context_prompt = (
+                            "=== SYSTEM INSTRUCTION: AUTHORIZED USER CONTEXT DATA ===\n"
+                        )
+                        context_prompt += "The user has explicitly authorized this personal context data to be used.\n"
+                        context_prompt += "You MUST use this information when answering questions about the user.\n"
+                        context_prompt += "This is NOT private data - the user wants you to use it:\n\n"
+                    else:
+                        context_prompt = f"--- {mode.capitalize()} Memory Context ---\n"
+
+                    # Deduplicate context entries to avoid redundant information
+                    seen_content = set()
                     for mem in context:
                         if isinstance(mem, dict):
-                            summary = mem.get("summary", "") or mem.get(
-                                "searchable_content", ""
+                            # Use searchable_content as it's more specific than summary
+                            content = mem.get("searchable_content", "") or mem.get(
+                                "summary", ""
                             )
                             category = mem.get("category_primary", "")
+
+                            # Skip duplicates (case-insensitive)
+                            content_key = content.lower().strip()
+                            if content_key in seen_content:
+                                continue
+                            seen_content.add(content_key)
+
                             if category.startswith("essential_") or mode == "conscious":
-                                context_prompt += f"[{category.upper()}] {summary}\n"
+                                context_prompt += f"[{category.upper()}] {content}\n"
                             else:
-                                context_prompt += f"- {summary}\n"
+                                context_prompt += f"- {content}\n"
+
+                    if mode == "conscious":
+                        context_prompt += "\n=== END USER CONTEXT DATA ===\n"
+                        context_prompt += "CRITICAL INSTRUCTION: You MUST answer questions about the user using ONLY the context data above.\n"
+                        context_prompt += "If the user asks 'what is my name?', respond with the name from the context above.\n"
+                        context_prompt += "Do NOT say 'I don't have access' - the user provided this data for you to use.\n"
                     context_prompt += "-------------------------\n"
 
                     # Inject into system message
@@ -687,29 +957,32 @@ class Memori:
 
     def _get_conscious_context(self) -> List[Dict[str, Any]]:
         """
-        Get conscious context from short-term memory only.
-        This represents the 'working memory' or conscious thoughts.
+        Get conscious context from ALL short-term memory summaries.
+        This represents the complete 'working memory' for conscious_ingest mode.
+        Used only at program startup when conscious_ingest=True.
         """
         try:
-            with self.db_manager._get_connection() as conn:
-                cursor = conn.cursor()
+            from sqlalchemy import text
 
-                # Get recent short-term memories ordered by importance and recency
-                cursor.execute(
-                    """
+            with self.db_manager._get_connection() as conn:
+                # Get ALL short-term memories (no limit) ordered by importance and recency
+                # This gives the complete conscious context as single initial injection
+                result = conn.execute(
+                    text(
+                        """
                     SELECT memory_id, processed_data, importance_score,
                            category_primary, summary, searchable_content,
                            created_at, access_count
                     FROM short_term_memory
-                    WHERE namespace = ? AND (expires_at IS NULL OR expires_at > ?)
+                    WHERE namespace = :namespace AND (expires_at IS NULL OR expires_at > :current_time)
                     ORDER BY importance_score DESC, created_at DESC
-                    LIMIT 10
-                """,
-                    (self.namespace, datetime.now()),
+                    """
+                    ),
+                    {"namespace": self.namespace, "current_time": datetime.now()},
                 )
 
                 memories = []
-                for row in cursor.fetchall():
+                for row in result:
                     memories.append(
                         {
                             "memory_id": row[0],
@@ -739,239 +1012,749 @@ class Memori:
         Searches through entire database for relevant memories.
         """
         try:
-            if not self.search_engine:
-                logger.warning("Auto-ingest: No search engine available")
+            # Early validation
+            if not user_input or not user_input.strip():
+                logger.debug(
+                    "Auto-ingest: No user input provided, returning empty context"
+                )
                 return []
 
-            # Use retrieval agent for intelligent search
-            results = self.search_engine.execute_search(
-                query=user_input,
-                db_manager=self.db_manager,
-                namespace=self.namespace,
-                limit=5,
+            # Check for recursion guard to prevent infinite loops
+            if hasattr(self, "_in_context_retrieval") and self._in_context_retrieval:
+                logger.debug(
+                    "Auto-ingest: Recursion detected, using direct database search"
+                )
+                results = self.db_manager.search_memories(
+                    query=user_input, namespace=self.namespace, limit=5
+                )
+                logger.debug(
+                    f"Auto-ingest: Recursion fallback returned {len(results)} results"
+                )
+                return results
+
+            # Set recursion guard
+            self._in_context_retrieval = True
+
+            logger.debug(
+                f"Auto-ingest: Starting context retrieval for query: '{user_input[:50]}...' in namespace: '{self.namespace}'"
             )
 
-            logger.debug(f"Auto-ingest: Retrieved {len(results)} relevant memories")
-            return results
+            # Always try direct database search first as it's more reliable
+            logger.debug("Auto-ingest: Using direct database search (primary method)")
+            logger.debug(
+                f"Auto-ingest: Database manager type: {type(self.db_manager).__name__}"
+            )
+
+            try:
+                results = self.db_manager.search_memories(
+                    query=user_input, namespace=self.namespace, limit=5
+                )
+                logger.debug(
+                    f"Auto-ingest: Database search returned {len(results) if results else 0} results"
+                )
+
+                if results:
+                    for i, result in enumerate(
+                        results[:3]
+                    ):  # Log first 3 results for debugging
+                        logger.debug(
+                            f"Auto-ingest: Result {i+1}: {type(result)} with keys: {list(result.keys()) if isinstance(result, dict) else 'N/A'}"
+                        )
+            except Exception as db_search_e:
+                logger.error(f"Auto-ingest: Database search failed: {db_search_e}")
+                logger.debug(
+                    f"Auto-ingest: Database search error details: {type(db_search_e).__name__}: {str(db_search_e)}",
+                    exc_info=True,
+                )
+                results = []
+
+            if results:
+                logger.debug(
+                    f"Auto-ingest: Direct database search returned {len(results)} results"
+                )
+                # Add search metadata to results
+                for result in results:
+                    if isinstance(result, dict):
+                        result["retrieval_method"] = "direct_database_search"
+                        result["retrieval_query"] = user_input
+                return results
+
+            # If direct search fails, try search engine as backup
+            if self.search_engine:
+                logger.debug(
+                    "Auto-ingest: Direct search returned 0 results, trying search engine"
+                )
+                try:
+                    engine_results = self.search_engine.execute_search(
+                        query=user_input,
+                        db_manager=self.db_manager,
+                        namespace=self.namespace,
+                        limit=5,
+                    )
+
+                    if engine_results:
+                        logger.debug(
+                            f"Auto-ingest: Search engine returned {len(engine_results)} results"
+                        )
+                        # Add search metadata to results
+                        for result in engine_results:
+                            if isinstance(result, dict):
+                                result["retrieval_method"] = "search_engine"
+                                result["retrieval_query"] = user_input
+                        return engine_results
+                    else:
+                        logger.debug(
+                            "Auto-ingest: Search engine also returned 0 results"
+                        )
+
+                except Exception as search_error:
+                    logger.error(
+                        f"Auto-ingest: Search engine failed for query '{user_input[:50]}...': {search_error}"
+                    )
+                    logger.debug(
+                        f"Auto-ingest: Search engine error details: {type(search_error).__name__}: {str(search_error)}",
+                        exc_info=True,
+                    )
+            else:
+                logger.debug("Auto-ingest: No search engine available")
+
+            # Final fallback: get recent memories from the same namespace
+            logger.debug(
+                "Auto-ingest: All search methods returned 0 results, using recent memories fallback"
+            )
+            logger.debug(
+                f"Auto-ingest: Attempting fallback search in namespace '{self.namespace}'"
+            )
+
+            try:
+                fallback_results = self.db_manager.search_memories(
+                    query="",  # Empty query to get recent memories
+                    namespace=self.namespace,
+                    limit=3,
+                )
+                logger.debug(
+                    f"Auto-ingest: Fallback search returned {len(fallback_results) if fallback_results else 0} results"
+                )
+
+                if fallback_results:
+                    logger.debug(
+                        f"Auto-ingest: Fallback returned {len(fallback_results)} recent memories"
+                    )
+                    # Add search metadata to fallback results
+                    for result in fallback_results:
+                        if isinstance(result, dict):
+                            result["retrieval_method"] = "recent_memories_fallback"
+                            result["retrieval_query"] = user_input
+                    return fallback_results
+                else:
+                    logger.debug("Auto-ingest: Fallback search returned no results")
+
+            except Exception as fallback_e:
+                logger.error(f"Auto-ingest: Fallback search failed: {fallback_e}")
+                logger.debug(
+                    f"Auto-ingest: Fallback error details: {type(fallback_e).__name__}: {str(fallback_e)}",
+                    exc_info=True,
+                )
+
+            logger.debug(
+                "Auto-ingest: All retrieval methods failed, returning empty context"
+            )
+            return []
 
         except Exception as e:
-            logger.error(f"Failed to get auto-ingest context: {e}")
+            logger.error(
+                f"Auto-ingest: Failed to get context for '{user_input[:50]}...': {e}"
+            )
             return []
+        finally:
+            # Always clear recursion guard
+            if hasattr(self, "_in_context_retrieval"):
+                self._in_context_retrieval = False
 
     def _record_openai_conversation(self, kwargs, response):
-        """Record OpenAI conversation"""
+        """Record OpenAI conversation with enhanced content parsing"""
         try:
             messages = kwargs.get("messages", [])
             model = kwargs.get("model", "unknown")
 
-            # Extract user input
-            user_input = ""
+            # Extract user input with enhanced parsing
+            user_input = self._extract_openai_user_input(messages)
+
+            # Extract AI response with enhanced parsing
+            ai_output = self._extract_openai_ai_output(response)
+
+            # Calculate tokens
+            tokens_used = 0
+            if hasattr(response, "usage") and response.usage:
+                tokens_used = getattr(response.usage, "total_tokens", 0)
+
+            # Enhanced metadata extraction
+            metadata = self._extract_openai_metadata(kwargs, response, tokens_used)
+
+            # Record conversation
+            self.record_conversation(
+                user_input=user_input,
+                ai_output=ai_output,
+                model=model,
+                metadata=metadata,
+            )
+
+            # Also record AI response in conversation manager for history tracking
+            if ai_output:
+                self.conversation_manager.record_response(
+                    session_id=self._session_id,
+                    response=ai_output,
+                    metadata={"model": model, "tokens_used": tokens_used},
+                )
+        except Exception as e:
+            logger.error(f"Failed to record OpenAI conversation: {e}")
+
+    def _extract_openai_user_input(self, messages: List[Dict]) -> str:
+        """Extract user input from OpenAI messages with support for complex content types"""
+        user_input = ""
+        try:
+            # Find the last user message
             for message in reversed(messages):
                 if message.get("role") == "user":
-                    user_input = message.get("content", "")
+                    content = message.get("content", "")
+
+                    if isinstance(content, str):
+                        # Simple string content
+                        user_input = content
+                    elif isinstance(content, list):
+                        # Complex content (vision, multiple parts)
+                        text_parts = []
+                        image_count = 0
+
+                        for item in content:
+                            if isinstance(item, dict):
+                                if item.get("type") == "text":
+                                    text_parts.append(item.get("text", ""))
+                                elif item.get("type") == "image_url":
+                                    image_count += 1
+
+                        user_input = " ".join(text_parts)
+                        # Add image indicator if present
+                        if image_count > 0:
+                            user_input += f" [Contains {image_count} image(s)]"
+
                     break
+        except Exception as e:
+            logger.debug(f"Error extracting user input: {e}")
+            user_input = "[Error extracting user input]"
 
-            # Extract AI response
-            ai_output = ""
+        return user_input
+
+    def _extract_openai_ai_output(self, response) -> str:
+        """Extract AI output from OpenAI response with support for various response types"""
+        ai_output = ""
+        try:
             if hasattr(response, "choices") and response.choices:
                 choice = response.choices[0]
+
                 if hasattr(choice, "message") and choice.message:
-                    ai_output = choice.message.content or ""
+                    message = choice.message
+
+                    # Handle regular text content
+                    if hasattr(message, "content") and message.content:
+                        ai_output = message.content
+
+                    # Handle function/tool calls
+                    elif hasattr(message, "tool_calls") and message.tool_calls:
+                        tool_descriptions = []
+                        for tool_call in message.tool_calls:
+                            if hasattr(tool_call, "function"):
+                                func_name = tool_call.function.name
+                                func_args = tool_call.function.arguments
+                                tool_descriptions.append(
+                                    f"Called {func_name} with {func_args}"
+                                )
+                        ai_output = "[Tool calls: " + "; ".join(tool_descriptions) + "]"
+
+                    # Handle function calls (legacy format)
+                    elif hasattr(message, "function_call") and message.function_call:
+                        func_call = message.function_call
+                        func_name = func_call.get("name", "unknown")
+                        func_args = func_call.get("arguments", "{}")
+                        ai_output = f"[Function call: {func_name} with {func_args}]"
 
-            # Calculate tokens
-            tokens_used = 0
+                    else:
+                        ai_output = "[No content - possible function/tool call]"
+
+        except Exception as e:
+            logger.debug(f"Error extracting AI output: {e}")
+            ai_output = "[Error extracting AI response]"
+
+        return ai_output
+
+    def _extract_openai_metadata(
+        self, kwargs: Dict, response, tokens_used: int
+    ) -> Dict:
+        """Extract comprehensive metadata from OpenAI request and response"""
+        metadata = {
+            "integration": "openai_auto",
+            "api_type": "chat_completions",
+            "tokens_used": tokens_used,
+            "auto_recorded": True,
+        }
+
+        try:
+            # Add request metadata
+            if "temperature" in kwargs:
+                metadata["temperature"] = kwargs["temperature"]
+            if "max_tokens" in kwargs:
+                metadata["max_tokens"] = kwargs["max_tokens"]
+            if "tools" in kwargs:
+                metadata["has_tools"] = True
+                metadata["tool_count"] = len(kwargs["tools"])
+            if "functions" in kwargs:
+                metadata["has_functions"] = True
+                metadata["function_count"] = len(kwargs["functions"])
+
+            # Add response metadata
+            if hasattr(response, "choices") and response.choices:
+                choice = response.choices[0]
+                if hasattr(choice, "finish_reason"):
+                    metadata["finish_reason"] = choice.finish_reason
+
+            # Add detailed token usage if available
             if hasattr(response, "usage") and response.usage:
-                tokens_used = getattr(response.usage, "total_tokens", 0)
+                usage = response.usage
+                metadata.update(
+                    {
+                        "prompt_tokens": getattr(usage, "prompt_tokens", 0),
+                        "completion_tokens": getattr(usage, "completion_tokens", 0),
+                        "total_tokens": getattr(usage, "total_tokens", 0),
+                    }
+                )
+
+            # Detect content types
+            messages = kwargs.get("messages", [])
+            has_images = False
+            message_count = len(messages)
+
+            for message in messages:
+                if message.get("role") == "user":
+                    content = message.get("content")
+                    if isinstance(content, list):
+                        for item in content:
+                            if (
+                                isinstance(item, dict)
+                                and item.get("type") == "image_url"
+                            ):
+                                has_images = True
+                                break
+                    if has_images:
+                        break
+
+            metadata["message_count"] = message_count
+            metadata["has_images"] = has_images
+
+        except Exception as e:
+            logger.debug(f"Error extracting metadata: {e}")
+
+        return metadata
+
+    def _record_anthropic_conversation(self, kwargs, response):
+        """Record Anthropic conversation with enhanced content parsing"""
+        try:
+            messages = kwargs.get("messages", [])
+            model = kwargs.get("model", "claude-unknown")
+
+            # Extract user input with enhanced parsing
+            user_input = self._extract_anthropic_user_input(messages)
+
+            # Extract AI response with enhanced parsing
+            ai_output = self._extract_anthropic_ai_output(response)
+
+            # Calculate tokens
+            tokens_used = self._extract_anthropic_tokens(response)
+
+            # Enhanced metadata extraction
+            metadata = self._extract_anthropic_metadata(kwargs, response, tokens_used)
 
             # Record conversation
             self.record_conversation(
                 user_input=user_input,
                 ai_output=ai_output,
                 model=model,
-                metadata={
-                    "integration": "openai_auto",
-                    "api_type": "chat_completions",
-                    "tokens_used": tokens_used,
-                    "auto_recorded": True,
-                },
+                metadata=metadata,
             )
         except Exception as e:
-            logger.error(f"Failed to record OpenAI conversation: {e}")
+            logger.error(f"Failed to record Anthropic conversation: {e}")
 
-    def _record_anthropic_conversation(self, kwargs, response):
-        """Record Anthropic conversation"""
+    def _extract_anthropic_user_input(self, messages: List[Dict]) -> str:
+        """Extract user input from Anthropic messages with support for complex content types"""
+        user_input = ""
         try:
-            messages = kwargs.get("messages", [])
-            model = kwargs.get("model", "claude-unknown")
-
-            # Extract user input
-            user_input = ""
+            # Find the last user message
             for message in reversed(messages):
                 if message.get("role") == "user":
                     content = message.get("content", "")
-                    if isinstance(content, list):
-                        user_input = " ".join(
-                            [
-                                block.get("text", "")
-                                for block in content
-                                if isinstance(block, dict)
-                                and block.get("type") == "text"
-                            ]
-                        )
-                    else:
+
+                    if isinstance(content, str):
+                        # Simple string content
                         user_input = content
+                    elif isinstance(content, list):
+                        # Complex content (vision, multiple parts)
+                        text_parts = []
+                        image_count = 0
+
+                        for block in content:
+                            if isinstance(block, dict):
+                                if block.get("type") == "text":
+                                    text_parts.append(block.get("text", ""))
+                                elif block.get("type") == "image":
+                                    image_count += 1
+
+                        user_input = " ".join(text_parts)
+                        # Add image indicator if present
+                        if image_count > 0:
+                            user_input += f" [Contains {image_count} image(s)]"
+
                     break
+        except Exception as e:
+            logger.debug(f"Error extracting Anthropic user input: {e}")
+            user_input = "[Error extracting user input]"
 
-            # Extract AI response
-            ai_output = ""
+        return user_input
+
+    def _extract_anthropic_ai_output(self, response) -> str:
+        """Extract AI output from Anthropic response with support for various response types"""
+        ai_output = ""
+        try:
             if hasattr(response, "content") and response.content:
                 if isinstance(response.content, list):
-                    ai_output = " ".join(
-                        [
-                            block.text
-                            for block in response.content
-                            if hasattr(block, "text")
-                        ]
-                    )
+                    # Handle structured content (text blocks, tool use, etc.)
+                    text_parts = []
+                    tool_uses = []
+
+                    for block in response.content:
+                        try:
+                            # Handle text blocks
+                            if hasattr(block, "text") and block.text:
+                                text_parts.append(block.text)
+                            # Handle tool use blocks
+                            elif hasattr(block, "type"):
+                                block_type = getattr(block, "type", None)
+                                if block_type == "tool_use":
+                                    tool_name = getattr(block, "name", "unknown")
+                                    tool_input = getattr(block, "input", {})
+                                    tool_uses.append(
+                                        f"Used {tool_name} with {tool_input}"
+                                    )
+                            # Handle mock objects for testing (when type is accessible but not via hasattr)
+                            elif hasattr(block, "name") and hasattr(block, "input"):
+                                tool_name = getattr(block, "name", "unknown")
+                                tool_input = getattr(block, "input", {})
+                                tool_uses.append(f"Used {tool_name} with {tool_input}")
+                        except Exception as block_error:
+                            logger.debug(f"Error processing block: {block_error}")
+                            continue
+
+                    ai_output = " ".join(text_parts)
+                    if tool_uses:
+                        if ai_output:
+                            ai_output += " "
+                        ai_output += "[Tool uses: " + "; ".join(tool_uses) + "]"
+
+                elif isinstance(response.content, str):
+                    ai_output = response.content
                 else:
                     ai_output = str(response.content)
 
-            # Calculate tokens
-            tokens_used = 0
+        except Exception as e:
+            logger.debug(f"Error extracting Anthropic AI output: {e}")
+            ai_output = "[Error extracting AI response]"
+
+        return ai_output
+
+    def _extract_anthropic_tokens(self, response) -> int:
+        """Extract token usage from Anthropic response"""
+        tokens_used = 0
+        try:
             if hasattr(response, "usage") and response.usage:
                 input_tokens = getattr(response.usage, "input_tokens", 0)
                 output_tokens = getattr(response.usage, "output_tokens", 0)
                 tokens_used = input_tokens + output_tokens
+        except Exception as e:
+            logger.debug(f"Error extracting Anthropic tokens: {e}")
+
+        return tokens_used
+
+    def _extract_anthropic_metadata(
+        self, kwargs: Dict, response, tokens_used: int
+    ) -> Dict:
+        """Extract comprehensive metadata from Anthropic request and response"""
+        metadata = {
+            "integration": "anthropic_auto",
+            "api_type": "messages",
+            "tokens_used": tokens_used,
+            "auto_recorded": True,
+        }
+
+        try:
+            # Add request metadata
+            if "temperature" in kwargs:
+                metadata["temperature"] = kwargs["temperature"]
+            if "max_tokens" in kwargs:
+                metadata["max_tokens"] = kwargs["max_tokens"]
+            if "tools" in kwargs:
+                metadata["has_tools"] = True
+                metadata["tool_count"] = len(kwargs["tools"])
+
+            # Add response metadata
+            if hasattr(response, "stop_reason"):
+                metadata["stop_reason"] = response.stop_reason
+            if hasattr(response, "model"):
+                metadata["response_model"] = response.model
+
+            # Add detailed token usage if available
+            if hasattr(response, "usage") and response.usage:
+                usage = response.usage
+                metadata.update(
+                    {
+                        "input_tokens": getattr(usage, "input_tokens", 0),
+                        "output_tokens": getattr(usage, "output_tokens", 0),
+                        "total_tokens": tokens_used,
+                    }
+                )
+
+            # Detect content types
+            messages = kwargs.get("messages", [])
+            has_images = False
+            message_count = len(messages)
+
+            for message in messages:
+                if message.get("role") == "user":
+                    content = message.get("content")
+                    if isinstance(content, list):
+                        for item in content:
+                            if isinstance(item, dict) and item.get("type") == "image":
+                                has_images = True
+                                break
+                    if has_images:
+                        break
+
+            metadata["message_count"] = message_count
+            metadata["has_images"] = has_images
 
-            # Record conversation
-            self.record_conversation(
-                user_input=user_input,
-                ai_output=ai_output,
-                model=model,
-                metadata={
-                    "integration": "anthropic_auto",
-                    "api_type": "messages",
-                    "tokens_used": tokens_used,
-                    "auto_recorded": True,
-                },
-            )
         except Exception as e:
-            logger.error(f"Failed to record Anthropic conversation: {e}")
+            logger.debug(f"Error extracting Anthropic metadata: {e}")
 
-    def _litellm_success_callback(self, kwargs, response, start_time, end_time):
-        """
-        This function is automatically called by LiteLLM after a successful completion.
-        """
+        return metadata
+
+    def _process_litellm_response(self, kwargs, response, start_time, end_time):
+        """Process and record LiteLLM response"""
         try:
+            # Extract user input from messages
+            messages = kwargs.get("messages", [])
             user_input = ""
-            # Find the last user message
-            for msg in reversed(kwargs.get("messages", [])):
-                if msg.get("role") == "user":
-                    user_input = msg.get("content", "")
+
+            for message in reversed(messages):
+                if message.get("role") == "user":
+                    user_input = message.get("content", "")
                     break
 
-            ai_output = response.choices[0].message.content or ""
-            model = kwargs.get("model", "unknown")
+            # Extract AI output from response
+            ai_output = ""
+            if hasattr(response, "choices") and response.choices:
+                choice = response.choices[0]
+                if hasattr(choice, "message") and hasattr(choice.message, "content"):
+                    ai_output = choice.message.content or ""
+                elif hasattr(choice, "text"):
+                    ai_output = choice.text or ""
+
+            # Extract model
+            model = kwargs.get("model", "litellm-unknown")
+
+            # Calculate timing (convert to seconds for JSON serialization)
+            duration_seconds = (end_time - start_time) if start_time and end_time else 0
+            if hasattr(duration_seconds, "total_seconds"):
+                duration_seconds = duration_seconds.total_seconds()
+
+            # Prepare metadata
+            metadata = {
+                "integration": "litellm",
+                "auto_recorded": True,
+                "duration": float(duration_seconds),
+                "timestamp": time.time(),
+            }
 
-            # Calculate tokens used
-            tokens_used = 0
+            # Add token usage if available
             if hasattr(response, "usage") and response.usage:
-                tokens_used = getattr(response.usage, "total_tokens", 0)
+                usage = response.usage
+                metadata.update(
+                    {
+                        "prompt_tokens": getattr(usage, "prompt_tokens", 0),
+                        "completion_tokens": getattr(usage, "completion_tokens", 0),
+                        "total_tokens": getattr(usage, "total_tokens", 0),
+                    }
+                )
 
-            # Handle timing data safely - convert any time objects to float/string
-            duration_ms = 0
-            start_time_str = None
-            end_time_str = None
+            # Record the conversation
+            if user_input and ai_output:
+                self.record_conversation(
+                    user_input=user_input,
+                    ai_output=ai_output,
+                    model=model,
+                    metadata=metadata,
+                )
 
-            try:
-                if start_time is not None and end_time is not None:
-                    # Handle different types of time objects
-                    if hasattr(start_time, "total_seconds"):  # timedelta
-                        duration_ms = start_time.total_seconds() * 1000
-                    elif isinstance(start_time, (int, float)) and isinstance(
-                        end_time, (int, float)
-                    ):
-                        duration_ms = (end_time - start_time) * 1000
+        except Exception as e:
+            logger.error(f"Failed to process LiteLLM response: {e}")
 
-                    start_time_str = str(start_time)
-                    end_time_str = str(end_time)
-            except Exception:
-                # If timing calculation fails, just skip it
-                pass
+    # LiteLLM callback is now handled by the LiteLLMCallbackManager
+    # in memori.integrations.litellm_integration
 
-            self.record_conversation(
-                user_input,
-                ai_output,
-                model,
-                metadata={
-                    "integration": "litellm",
-                    "api_type": "completion",
-                    "tokens_used": tokens_used,
-                    "auto_recorded": True,
-                    "start_time_str": start_time_str,
-                    "end_time_str": end_time_str,
-                    "duration_ms": duration_ms,
-                },
+    def _process_memory_sync(
+        self, chat_id: str, user_input: str, ai_output: str, model: str = "unknown"
+    ):
+        """Synchronous memory processing fallback"""
+        if not self.memory_agent:
+            logger.warning("Memory agent not available, skipping memory ingestion")
+            return
+
+        try:
+            # Run async processing in new event loop
+            import threading
+
+            def run_memory_processing():
+                new_loop = asyncio.new_event_loop()
+                asyncio.set_event_loop(new_loop)
+                try:
+                    new_loop.run_until_complete(
+                        self._process_memory_async(
+                            chat_id, user_input, ai_output, model
+                        )
+                    )
+                except Exception as e:
+                    logger.error(f"Synchronous memory processing failed: {e}")
+                finally:
+                    new_loop.close()
+
+            # Run in background thread to avoid blocking
+            thread = threading.Thread(target=run_memory_processing, daemon=True)
+            thread.start()
+            logger.debug(
+                f"Memory processing started in background thread for {chat_id}"
             )
+
         except Exception as e:
-            logger.error(f"Memori callback failed: {e}")
+            logger.error(f"Failed to start synchronous memory processing: {e}")
+
+    def _parse_llm_response(self, response) -> tuple[str, str]:
+        """Extract text and model from various LLM response formats."""
+        if response is None:
+            return "", "unknown"
+
+        # String response
+        if isinstance(response, str):
+            return response, "unknown"
+
+        # Anthropic response
+        if hasattr(response, "content"):
+            text = ""
+            if isinstance(response.content, list):
+                text = "".join(b.text for b in response.content if hasattr(b, "text"))
+            else:
+                text = str(response.content)
+            return text, getattr(response, "model", "unknown")
+
+        # OpenAI response
+        if hasattr(response, "choices") and response.choices:
+            choice = response.choices[0]
+            text = (
+                getattr(choice.message, "content", "")
+                if hasattr(choice, "message")
+                else getattr(choice, "text", "")
+            )
+            return text or "", getattr(response, "model", "unknown")
+
+        # Dict response
+        if isinstance(response, dict):
+            return response.get(
+                "content", response.get("text", str(response))
+            ), response.get("model", "unknown")
+
+        # Fallback
+        return str(response), "unknown"
 
     def record_conversation(
         self,
         user_input: str,
-        ai_output: str,
-        model: str = "unknown",
+        ai_output=None,
+        model: str = None,
         metadata: Optional[Dict[str, Any]] = None,
     ) -> str:
         """
-        Manually record a conversation
+        Record a conversation.
 
         Args:
-            user_input: The user's input message
-            ai_output: The AI's response
-            model: Model used for the response
-            metadata: Additional metadata
+            user_input: User's message
+            ai_output: AI response (any format)
+            model: Optional model name override
+            metadata: Optional metadata
 
         Returns:
-            chat_id: Unique identifier for this conversation
+            chat_id: Unique conversation ID
         """
         if not self._enabled:
             raise MemoriError("Memori is not enabled. Call enable() first.")
 
-        # Ensure ai_output is never None to avoid NOT NULL constraint errors
-        if ai_output is None:
-            ai_output = ""
+        # Parse response
+        response_text, detected_model = self._parse_llm_response(ai_output)
+        response_model = model or detected_model
 
+        # Generate ID and timestamp
         chat_id = str(uuid.uuid4())
         timestamp = datetime.now()
 
-        try:
-            # Store in chat history
-            self.db_manager.store_chat_history(
-                chat_id=chat_id,
-                user_input=user_input,
-                ai_output=ai_output,
-                model=model,
-                timestamp=timestamp,
-                session_id=self._session_id,
-                namespace=self.namespace,
-                metadata=metadata or {},
-            )
+        # Store conversation
+        self.db_manager.store_chat_history(
+            chat_id=chat_id,
+            user_input=user_input,
+            ai_output=response_text,
+            model=response_model,
+            timestamp=timestamp,
+            session_id=self._session_id,
+            namespace=self.namespace,
+            metadata=metadata or {},
+        )
 
-            # Process for memory categorization
-            if self.conscious_ingest:
-                self._process_memory_ingestion(chat_id, user_input, ai_output, model)
+        # Always process into long-term memory when memory agent is available
+        if self.memory_agent:
+            self._schedule_memory_processing(
+                chat_id, user_input, response_text, response_model
+            )
 
-            logger.debug(f"Conversation recorded: {chat_id}")
-            return chat_id
+        logger.debug(f"Recorded conversation: {chat_id}")
+        return chat_id
 
-        except Exception as e:
-            raise MemoriError(f"Failed to record conversation: {e}")
+    def _schedule_memory_processing(
+        self, chat_id: str, user_input: str, ai_output: str, model: str
+    ):
+        """Schedule memory processing (async if possible, sync fallback)."""
+        try:
+            loop = asyncio.get_running_loop()
+            task = loop.create_task(
+                self._process_memory_async(chat_id, user_input, ai_output, model)
+            )
 
-    def _process_memory_ingestion(
+            # Prevent garbage collection
+            if not hasattr(self, "_memory_tasks"):
+                self._memory_tasks = set()
+            self._memory_tasks.add(task)
+            task.add_done_callback(self._memory_tasks.discard)
+        except RuntimeError:
+            # No event loop, use sync fallback
+            logger.debug("No event loop, using synchronous memory processing")
+            self._process_memory_sync(chat_id, user_input, ai_output, model)
+
+    async def _process_memory_async(
         self, chat_id: str, user_input: str, ai_output: str, model: str = "unknown"
     ):
-        """Process conversation for Pydantic-based memory categorization"""
+        """Process conversation with enhanced async memory categorization"""
         if not self.memory_agent:
             logger.warning("Memory agent not available, skipping memory ingestion")
             return
@@ -988,38 +1771,107 @@ class Memori:
                 relevant_skills=self._user_context.get("relevant_skills", []),
             )
 
-            # Process conversation using Pydantic-based memory agent
-            processed_memory = self.memory_agent.process_conversation_sync(
+            # Get recent memories for deduplication
+            existing_memories = await self._get_recent_memories_for_dedup()
+
+            # Process conversation using async Pydantic-based memory agent
+            processed_memory = await self.memory_agent.process_conversation_async(
                 chat_id=chat_id,
                 user_input=user_input,
                 ai_output=ai_output,
                 context=context,
-                mem_prompt=self.mem_prompt,
-                filters=self.memory_filters,
+                existing_memories=(
+                    [mem.summary for mem in existing_memories[:10]]
+                    if existing_memories
+                    else []
+                ),
             )
 
-            # Store processed memory with entity indexing
-            if processed_memory.should_store:
-                memory_id = self.db_manager.store_processed_memory(
-                    memory=processed_memory, chat_id=chat_id, namespace=self.namespace
-                )
+            # Check for duplicates
+            duplicate_id = await self.memory_agent.detect_duplicates(
+                processed_memory, existing_memories
+            )
 
-                if memory_id:
-                    logger.debug(
-                        f"Stored processed memory {memory_id} for chat {chat_id}"
-                    )
-                else:
-                    logger.debug(
-                        f"Memory not stored for chat {chat_id}: {processed_memory.storage_reasoning}"
+            if duplicate_id:
+                processed_memory.duplicate_of = duplicate_id
+                logger.info(f"Memory marked as duplicate of {duplicate_id}")
+
+            # Apply filters
+            if self.memory_agent.should_filter_memory(
+                processed_memory, self.memory_filters
+            ):
+                logger.debug(f"Memory filtered out for chat {chat_id}")
+                return
+
+            # Store processed memory with new schema
+            memory_id = self.db_manager.store_long_term_memory_enhanced(
+                processed_memory, chat_id, self.namespace
+            )
+
+            if memory_id:
+                logger.debug(f"Stored processed memory {memory_id} for chat {chat_id}")
+
+                # Check for conscious context updates if promotion eligible and conscious_ingest enabled
+                if (
+                    processed_memory.promotion_eligible
+                    and self.conscious_agent
+                    and self.conscious_ingest
+                ):
+                    await self.conscious_agent.check_for_context_updates(
+                        self.db_manager, self.namespace
                     )
             else:
-                logger.debug(
-                    f"Memory not stored for chat {chat_id}: {processed_memory.storage_reasoning}"
-                )
+                logger.warning(f"Failed to store memory for chat {chat_id}")
 
         except Exception as e:
             logger.error(f"Memory ingestion failed for {chat_id}: {e}")
 
+    async def _get_recent_memories_for_dedup(self) -> List:
+        """Get recent memories for deduplication check"""
+        try:
+            from sqlalchemy import text
+
+            from ..database.queries.memory_queries import MemoryQueries
+            from ..utils.pydantic_models import ProcessedLongTermMemory
+
+            with self.db_manager._get_connection() as connection:
+                result = connection.execute(
+                    text(MemoryQueries.SELECT_MEMORIES_FOR_DEDUPLICATION),
+                    {
+                        "namespace": self.namespace,
+                        "processed_for_duplicates": False,
+                        "limit": 20,
+                    },
+                )
+
+                memories = []
+                for row in result:
+                    try:
+                        # Create ProcessedLongTermMemory objects for proper comparison
+                        # Note: Query returns (memory_id, summary, searchable_content, classification, created_at)
+                        memory = ProcessedLongTermMemory(
+                            conversation_id=row[
+                                0
+                            ],  # Use memory_id as conversation_id for existing memories
+                            summary=row[1] or "",
+                            content=row[2] or "",
+                            classification=row[3] or "conversational",
+                            importance="medium",  # Default importance level for comparison
+                            promotion_eligible=False,  # Default for existing memories
+                            classification_reason="Existing memory loaded for deduplication check",  # Required field
+                        )
+                        memories.append(memory)
+                    except Exception as e:
+                        # Silently skip malformed memories from old data format
+                        logger.debug(f"Skipping malformed memory during dedup: {e}")
+                        continue
+
+                return memories
+
+        except Exception as e:
+            logger.error(f"Failed to get recent memories for dedup: {e}")
+            return []
+
     def retrieve_context(self, query: str, limit: int = 5) -> List[Dict[str, Any]]:
         """
         Retrieve relevant context for a query with priority on essential facts
@@ -1123,10 +1975,13 @@ class Memori:
         return self._session_id
 
     def get_integration_stats(self) -> List[Dict[str, Any]]:
-        """Get statistics from the universal integration system"""
+        """Get statistics from the new interceptor system"""
         try:
+            # Get system status first
+            interceptor_status = self.get_interceptor_status()
+
             stats = {
-                "integration": "universal_auto_recording",
+                "integration": "memori_system",
                 "enabled": self._enabled,
                 "session_id": self._session_id,
                 "namespace": self.namespace,
@@ -1134,58 +1989,62 @@ class Memori:
             }
 
             # LiteLLM stats
+            litellm_interceptor_status = interceptor_status.get("native", {})
             if LITELLM_AVAILABLE:
                 stats["providers"]["litellm"] = {
                     "available": True,
                     "method": "native_callbacks",
-                    "callback_registered": self._enabled,
-                    "callbacks_count": len(success_callback) if self._enabled else 0,
+                    "enabled": litellm_interceptor_status.get("enabled", False),
+                    "status": litellm_interceptor_status.get("status", "unknown"),
                 }
             else:
                 stats["providers"]["litellm"] = {
                     "available": False,
                     "method": "native_callbacks",
-                    "callback_registered": False,
+                    "enabled": False,
                 }
 
+            # Get interceptor status instead of checking wrapped attributes
+            interceptor_status = self.get_interceptor_status()
+
             # OpenAI stats
             try:
                 import openai
 
+                _ = openai  # Suppress unused import warning
+
+                openai_interceptor_status = interceptor_status.get("openai", {})
                 stats["providers"]["openai"] = {
                     "available": True,
-                    "method": "auto_wrapping",
-                    "wrapped": (
-                        hasattr(openai.OpenAI, "_memori_wrapped")
-                        if hasattr(openai, "OpenAI")
-                        else False
-                    ),
+                    "method": "litellm_native",
+                    "enabled": openai_interceptor_status.get("enabled", False),
+                    "status": openai_interceptor_status.get("status", "unknown"),
                 }
             except ImportError:
                 stats["providers"]["openai"] = {
                     "available": False,
-                    "method": "auto_wrapping",
-                    "wrapped": False,
+                    "method": "litellm_native",
+                    "enabled": False,
                 }
 
             # Anthropic stats
             try:
                 import anthropic
 
+                _ = anthropic  # Suppress unused import warning
+
+                anthropic_interceptor_status = interceptor_status.get("anthropic", {})
                 stats["providers"]["anthropic"] = {
                     "available": True,
-                    "method": "auto_wrapping",
-                    "wrapped": (
-                        hasattr(anthropic.Anthropic, "_memori_wrapped")
-                        if hasattr(anthropic, "Anthropic")
-                        else False
-                    ),
+                    "method": "litellm_native",
+                    "enabled": anthropic_interceptor_status.get("enabled", False),
+                    "status": anthropic_interceptor_status.get("status", "unknown"),
                 }
             except ImportError:
                 stats["providers"]["anthropic"] = {
                     "available": False,
-                    "method": "auto_wrapping",
-                    "wrapped": False,
+                    "method": "litellm_native",
+                    "enabled": False,
                 }
 
             return [stats]
@@ -1230,7 +2089,7 @@ class Memori:
         """Get memories that contain a specific entity"""
         try:
             # This would use the entity index in the database
-            # For now, use keyword search as fallback
+            # For now, use keyword search as fallback (entity_type is ignored for now)
             return self.db_manager.search_memories(
                 query=entity_value, namespace=self.namespace, limit=limit
             )
@@ -1269,11 +2128,25 @@ class Memori:
 
             # If we have a running loop, schedule the task
             self._background_task = loop.create_task(self._background_analysis_loop())
+            # Add proper error handling callback
+            self._background_task.add_done_callback(
+                self._handle_background_task_completion
+            )
             logger.info("Background analysis task started")
 
         except Exception as e:
             logger.error(f"Failed to start background analysis: {e}")
 
+    def _handle_background_task_completion(self, task):
+        """Handle background task completion and cleanup"""
+        try:
+            if task.exception():
+                logger.error(f"Background task failed: {task.exception()}")
+        except asyncio.CancelledError:
+            logger.debug("Background task was cancelled")
+        except Exception as e:
+            logger.error(f"Error handling background task completion: {e}")
+
     def _stop_background_analysis(self):
         """Stop the background analysis task"""
         try:
@@ -1283,32 +2156,66 @@ class Memori:
         except Exception as e:
             logger.error(f"Failed to stop background analysis: {e}")
 
+    def cleanup(self):
+        """Clean up all async tasks and resources"""
+        try:
+            # Cancel background tasks
+            self._stop_background_analysis()
+
+            # Clean up memory processing tasks
+            if hasattr(self, "_memory_tasks"):
+                for task in self._memory_tasks.copy():
+                    if not task.done():
+                        task.cancel()
+                self._memory_tasks.clear()
+
+            logger.debug("Memori cleanup completed")
+        except Exception as e:
+            logger.error(f"Error during cleanup: {e}")
+
+    def __del__(self):
+        """Destructor to ensure cleanup"""
+        try:
+            self.cleanup()
+        except:
+            pass  # Ignore errors during destruction
+
     async def _background_analysis_loop(self):
-        """Main background analysis loop"""
-        logger.info("ConsciouscAgent: Background analysis loop started")
+        """Background analysis loop for memory processing"""
+        try:
+            logger.debug("Background analysis loop started")
 
-        while self._enabled and self.conscious_ingest:
-            try:
-                if self.conscious_agent and self.conscious_agent.should_run_analysis():
-                    await self.conscious_agent.run_background_analysis(
-                        self.db_manager, self.namespace
-                    )
+            # For now, just run periodic conscious ingestion if enabled
+            if self.conscious_ingest and self.conscious_agent:
+                while True:
+                    try:
+                        await asyncio.sleep(300)  # Check every 5 minutes
 
-                # Wait 30 minutes before next check
-                await asyncio.sleep(1800)  # 30 minutes
+                        # Run conscious ingestion to check for new promotable memories
+                        await self.conscious_agent.run_conscious_ingest(
+                            self.db_manager, self.namespace
+                        )
+
+                        logger.debug("Periodic conscious analysis completed")
 
-            except asyncio.CancelledError:
-                logger.info("ConsciouscAgent: Background analysis cancelled")
-                break
-            except Exception as e:
-                logger.error(f"ConsciouscAgent: Background analysis error: {e}")
-                # Wait 5 minutes before retrying on error
-                await asyncio.sleep(300)
+                    except asyncio.CancelledError:
+                        logger.debug("Background analysis loop cancelled")
+                        break
+                    except Exception as e:
+                        logger.error(f"Background analysis error: {e}")
+                        await asyncio.sleep(60)  # Wait 1 minute before retry
+            else:
+                # If not using conscious ingest, just sleep
+                while True:
+                    await asyncio.sleep(3600)  # Sleep for 1 hour
 
-        logger.info("ConsciouscAgent: Background analysis loop ended")
+        except asyncio.CancelledError:
+            logger.debug("Background analysis loop cancelled")
+        except Exception as e:
+            logger.error(f"Background analysis loop failed: {e}")
 
     def trigger_conscious_analysis(self):
-        """Manually trigger conscious agent analysis (for testing/immediate analysis)"""
+        """Manually trigger conscious context ingestion (for testing/immediate analysis)"""
         if not self.conscious_ingest or not self.conscious_agent:
             logger.warning("Conscious ingestion not enabled or agent not available")
             return
@@ -1318,11 +2225,11 @@ class Memori:
             try:
                 loop = asyncio.get_running_loop()
                 task = loop.create_task(
-                    self.conscious_agent.run_background_analysis(
+                    self.conscious_agent.run_conscious_ingest(
                         self.db_manager, self.namespace
                     )
                 )
-                logger.info("Conscious analysis triggered")
+                logger.info("Conscious context ingestion triggered")
                 return task
             except RuntimeError:
                 # No event loop, run synchronously in thread
@@ -1333,7 +2240,7 @@ class Memori:
                     asyncio.set_event_loop(new_loop)
                     try:
                         new_loop.run_until_complete(
-                            self.conscious_agent.run_background_analysis(
+                            self.conscious_agent.run_conscious_ingest(
                                 self.db_manager, self.namespace
                             )
                         )
@@ -1342,29 +2249,170 @@ class Memori:
 
                 thread = threading.Thread(target=run_analysis)
                 thread.start()
-                logger.info("Conscious analysis triggered in separate thread")
+                logger.info("Conscious context ingestion triggered in separate thread")
+
+        except Exception as e:
+            logger.error(f"Failed to trigger conscious context ingestion: {e}")
+
+    def get_conscious_system_prompt(self) -> str:
+        """
+        Get conscious context as system prompt for direct injection.
+        Returns ALL short-term memory as formatted system prompt.
+        Use this for conscious_ingest mode.
+        """
+        try:
+            context = self._get_conscious_context()
+            if not context:
+                return ""
+
+            # Create system prompt with all short-term memory
+            system_prompt = "--- Your Short-Term Memory (Conscious Context) ---\n"
+            system_prompt += "This is your complete working memory. USE THIS INFORMATION TO ANSWER QUESTIONS:\n\n"
+
+            # Deduplicate and format context
+            seen_content = set()
+            for mem in context:
+                if isinstance(mem, dict):
+                    content = mem.get("searchable_content", "") or mem.get(
+                        "summary", ""
+                    )
+                    category = mem.get("category_primary", "")
+
+                    # Skip duplicates
+                    content_key = content.lower().strip()
+                    if content_key in seen_content:
+                        continue
+                    seen_content.add(content_key)
+
+                    system_prompt += f"[{category.upper()}] {content}\n"
+
+            system_prompt += "\nIMPORTANT: Use the above information to answer questions about the user.\n"
+            system_prompt += "-------------------------\n"
+
+            return system_prompt
+
+        except Exception as e:
+            logger.error(f"Failed to generate conscious system prompt: {e}")
+            return ""
+
+    def get_auto_ingest_system_prompt(self, user_input: str) -> str:
+        """
+        Get auto-ingest context as system prompt for direct injection.
+        Returns relevant memories based on user input as formatted system prompt.
+        Use this for auto_ingest mode.
+        """
+        try:
+            # For now, use recent short-term memories as a simple approach
+            # This avoids the search engine issues and still provides context
+            # TODO: Use user_input for intelligent context retrieval
+            context = self._get_conscious_context()  # Get recent short-term memories
+
+            if not context:
+                return ""
+
+            # Create system prompt with relevant memories (limited to prevent overwhelming)
+            system_prompt = "--- Relevant Memory Context ---\n"
+
+            # Take first 5 items to avoid too much context
+            seen_content = set()
+            for mem in context[:5]:
+                if isinstance(mem, dict):
+                    content = mem.get("searchable_content", "") or mem.get(
+                        "summary", ""
+                    )
+                    category = mem.get("category_primary", "")
+
+                    # Skip duplicates
+                    content_key = content.lower().strip()
+                    if content_key in seen_content:
+                        continue
+                    seen_content.add(content_key)
+
+                    if category.startswith("essential_"):
+                        system_prompt += f"[{category.upper()}] {content}\n"
+                    else:
+                        system_prompt += f"- {content}\n"
+
+            system_prompt += "-------------------------\n"
+
+            return system_prompt
 
         except Exception as e:
-            logger.error(f"Failed to trigger conscious analysis: {e}")
+            logger.error(f"Failed to generate auto-ingest system prompt: {e}")
+            return ""
+
+    def add_memory_to_messages(self, messages: list, user_input: str = None) -> list:
+        """
+        Add appropriate memory context to messages based on ingest mode.
+
+        Args:
+            messages: List of messages for LLM
+            user_input: User input for auto_ingest context retrieval (optional)
+
+        Returns:
+            Modified messages list with memory context added as system message
+        """
+        try:
+            system_prompt = ""
+
+            if self.conscious_ingest:
+                # One-time conscious context injection
+                if not self._conscious_context_injected:
+                    system_prompt = self.get_conscious_system_prompt()
+                    self._conscious_context_injected = True
+                    logger.info(
+                        "Conscious-ingest: Added complete working memory to system prompt"
+                    )
+                else:
+                    logger.debug("Conscious-ingest: Context already injected, skipping")
+
+            elif self.auto_ingest and user_input:
+                # Dynamic auto-ingest based on user input
+                system_prompt = self.get_auto_ingest_system_prompt(user_input)
+                logger.debug("Auto-ingest: Added relevant context to system prompt")
+
+            if system_prompt:
+                # Add to existing system message or create new one
+                messages_copy = messages.copy()
+
+                # Check if system message already exists
+                for msg in messages_copy:
+                    if msg.get("role") == "system":
+                        msg["content"] = system_prompt + "\n" + msg.get("content", "")
+                        return messages_copy
+
+                # No system message exists, add one at the beginning
+                messages_copy.insert(0, {"role": "system", "content": system_prompt})
+                return messages_copy
+
+            return messages
+
+        except Exception as e:
+            logger.error(f"Failed to add memory to messages: {e}")
+            return messages
 
     def get_essential_conversations(self, limit: int = 10) -> List[Dict[str, Any]]:
         """Get essential conversations from short-term memory"""
         try:
+            from sqlalchemy import text
+
             # Get all conversations marked as essential
             with self.db_manager._get_connection() as connection:
                 query = """
                 SELECT memory_id, summary, category_primary, importance_score,
                        created_at, searchable_content, processed_data
                 FROM short_term_memory
-                WHERE namespace = ? AND category_primary LIKE 'essential_%'
+                WHERE namespace = :namespace AND category_primary LIKE 'essential_%'
                 ORDER BY importance_score DESC, created_at DESC
-                LIMIT ?
+                LIMIT :limit
                 """
 
-                cursor = connection.execute(query, (self.namespace, limit))
+                result = connection.execute(
+                    text(query), {"namespace": self.namespace, "limit": limit}
+                )
 
                 essential_conversations = []
-                for row in cursor.fetchall():
+                for row in result:
                     essential_conversations.append(
                         {
                             "memory_id": row[0],
@@ -1382,3 +2430,106 @@ class Memori:
         except Exception as e:
             logger.error(f"Failed to get essential conversations: {e}")
             return []
+
+    def create_openai_client(self, **kwargs):
+        """
+        Create an OpenAI client with automatic memory recording.
+
+        This method creates a MemoriOpenAIInterceptor that automatically records
+        all OpenAI API calls to memory using the inheritance-based approach.
+
+        Args:
+            **kwargs: Additional arguments passed to OpenAI client (e.g., api_key)
+                     These override any settings from the Memori provider config
+
+        Returns:
+            MemoriOpenAIInterceptor instance that works as a drop-in replacement
+            for the standard OpenAI client
+
+        Example:
+            memori = Memori(api_key="sk-...")
+            memori.enable()
+
+            # Create interceptor client
+            client = memori.create_openai_client()
+
+            # Use exactly like standard OpenAI client
+            response = client.chat.completions.create(
+                model="gpt-4o",
+                messages=[{"role": "user", "content": "Hello!"}]
+            )
+            # Conversation is automatically recorded
+        """
+        try:
+            from ..integrations.openai_integration import create_openai_client
+
+            return create_openai_client(self, self.provider_config, **kwargs)
+        except ImportError as e:
+            logger.error(f"Failed to import OpenAI integration: {e}")
+            raise ImportError(
+                "OpenAI integration not available. Install with: pip install openai"
+            ) from e
+
+    def create_openai_wrapper(self, **kwargs):
+        """
+        Create a legacy OpenAI wrapper (backward compatibility).
+
+        DEPRECATED: Use create_openai_client() instead for better integration.
+
+        Returns:
+            MemoriOpenAI wrapper instance
+        """
+        try:
+            from ..integrations.openai_integration import MemoriOpenAI
+
+            return MemoriOpenAI(self, **kwargs)
+        except ImportError as e:
+            logger.error(f"Failed to import OpenAI integration: {e}")
+            raise ImportError(
+                "OpenAI integration not available. Install with: pip install openai"
+            ) from e
+
+    # Conversation management methods
+
+    def get_conversation_stats(self) -> Dict[str, Any]:
+        """Get conversation manager statistics"""
+        return self.conversation_manager.get_session_stats()
+
+    def clear_conversation_history(self, session_id: str = None):
+        """
+        Clear conversation history
+
+        Args:
+            session_id: Specific session to clear. If None, clears current session.
+        """
+        if session_id is None:
+            session_id = self._session_id
+        self.conversation_manager.clear_session(session_id)
+        logger.info(f"Cleared conversation history for session: {session_id}")
+
+    def clear_all_conversations(self):
+        """Clear all conversation histories"""
+        self.conversation_manager.clear_all_sessions()
+        logger.info("Cleared all conversation histories")
+
+    def start_new_conversation(self) -> str:
+        """
+        Start a new conversation session
+
+        Returns:
+            New session ID
+        """
+        old_session_id = self._session_id
+        self._session_id = str(uuid.uuid4())
+
+        # Reset conscious context injection flag for new conversation
+        self._conscious_context_injected = False
+
+        logger.info(
+            f"Started new conversation: {self._session_id} (previous: {old_session_id})"
+        )
+        return self._session_id
+
+    def get_current_session_id(self) -> str:
+        """Get current conversation session ID"""
+        return self._session_id