Update memory_logic.py

memory_logic.py

@@ -1,163 +1,508 @@
+# memory_logic.py
 import os
 import json
-import logging
+import time
 from datetime import datetime
-import re # For insight format validation
+import logging
+import re
+import threading
+
+# Conditionally import heavy dependencies
+try:
+    from sentence_transformers import SentenceTransformer
+    import faiss
+    import numpy as np
+except ImportError:
+    SentenceTransformer, faiss, np = None, None, None
+    logging.warning("SentenceTransformers, FAISS, or NumPy not installed. Semantic search will be unavailable.")
+
+try:
+    import sqlite3
+except ImportError:
+    sqlite3 = None
+    logging.warning("sqlite3 module not available. SQLite backend will be unavailable.")
+
+try:
+    from datasets import load_dataset, Dataset
+except ImportError:
+    load_dataset, Dataset = None, None
+    logging.warning("datasets library not installed. Hugging Face Dataset backend will be unavailable.")
+
 
 logger = logging.getLogger(__name__)
+# Suppress verbose logs from dependencies
+for lib_name in ["sentence_transformers", "faiss", "datasets", "huggingface_hub"]:
+    if logging.getLogger(lib_name): # Check if logger exists
+        logging.getLogger(lib_name).setLevel(logging.WARNING)
+
 
-DATA_DIR = "app_data"
-MEMORIES_FILE = os.path.join(DATA_DIR, "conversation_memories.jsonl") # JSON Lines format
-RULES_FILE = os.path.join(DATA_DIR, "learned_rules.jsonl") # Rules/Insights, also JSON Lines
+# --- Configuration (Read directly from environment variables) ---
+STORAGE_BACKEND = os.getenv("STORAGE_BACKEND", "RAM").upper()
+SQLITE_DB_PATH = os.getenv("SQLITE_DB_PATH", "app_data/ai_memory.db") # Changed default path
+HF_TOKEN = os.getenv("HF_TOKEN")
+HF_MEMORY_DATASET_REPO = os.getenv("HF_MEMORY_DATASET_REPO", "your_hf_username/app-memories") # Example
+HF_RULES_DATASET_REPO = os.getenv("HF_RULES_DATASET_REPO", "your_hf_username/app-rules")     # Example
 
-# Ensure data directory exists
-os.makedirs(DATA_DIR, exist_ok=True)
+# --- Globals for RAG within this module ---
+_embedder = None
+_dimension = 384 # Default, will be set by embedder
+_faiss_memory_index = None
+_memory_items_list = []  # Stores JSON strings of memory objects for RAM, or loaded from DB/HF
+_faiss_rules_index = None
+_rules_items_list = []   # Stores rule text strings
 
-# --- Rules/Insights Management ---
+_initialized = False
+_init_lock = threading.Lock()
 
-def load_rules_from_file() -> list[str]:
-    """Loads rules (insights) from the JSON Lines file."""
-    rules = []
-    if not os.path.exists(RULES_FILE):
-        return rules
+# --- Helper: SQLite Connection ---
+def _get_sqlite_connection():
+    if not sqlite3:
+        raise ImportError("sqlite3 module is required for SQLite backend but not found.")
+    db_dir = os.path.dirname(SQLITE_DB_PATH)
+    if db_dir and not os.path.exists(db_dir):
+        os.makedirs(db_dir, exist_ok=True)
+    return sqlite3.connect(SQLITE_DB_PATH, timeout=10) # Added timeout
+
+def _init_sqlite_tables():
+    if STORAGE_BACKEND != "SQLITE" or not sqlite3:
+        return
     try:
-        with open(RULES_FILE, 'r', encoding='utf-8') as f:
-            for line in f:
-                if line.strip():
-                    try:
-                        # Assuming each line is a JSON object like {"rule_text": "...", "timestamp": "..."}
-                        # For simplicity, if we only stored the text previously, adapt here.
-                        # Let's assume we store {"text": "rule_text_content"}
-                        data = json.loads(line)
-                        if "text" in data and isinstance(data["text"], str) and data["text"].strip():
-                             rules.append(data["text"].strip())
-                        elif isinstance(data, str): # If old format was just text per line
-                             rules.append(data.strip())
-
-                    except json.JSONDecodeError:
-                        logger.warning(f"Skipping malformed JSON line in rules file: {line.strip()}")
-        logger.info(f"Loaded {len(rules)} rules from {RULES_FILE}")
+        with _get_sqlite_connection() as conn:
+            cursor = conn.cursor()
+            # Stores JSON string of the memory object
+            cursor.execute("""
+                CREATE TABLE IF NOT EXISTS memories (
+                    id INTEGER PRIMARY KEY AUTOINCREMENT,
+                    memory_json TEXT NOT NULL,
+                    # Optionally add embedding here if not using separate FAISS index
+                    # embedding BLOB, 
+                    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
+                )
+            """)
+            # Stores the rule text directly
+            cursor.execute("""
+                CREATE TABLE IF NOT EXISTS rules (
+                    id INTEGER PRIMARY KEY AUTOINCREMENT,
+                    rule_text TEXT NOT NULL UNIQUE,
+                    # embedding BLOB,
+                    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
+                )
+            """)
+            conn.commit()
+        logger.info("SQLite tables for memories and rules checked/created.")
     except Exception as e:
-        logger.error(f"Error loading rules from {RULES_FILE}: {e}", exc_info=True)
-    return sorted(list(set(rules))) # Ensure unique and sorted
+        logger.error(f"SQLite table initialization error: {e}", exc_info=True)
 
-def save_rule_to_file(rule_text: str) -> bool:
-    """Saves a single rule (insight) to the JSON Lines file if it's new and valid."""
-    rule_text = rule_text.strip()
-    if not rule_text:
-        logger.warning("Attempted to save an empty rule.")
-        return False
+# --- Initialization ---
+def initialize_memory_system():
+    global _initialized, _embedder, _dimension, _faiss_memory_index, _memory_items_list, _faiss_rules_index, _rules_items_list
+    
+    with _init_lock:
+        if _initialized:
+            logger.info("Memory system already initialized.")
+            return
+
+        logger.info(f"Initializing memory system with backend: {STORAGE_BACKEND}")
+        init_start_time = time.time()
+
+        # 1. Load Sentence Transformer Model (always needed for semantic operations)
+        if not SentenceTransformer or not faiss or not np:
+            logger.error("Core RAG libraries (SentenceTransformers, FAISS, NumPy) not available. Cannot initialize semantic memory.")
+            _initialized = False # Mark as not properly initialized
+            return
+        
+        if not _embedder:
+            try:
+                logger.info("Loading SentenceTransformer model (all-MiniLM-L6-v2)...")
+                _embedder = SentenceTransformer('all-MiniLM-L6-v2', cache_folder="./sentence_transformer_cache")
+                _dimension = _embedder.get_sentence_embedding_dimension() or 384
+                logger.info(f"SentenceTransformer loaded. Dimension: {_dimension}")
+            except Exception as e:
+                logger.critical(f"FATAL: Error loading SentenceTransformer: {e}", exc_info=True)
+                _initialized = False
+                return # Cannot proceed without embedder
+
+        # 2. Initialize SQLite if used
+        if STORAGE_BACKEND == "SQLITE":
+            _init_sqlite_tables()
+
+        # 3. Load Memories
+        logger.info("Loading memories...")
+        temp_memories_json = []
+        if STORAGE_BACKEND == "RAM":
+            _memory_items_list = [] # Start fresh for RAM backend
+        elif STORAGE_BACKEND == "SQLITE" and sqlite3:
+            try:
+                with _get_sqlite_connection() as conn:
+                    temp_memories_json = [row[0] for row in conn.execute("SELECT memory_json FROM memories ORDER BY created_at ASC")]
+            except Exception as e: logger.error(f"Error loading memories from SQLite: {e}")
+        elif STORAGE_BACKEND == "HF_DATASET" and HF_TOKEN and Dataset and load_dataset:
+            try:
+                logger.info(f"Attempting to load memories from HF Dataset: {HF_MEMORY_DATASET_REPO}")
+                dataset = load_dataset(HF_MEMORY_DATASET_REPO, token=HF_TOKEN, trust_remote_code=True) # Add download_mode if needed
+                if "train" in dataset and "memory_json" in dataset["train"].column_names: # Assuming 'memory_json' column
+                    temp_memories_json = [m_json for m_json in dataset["train"]["memory_json"] if isinstance(m_json, str)]
+                else: logger.warning(f"HF Dataset {HF_MEMORY_DATASET_REPO} for memories not found or 'memory_json' column missing.")
+            except Exception as e: logger.error(f"Error loading memories from HF Dataset ({HF_MEMORY_DATASET_REPO}): {e}")
+        
+        _memory_items_list = temp_memories_json
+        logger.info(f"Loaded {len(_memory_items_list)} memory items from {STORAGE_BACKEND}.")
+
+        # 4. Build/Load FAISS Memory Index
+        _faiss_memory_index = faiss.IndexFlatL2(_dimension)
+        if _memory_items_list:
+            logger.info(f"Building FAISS index for {len(_memory_items_list)} memories...")
+            # Extract text to embed from memory JSON objects
+            texts_to_embed_mem = []
+            for mem_json_str in _memory_items_list:
+                try:
+                    mem_obj = json.loads(mem_json_str)
+                    # Consistent embedding strategy: user input + bot response + takeaway
+                    text = f"User: {mem_obj.get('user_input','')}\nAI: {mem_obj.get('bot_response','')}\nTakeaway: {mem_obj.get('metrics',{}).get('takeaway','N/A')}"
+                    texts_to_embed_mem.append(text)
+                except json.JSONDecodeError:
+                    logger.warning(f"Skipping malformed memory JSON for FAISS indexing: {mem_json_str[:100]}")
+            
+            if texts_to_embed_mem:
+                try:
+                    embeddings = _embedder.encode(texts_to_embed_mem, convert_to_tensor=False, show_progress_bar=False) # convert_to_numpy=True
+                    embeddings_np = np.array(embeddings, dtype=np.float32)
+                    if embeddings_np.ndim == 2 and embeddings_np.shape[0] == len(texts_to_embed_mem) and embeddings_np.shape[1] == _dimension:
+                        _faiss_memory_index.add(embeddings_np)
+                    else: logger.error(f"Memory embeddings shape error. Expected ({len(texts_to_embed_mem)}, {_dimension}), Got {embeddings_np.shape if hasattr(embeddings_np, 'shape') else 'N/A'}")
+                except Exception as e_faiss_mem: logger.error(f"Error building FAISS memory index: {e_faiss_mem}")
+        logger.info(f"FAISS memory index built. Total items: {_faiss_memory_index.ntotal if _faiss_memory_index else 'N/A'}")
 
-    # Validate format: [TYPE|SCORE] Text
-    if not re.match(r"\[(CORE_RULE|RESPONSE_PRINCIPLE|BEHAVIORAL_ADJUSTMENT|GENERAL_LEARNING)\|([\d\.]+?)\](.*)", rule_text, re.I|re.DOTALL):
-        logger.warning(f"Rule '{rule_text[:50]}...' has invalid format. Not saving.")
-        return False
 
-    current_rules = load_rules_from_file()
-    if rule_text in current_rules:
-        logger.info(f"Rule '{rule_text[:50]}...' already exists. Not saving duplicate.")
-        return False # Or True if "already exists" is considered success
+        # 5. Load Rules
+        logger.info("Loading rules...")
+        temp_rules_text = []
+        if STORAGE_BACKEND == "RAM":
+            _rules_items_list = []
+        elif STORAGE_BACKEND == "SQLITE" and sqlite3:
+            try:
+                with _get_sqlite_connection() as conn:
+                    temp_rules_text = [row[0] for row in conn.execute("SELECT rule_text FROM rules ORDER BY created_at ASC")]
+            except Exception as e: logger.error(f"Error loading rules from SQLite: {e}")
+        elif STORAGE_BACKEND == "HF_DATASET" and HF_TOKEN and Dataset and load_dataset:
+            try:
+                logger.info(f"Attempting to load rules from HF Dataset: {HF_RULES_DATASET_REPO}")
+                dataset = load_dataset(HF_RULES_DATASET_REPO, token=HF_TOKEN, trust_remote_code=True)
+                if "train" in dataset and "rule_text" in dataset["train"].column_names:
+                    temp_rules_text = [r_text for r_text in dataset["train"]["rule_text"] if isinstance(r_text, str) and r_text.strip()]
+                else: logger.warning(f"HF Dataset {HF_RULES_DATASET_REPO} for rules not found or 'rule_text' column missing.")
+            except Exception as e: logger.error(f"Error loading rules from HF Dataset ({HF_RULES_DATASET_REPO}): {e}")
+        
+        _rules_items_list = sorted(list(set(temp_rules_text))) # Ensure unique and sorted
+        logger.info(f"Loaded {len(_rules_items_list)} rule items from {STORAGE_BACKEND}.")
 
+        # 6. Build/Load FAISS Rules Index
+        _faiss_rules_index = faiss.IndexFlatL2(_dimension)
+        if _rules_items_list:
+            logger.info(f"Building FAISS index for {len(_rules_items_list)} rules...")
+            if _rules_items_list: # Check again in case it became empty after filtering
+                try:
+                    embeddings = _embedder.encode(_rules_items_list, convert_to_tensor=False, show_progress_bar=False)
+                    embeddings_np = np.array(embeddings, dtype=np.float32)
+                    if embeddings_np.ndim == 2 and embeddings_np.shape[0] == len(_rules_items_list) and embeddings_np.shape[1] == _dimension:
+                        _faiss_rules_index.add(embeddings_np)
+                    else: logger.error(f"Rule embeddings shape error. Expected ({len(_rules_items_list)}, {_dimension}), Got {embeddings_np.shape if hasattr(embeddings_np, 'shape') else 'N/A'}")
+                except Exception as e_faiss_rule: logger.error(f"Error building FAISS rule index: {e_faiss_rule}")
+        logger.info(f"FAISS rules index built. Total items: {_faiss_rules_index.ntotal if _faiss_rules_index else 'N/A'}")
+
+        _initialized = True
+        logger.info(f"Memory system initialization complete in {time.time() - init_start_time:.2f}s")
+
+
+# --- Memory Operations (Semantic) ---
+def add_memory_entry(user_input: str, metrics: dict, bot_response: str) -> tuple[bool, str]:
+    """Adds a memory entry to the configured backend and FAISS index."""
+    global _memory_items_list, _faiss_memory_index
+    if not _initialized: initialize_memory_system()
+    if not _embedder or not _faiss_memory_index:
+        return False, "Memory system or embedder not initialized for adding memory."
+
+    memory_obj = {
+        "user_input": user_input,
+        "metrics": metrics,
+        "bot_response": bot_response,
+        "timestamp": datetime.utcnow().isoformat()
+    }
+    memory_json_str = json.dumps(memory_obj)
+    
+    text_to_embed = f"User: {user_input}\nAI: {bot_response}\nTakeaway: {metrics.get('takeaway', 'N/A')}"
+    
     try:
-        with open(RULES_FILE, 'a', encoding='utf-8') as f:
-            # Store as JSON object for potential future metadata
-            json.dump({"text": rule_text, "added_at": datetime.utcnow().isoformat()}, f)
-            f.write('\n')
-        logger.info(f"Saved new rule: {rule_text[:70]}...")
-        return True
-    except Exception as e:
-        logger.error(f"Error saving rule '{rule_text[:50]}...' to {RULES_FILE}: {e}", exc_info=True)
-        return False
+        embedding = _embedder.encode([text_to_embed], convert_to_tensor=False)
+        embedding_np = np.array(embedding, dtype=np.float32).reshape(1, -1)
 
-def delete_rule_from_file(rule_text_to_delete: str) -> bool:
-    """Deletes a rule from the file."""
-    rule_text_to_delete = rule_text_to_delete.strip()
-    if not rule_text_to_delete: return False
+        if embedding_np.shape != (1, _dimension):
+            logger.error(f"Memory embedding shape error: {embedding_np.shape}. Expected (1, {_dimension})")
+            return False, "Embedding shape error."
 
-    current_rules = load_rules_from_file()
-    if rule_text_to_delete not in current_rules:
-        logger.info(f"Rule '{rule_text_to_delete[:50]}...' not found for deletion.")
-        return False
+        # Add to FAISS
+        _faiss_memory_index.add(embedding_np)
+        
+        # Add to in-memory list
+        _memory_items_list.append(memory_json_str)
+        
+        # Add to persistent storage
+        if STORAGE_BACKEND == "SQLITE" and sqlite3:
+            with _get_sqlite_connection() as conn:
+                conn.execute("INSERT INTO memories (memory_json) VALUES (?)", (memory_json_str,))
+                conn.commit()
+        elif STORAGE_BACKEND == "HF_DATASET" and HF_TOKEN and Dataset:
+            # This can be slow, consider batching or async push
+            logger.info(f"Pushing {len(_memory_items_list)} memories to HF Hub: {HF_MEMORY_DATASET_REPO}")
+            Dataset.from_dict({"memory_json": list(_memory_items_list)}).push_to_hub(HF_MEMORY_DATASET_REPO, token=HF_TOKEN, private=True) # Ensure 'private' as needed
+        
+        logger.info(f"Added memory. RAM: {len(_memory_items_list)}, FAISS: {_faiss_memory_index.ntotal}")
+        return True, "Memory added successfully."
+    except Exception as e:
+        logger.error(f"Error adding memory entry: {e}", exc_info=True)
+        # TODO: Potential rollback logic if FAISS add succeeded but backend failed (complex)
+        return False, f"Error adding memory: {e}"
+
+def retrieve_memories_semantic(query: str, k: int = 3) -> list[dict]:
+    """Retrieves k most relevant memories using semantic search."""
+    if not _initialized: initialize_memory_system()
+    if not _embedder or not _faiss_memory_index or _faiss_memory_index.ntotal == 0:
+        logger.debug("Cannot retrieve memories: Embedder, FAISS index not ready, or index is empty.")
+        return []
 
-    updated_rules = [rule for rule in current_rules if rule != rule_text_to_delete]
     try:
-        with open(RULES_FILE, 'w', encoding='utf-8') as f: # Overwrite with updated list
-            for rule_text in updated_rules:
-                json.dump({"text": rule_text, "added_at": "unknown"}, f) # timestamp lost on rewrite this way
-                f.write('\n')
-        logger.info(f"Deleted rule: {rule_text_to_delete[:70]}...")
-        return True
+        query_embedding = _embedder.encode([query], convert_to_tensor=False)
+        query_embedding_np = np.array(query_embedding, dtype=np.float32).reshape(1, -1)
+
+        if query_embedding_np.shape[1] != _dimension:
+            logger.error(f"Query embedding dimension mismatch. Expected {_dimension}, got {query_embedding_np.shape[1]}")
+            return []
+            
+        distances, indices = _faiss_memory_index.search(query_embedding_np, min(k, _faiss_memory_index.ntotal))
+        
+        results = []
+        for i in indices[0]:
+            if 0 <= i < len(_memory_items_list):
+                try:
+                    results.append(json.loads(_memory_items_list[i]))
+                except json.JSONDecodeError:
+                    logger.warning(f"Could not parse memory JSON from list at index {i}")
+            else:
+                logger.warning(f"FAISS index {i} out of bounds for memory_items_list (len: {len(_memory_items_list)})")
+
+        logger.debug(f"Retrieved {len(results)} memories semantically for query: '{query[:50]}...'")
+        return results
     except Exception as e:
-        logger.error(f"Error deleting rule '{rule_text_to_delete[:50]}...' from {RULES_FILE}: {e}", exc_info=True)
-        return False
+        logger.error(f"Error retrieving memories semantically: {e}", exc_info=True)
+        return []
 
 
-# --- Conversation Memories Management ---
+# --- Rule (Insight) Operations (Semantic) ---
+def add_rule_entry(rule_text: str) -> tuple[bool, str]:
+    """Adds a rule if valid and not a duplicate. Updates backend and FAISS."""
+    global _rules_items_list, _faiss_rules_index
+    if not _initialized: initialize_memory_system()
+    if not _embedder or not _faiss_rules_index:
+        return False, "Rule system or embedder not initialized."
+
+    rule_text = rule_text.strip()
+    if not rule_text: return False, "Rule text cannot be empty."
+    if not re.match(r"\[(CORE_RULE|RESPONSE_PRINCIPLE|BEHAVIORAL_ADJUSTMENT|GENERAL_LEARNING)\|([\d\.]+?)\](.*)", rule_text, re.I|re.DOTALL):
+        return False, "Invalid rule format."
+    if rule_text in _rules_items_list:
+        return False, "duplicate"
 
-def load_memories_from_file() -> list[dict]:
-    """Loads conversation memories from the JSON Lines file."""
-    memories = []
-    if not os.path.exists(MEMORIES_FILE):
-        return memories
     try:
-        with open(MEMORIES_FILE, 'r', encoding='utf-8') as f:
-            for line in f:
-                if line.strip():
-                    try:
-                        mem_obj = json.loads(line)
-                        # Basic validation for expected keys
-                        if all(k in mem_obj for k in ["user_input", "bot_response", "metrics", "timestamp"]):
-                            memories.append(mem_obj)
-                        else:
-                            logger.warning(f"Skipping memory object with missing keys: {line.strip()}")
-                    except json.JSONDecodeError:
-                        logger.warning(f"Skipping malformed JSON line in memories file: {line.strip()}")
-        logger.info(f"Loaded {len(memories)} memories from {MEMORIES_FILE}")
+        embedding = _embedder.encode([rule_text], convert_to_tensor=False)
+        embedding_np = np.array(embedding, dtype=np.float32).reshape(1, -1)
+
+        if embedding_np.shape != (1, _dimension):
+            return False, "Rule embedding shape error."
+
+        _faiss_rules_index.add(embedding_np)
+        _rules_items_list.append(rule_text)
+        _rules_items_list.sort()
+
+        if STORAGE_BACKEND == "SQLITE" and sqlite3:
+            with _get_sqlite_connection() as conn:
+                conn.execute("INSERT OR IGNORE INTO rules (rule_text) VALUES (?)", (rule_text,))
+                conn.commit()
+        elif STORAGE_BACKEND == "HF_DATASET" and HF_TOKEN and Dataset:
+            logger.info(f"Pushing {len(_rules_items_list)} rules to HF Hub: {HF_RULES_DATASET_REPO}")
+            Dataset.from_dict({"rule_text": list(_rules_items_list)}).push_to_hub(HF_RULES_DATASET_REPO, token=HF_TOKEN, private=True)
+        
+        logger.info(f"Added rule. RAM: {len(_rules_items_list)}, FAISS: {_faiss_rules_index.ntotal}")
+        return True, "Rule added successfully."
     except Exception as e:
-        logger.error(f"Error loading memories from {MEMORIES_FILE}: {e}", exc_info=True)
-    # Sort by timestamp if needed, though append-only usually keeps order
-    return sorted(memories, key=lambda x: x.get("timestamp", ""))
+        logger.error(f"Error adding rule entry: {e}", exc_info=True)
+        # Basic rollback if FAISS add succeeded
+        if rule_text in _rules_items_list and _faiss_rules_index.ntotal > 0: # Crude check
+             # A full rollback would involve rebuilding FAISS index from _rules_items_list before append.
+             # For simplicity, this is omitted here. State could be inconsistent on error.
+             pass
+        return False, f"Error adding rule: {e}"
 
+def retrieve_rules_semantic(query: str, k: int = 5) -> list[str]:
+    """Retrieves k most relevant rules using semantic search."""
+    if not _initialized: initialize_memory_system()
+    if not _embedder or not _faiss_rules_index or _faiss_rules_index.ntotal == 0:
+        return []
+    try:
+        query_embedding = _embedder.encode([query], convert_to_tensor=False)
+        query_embedding_np = np.array(query_embedding, dtype=np.float32).reshape(1, -1)
+        
+        if query_embedding_np.shape[1] != _dimension: return []
 
-def save_memory_to_file(user_input: str, bot_response: str, metrics: dict) -> bool:
-    """Saves a conversation memory to the JSON Lines file."""
-    if not user_input or not bot_response: # Metrics can be empty
-        logger.warning("Attempted to save memory with empty user input or bot response.")
-        return False
+        distances, indices = _faiss_rules_index.search(query_embedding_np, min(k, _faiss_rules_index.ntotal))
+        results = [_rules_items_list[i] for i in indices[0] if 0 <= i < len(_rules_items_list)]
+        logger.debug(f"Retrieved {len(results)} rules semantically for query: '{query[:50]}...'")
+        return results
+    except Exception as e:
+        logger.error(f"Error retrieving rules semantically: {e}", exc_info=True)
+        return []
 
-    memory_entry = {
-        "user_input": user_input,
-        "bot_response": bot_response,
-        "metrics": metrics,
-        "timestamp": datetime.utcnow().isoformat()
-    }
+def remove_rule_entry(rule_text_to_delete: str) -> bool:
+    """Removes a rule from backend and rebuilds FAISS for rules."""
+    global _rules_items_list, _faiss_rules_index
+    if not _initialized: initialize_memory_system()
+    if not _embedder or not _faiss_rules_index: return False
+
+    rule_text_to_delete = rule_text_to_delete.strip()
+    if rule_text_to_delete not in _rules_items_list:
+        return False # Not found
 
     try:
-        with open(MEMORIES_FILE, 'a', encoding='utf-8') as f:
-            json.dump(memory_entry, f)
-            f.write('\n')
-        logger.info(f"Saved new memory. User: {user_input[:50]}...")
+        _rules_items_list.remove(rule_text_to_delete)
+        _rules_items_list.sort() # Maintain sorted order
+
+        # Rebuild FAISS index for rules (simplest way to ensure consistency after removal)
+        new_faiss_rules_index = faiss.IndexFlatL2(_dimension)
+        if _rules_items_list:
+            embeddings = _embedder.encode(_rules_items_list, convert_to_tensor=False)
+            embeddings_np = np.array(embeddings, dtype=np.float32)
+            if embeddings_np.ndim == 2 and embeddings_np.shape[0] == len(_rules_items_list) and embeddings_np.shape[1] == _dimension:
+                 new_faiss_rules_index.add(embeddings_np)
+            else: # Should not happen if list is consistent
+                logger.error("Error rebuilding FAISS for rules after removal: Embedding shape error. State might be inconsistent.")
+                # Attempt to revert _rules_items_list (add back the rule)
+                _rules_items_list.append(rule_text_to_delete)
+                _rules_items_list.sort()
+                return False # Indicate failure
+        _faiss_rules_index = new_faiss_rules_index
+        
+        # Remove from persistent storage
+        if STORAGE_BACKEND == "SQLITE" and sqlite3:
+            with _get_sqlite_connection() as conn:
+                conn.execute("DELETE FROM rules WHERE rule_text = ?", (rule_text_to_delete,))
+                conn.commit()
+        elif STORAGE_BACKEND == "HF_DATASET" and HF_TOKEN and Dataset:
+            Dataset.from_dict({"rule_text": list(_rules_items_list)}).push_to_hub(HF_RULES_DATASET_REPO, token=HF_TOKEN, private=True)
+        
+        logger.info(f"Removed rule. RAM: {len(_rules_items_list)}, FAISS: {_faiss_rules_index.ntotal}")
         return True
     except Exception as e:
-        logger.error(f"Error saving memory to {MEMORIES_FILE}: {e}", exc_info=True)
+        logger.error(f"Error removing rule entry: {e}", exc_info=True)
+        # Potential partial failure, state might be inconsistent.
         return False
 
-def clear_all_rules() -> bool:
+# --- Utility functions to get all data (for UI display, etc.) ---
+def get_all_rules_cached() -> list[str]:
+    if not _initialized: initialize_memory_system()
+    return list(_rules_items_list)
+
+def get_all_memories_cached() -> list[dict]:
+    if not _initialized: initialize_memory_system()
+    # Convert JSON strings to dicts for easier use by UI
+    mem_dicts = []
+    for mem_json_str in _memory_items_list:
+        try: mem_dicts.append(json.loads(mem_json_str))
+        except: pass # Ignore parse errors for display
+    return mem_dicts
+
+def clear_all_memory_data_backend() -> bool:
+    """Clears all memories from backend and resets in-memory FAISS/list."""
+    global _memory_items_list, _faiss_memory_index
+    if not _initialized: initialize_memory_system()
+    
+    success = True
     try:
-        if os.path.exists(RULES_FILE):
-            os.remove(RULES_FILE)
-        logger.info("All rules cleared.")
-        return True
+        if STORAGE_BACKEND == "SQLITE" and sqlite3:
+            with _get_sqlite_connection() as conn: conn.execute("DELETE FROM memories"); conn.commit()
+        elif STORAGE_BACKEND == "HF_DATASET" and HF_TOKEN and Dataset:
+            # Deleting from HF usually means pushing an empty dataset
+            Dataset.from_dict({"memory_json": []}).push_to_hub(HF_MEMORY_DATASET_REPO, token=HF_TOKEN, private=True)
+        
+        _memory_items_list = []
+        if _faiss_memory_index: _faiss_memory_index.reset() # Clear FAISS index
+        logger.info("All memories cleared from backend and in-memory stores.")
     except Exception as e:
-        logger.error(f"Error clearing rules file {RULES_FILE}: {e}")
-        return False
+        logger.error(f"Error clearing all memory data: {e}")
+        success = False
+    return success
 
-def clear_all_memories() -> bool:
+def clear_all_rules_data_backend() -> bool:
+    """Clears all rules from backend and resets in-memory FAISS/list."""
+    global _rules_items_list, _faiss_rules_index
+    if not _initialized: initialize_memory_system()
+    
+    success = True
     try:
-        if os.path.exists(MEMORIES_FILE):
-            os.remove(MEMORIES_FILE)
-        logger.info("All memories cleared.")
-        return True
+        if STORAGE_BACKEND == "SQLITE" and sqlite3:
+            with _get_sqlite_connection() as conn: conn.execute("DELETE FROM rules"); conn.commit()
+        elif STORAGE_BACKEND == "HF_DATASET" and HF_TOKEN and Dataset:
+            Dataset.from_dict({"rule_text": []}).push_to_hub(HF_RULES_DATASET_REPO, token=HF_TOKEN, private=True)
+
+        _rules_items_list = []
+        if _faiss_rules_index: _faiss_rules_index.reset()
+        logger.info("All rules cleared from backend and in-memory stores.")
     except Exception as e:
-        logger.error(f"Error clearing memories file {MEMORIES_FILE}: {e}")
-        return False
+        logger.error(f"Error clearing all rules data: {e}")
+        success = False
+    return success
+
+# Optional: Function to save FAISS indices to disk (from ai-learn, if needed for persistence between app runs with RAM backend)
+FAISS_MEMORY_PATH = os.path.join(os.getenv("FAISS_STORAGE_PATH", "app_data/faiss_indices"), "memory_index.faiss")
+FAISS_RULES_PATH = os.path.join(os.getenv("FAISS_STORAGE_PATH", "app_data/faiss_indices"), "rules_index.faiss")
+
+def save_faiss_indices_to_disk():
+    if not _initialized or not faiss: return
+    
+    faiss_dir = os.path.dirname(FAISS_MEMORY_PATH)
+    if not os.path.exists(faiss_dir): os.makedirs(faiss_dir, exist_ok=True)
+
+    if _faiss_memory_index and _faiss_memory_index.ntotal > 0:
+        try:
+            faiss.write_index(_faiss_memory_index, FAISS_MEMORY_PATH)
+            logger.info(f"Memory FAISS index saved to disk ({_faiss_memory_index.ntotal} items).")
+        except Exception as e: logger.error(f"Error saving memory FAISS index: {e}")
+    
+    if _faiss_rules_index and _faiss_rules_index.ntotal > 0:
+        try:
+            faiss.write_index(_faiss_rules_index, FAISS_RULES_PATH)
+            logger.info(f"Rules FAISS index saved to disk ({_faiss_rules_index.ntotal} items).")
+        except Exception as e: logger.error(f"Error saving rules FAISS index: {e}")
+
+def load_faiss_indices_from_disk():
+    global _faiss_memory_index, _faiss_rules_index
+    if not _initialized or not faiss: return
+
+    if os.path.exists(FAISS_MEMORY_PATH) and _faiss_memory_index: # Check if index object exists
+        try:
+            logger.info(f"Loading memory FAISS index from {FAISS_MEMORY_PATH}...")
+            _faiss_memory_index = faiss.read_index(FAISS_MEMORY_PATH)
+            logger.info(f"Memory FAISS index loaded ({_faiss_memory_index.ntotal} items).")
+            # Consistency check: FAISS ntotal vs len(_memory_items_list)
+            if _faiss_memory_index.ntotal != len(_memory_items_list) and len(_memory_items_list) > 0:
+                 logger.warning(f"Memory FAISS index count ({_faiss_memory_index.ntotal}) differs from loaded texts ({len(_memory_items_list)}). Consider rebuilding FAISS.")
+        except Exception as e: logger.error(f"Error loading memory FAISS index: {e}. Will use fresh index.")
+    
+    if os.path.exists(FAISS_RULES_PATH) and _faiss_rules_index:
+        try:
+            logger.info(f"Loading rules FAISS index from {FAISS_RULES_PATH}...")
+            _faiss_rules_index = faiss.read_index(FAISS_RULES_PATH)
+            logger.info(f"Rules FAISS index loaded ({_faiss_rules_index.ntotal} items).")
+            if _faiss_rules_index.ntotal != len(_rules_items_list) and len(_rules_items_list) > 0:
+                 logger.warning(f"Rules FAISS index count ({_faiss_rules_index.ntotal}) differs from loaded texts ({len(_rules_items_list)}). Consider rebuilding FAISS.")
+        except Exception as e: logger.error(f"Error loading rules FAISS index: {e}. Will use fresh index.")
+
+