""" Temporal + Semantic + Entity Memory System for AI Agents. This implements a sophisticated memory architecture that combines: 1. Temporal links: Memories connected by time proximity 2. Semantic links: Memories connected by meaning/similarity 3. Entity links: Memories connected by shared entities (PERSON, ORG, etc.) 4. Spreading activation: Search through the graph with activation decay 5. Dynamic weighting: Recency and frequency-based importance """ import os from datetime import datetime, timedelta, timezone from typing import Any, Dict, List, Optional, Tuple import asyncpg from dotenv import load_dotenv import asyncio from .embeddings import Embeddings, SentenceTransformersEmbeddings import time import numpy as np import uuid import logging from .utils import ( extract_facts, calculate_recency_weight, calculate_frequency_weight, ) from .entity_resolver import EntityResolver from .operations import EmbeddingOperationsMixin, LinkOperationsMixin, ThinkOperationsMixin def utcnow(): """Get current UTC time with timezone info.""" return datetime.now(timezone.utc) # Logger for memory system logger = logging.getLogger(__name__) class TemporalSemanticMemory( EmbeddingOperationsMixin, LinkOperationsMixin, ThinkOperationsMixin, ): """ Advanced memory system using temporal and semantic linking with PostgreSQL. Uses mixin architecture for code organization: - EmbeddingOperationsMixin: Embedding generation - LinkOperationsMixin: Entity, temporal, and semantic link creation - ThinkOperationsMixin: Think operations for formulating answers with opinions """ def __init__( self, db_url: Optional[str] = None, embeddings: Optional[Embeddings] = None, embedding_model: Optional[str] = None, pool_min_size: int = 5, pool_max_size: int = 100, ): """ Initialize the temporal + semantic memory system. Args: db_url: PostgreSQL connection URL (postgresql://user:pass@host:port/dbname) embeddings: Embeddings implementation to use. If not provided, uses SentenceTransformersEmbeddings embedding_model: (Deprecated) Name of the SentenceTransformer model to use. Use embeddings parameter instead. pool_min_size: Minimum number of connections in the pool (default: 5) pool_max_size: Maximum number of connections in the pool (default: 100) Increase for parallel think/search operations (e.g., 200-300 for 100+ parallel thinks) """ load_dotenv() # Initialize PostgreSQL connection URL self.db_url = db_url or os.getenv("DATABASE_URL") if not self.db_url: raise ValueError( "Database URL not found. " "Set DATABASE_URL environment variable." ) # Connection pool (will be created in initialize()) self._pool = None self._initialized = False self._pool_min_size = pool_min_size self._pool_max_size = pool_max_size # Initialize entity resolver (will be created in initialize()) self.entity_resolver = None # Initialize embeddings if embeddings is not None: self.embeddings = embeddings else: # Default to SentenceTransformersEmbeddings model_name = embedding_model or "BAAI/bge-small-en-v1.5" self.embeddings = SentenceTransformersEmbeddings(model_name) # Initialize LLM client (cached for reuse across operations) from openai import AsyncOpenAI groq_api_key = os.getenv("GROQ_API_KEY") if groq_api_key: self._llm_client = AsyncOpenAI( api_key=groq_api_key, base_url="https://api.groq.com/openai/v1" ) else: self._llm_client = None # Will be created on-demand if needed # Background queue for access count updates (to avoid blocking searches) self._access_count_queue = asyncio.Queue() self._access_count_worker_task = None self._shutdown_event = asyncio.Event() # Track background opinion PUT tasks to ensure clean shutdown self._background_tasks = set() # Backpressure mechanism: limit concurrent searches to prevent overwhelming the database self._search_semaphore = asyncio.Semaphore(32) async def _access_count_worker(self): """Background worker that processes access count updates in batches.""" pool = self._pool # Pool is guaranteed to exist when worker starts while not self._shutdown_event.is_set(): try: # Collect updates for up to 1 second or 1000 items updates = {} deadline = asyncio.get_event_loop().time() + 1.0 while len(updates) < 1000 and asyncio.get_event_loop().time() < deadline: try: # Wait for items with short timeout remaining_time = max(0.1, deadline - asyncio.get_event_loop().time()) node_ids = await asyncio.wait_for( self._access_count_queue.get(), timeout=remaining_time ) # Deduplicate by adding to set for node_id in node_ids: updates[node_id] = True except asyncio.TimeoutError: break # Process batch if we have updates if updates: node_id_list = list(updates.keys()) try: # Convert string UUIDs to UUID type for faster matching uuid_list = [uuid.UUID(nid) for nid in node_id_list] async with pool.acquire() as conn: await conn.execute( "UPDATE memory_units SET access_count = access_count + 1 WHERE id = ANY($1::uuid[])", uuid_list ) except Exception as e: logger.error(f"Access count worker: Error updating access counts: {e}") except asyncio.CancelledError: break except Exception as e: logger.error(f"Access count worker: Unexpected error: {e}") await asyncio.sleep(1) # Backoff on error async def initialize(self): """Initialize the connection pool and background workers.""" if self._initialized: return # Create connection pool # For read-heavy workloads with many parallel think/search operations, # we need a larger pool. Read operations don't need strong isolation. self._pool = await asyncpg.create_pool( self.db_url, min_size=self._pool_min_size, max_size=self._pool_max_size, command_timeout=60, statement_cache_size=0 # Disable prepared statement cache ) # Initialize entity resolver with pool self.entity_resolver = EntityResolver(self._pool) # Start access count worker self._access_count_worker_task = asyncio.create_task(self._access_count_worker()) self._initialized = True logger.info("Memory system initialized (pool and workers started)") async def _get_pool(self) -> asyncpg.Pool: """Get the connection pool (must call initialize() first).""" if not self._initialized: await self.initialize() return self._pool async def wait_for_background_tasks(self): """Wait for all background tasks (e.g., opinion PUTs) to complete.""" if self._background_tasks: await asyncio.gather(*self._background_tasks, return_exceptions=True) async def close(self): """Close the connection pool and shutdown background workers.""" logger.info("close() started") # Signal shutdown to worker self._shutdown_event.set() logger.info("shutdown event set") # Wait for background opinion PUT tasks to complete if self._background_tasks: logger.debug(f"waiting for {len(self._background_tasks)} background tasks to complete") await self.wait_for_background_tasks() logger.debug("background tasks completed") # Cancel and wait for worker task if self._access_count_worker_task is not None: logger.debug("cancelling worker task") self._access_count_worker_task.cancel() try: logger.debug("waiting for worker task to finish") await self._access_count_worker_task logger.debug("worker task finished") except asyncio.CancelledError: logger.debug("worker task cancelled successfully") else: logger.debug("no worker task to cancel") # Close pool if self._pool is not None: logger.debug("closing connection pool") self._pool.terminate() logger.debug("connection pool closed") self._pool = None else: logger.debug("no pool to close") logger.debug("close() completed") async def _find_duplicate_facts_batch( self, conn, agent_id: str, texts: List[str], embeddings: List[List[float]], event_date: datetime, time_window_hours: int = 24, similarity_threshold: float = 0.95 ) -> List[bool]: """ Check which facts are duplicates using semantic similarity + temporal window. For each new fact, checks if a semantically similar fact already exists within the time window. Uses pgvector cosine similarity for efficiency. Args: conn: Database connection agent_id: Agent identifier texts: List of fact texts to check embeddings: Corresponding embeddings event_date: Event date for temporal filtering time_window_hours: Hours before/after event_date to search (default: 24) similarity_threshold: Minimum cosine similarity to consider duplicate (default: 0.95) Returns: List of booleans - True if fact is a duplicate (should skip), False if new """ if not texts: return [] time_lower = event_date - timedelta(hours=time_window_hours) time_upper = event_date + timedelta(hours=time_window_hours) # Fetch ALL existing facts in time window ONCE (much faster than N queries) import time as time_mod fetch_start = time_mod.time() existing_facts = await conn.fetch( """ SELECT id, text, embedding FROM memory_units WHERE agent_id = $1 AND event_date BETWEEN $2 AND $3 """, agent_id, time_lower, time_upper ) logger.debug(f" [3.X] Fetched {len(existing_facts)} existing facts in {time_mod.time() - fetch_start:.3f}s") # If no existing facts, nothing is duplicate if not existing_facts: return [False] * len(texts) # Compute similarities in Python (vectorized with numpy) import numpy as np is_duplicate = [] # Convert existing embeddings to numpy for faster computation embedding_arrays = [] for row in existing_facts: raw_emb = row['embedding'] # Handle different pgvector formats if isinstance(raw_emb, str): # Parse string format: "[1.0, 2.0, ...]" import json emb = np.array(json.loads(raw_emb), dtype=np.float32) elif isinstance(raw_emb, (list, tuple)): emb = np.array(raw_emb, dtype=np.float32) else: # Try direct conversion emb = np.array(raw_emb, dtype=np.float32) embedding_arrays.append(emb) if not embedding_arrays: existing_embeddings = np.array([]) elif len(embedding_arrays) == 1: # Single embedding: reshape to (1, dim) existing_embeddings = embedding_arrays[0].reshape(1, -1) else: # Multiple embeddings: vstack existing_embeddings = np.vstack(embedding_arrays) comp_start = time_mod.time() for embedding in embeddings: # Compute cosine similarity with all existing facts emb_array = np.array(embedding) # Cosine similarity = 1 - cosine distance # For normalized vectors: cosine_sim = dot product similarities = np.dot(existing_embeddings, emb_array) # Check if any existing fact is too similar max_similarity = np.max(similarities) if len(similarities) > 0 else 0 is_duplicate.append(max_similarity > similarity_threshold) logger.debug(f" [3.X] Computed {len(texts)} x {len(existing_facts)} similarities in {time_mod.time() - comp_start:.3f}s") return is_duplicate def put( self, agent_id: str, content: str, context: str = "", event_date: Optional[datetime] = None, ) -> List[str]: """ Store content as memory units (synchronous wrapper). This is a synchronous wrapper around put_async() for convenience. For best performance, use put_async() directly. Args: agent_id: Unique identifier for the agent content: Text content to store context: Context about when/why this memory was formed event_date: When the event occurred (defaults to now) Returns: List of created unit IDs """ # Run async version synchronously return asyncio.run(self.put_async(agent_id, content, context, event_date)) async def put_async( self, agent_id: str, content: str, context: str = "", event_date: Optional[datetime] = None, document_id: Optional[str] = None, document_metadata: Optional[Dict[str, Any]] = None, upsert: bool = False, fact_type_override: Optional[str] = None, confidence_score: Optional[float] = None, ) -> List[str]: """ Store content as memory units with temporal and semantic links (ASYNC version). This is a convenience wrapper around put_batch_async for a single content item. Args: agent_id: Unique identifier for the agent content: Text content to store context: Context about when/why this memory was formed event_date: When the event occurred (defaults to now) document_id: Optional document ID for tracking and upsert document_metadata: Optional metadata about the document upsert: If True and document_id exists, delete old units and create new ones fact_type_override: Override fact type ('world', 'agent', 'opinion') confidence_score: Confidence score for opinions (0.0 to 1.0) Returns: List of created unit IDs """ # Use put_batch_async with a single item (avoids code duplication) result = await self.put_batch_async( agent_id=agent_id, contents=[{ "content": content, "context": context, "event_date": event_date }], document_id=document_id, document_metadata=document_metadata, upsert=upsert, fact_type_override=fact_type_override, confidence_score=confidence_score ) # Return the first (and only) list of unit IDs return result[0] if result else [] async def put_batch_async( self, agent_id: str, contents: List[Dict[str, Any]], document_id: Optional[str] = None, document_metadata: Optional[Dict[str, Any]] = None, upsert: bool = False, fact_type_override: Optional[str] = None, confidence_score: Optional[float] = None, ) -> List[List[str]]: """ Store multiple content items as memory units in ONE batch operation. This is MUCH more efficient than calling put_async multiple times: - Extracts facts from all contents in parallel - Generates ALL embeddings in ONE batch - Does ALL database operations in ONE transaction Args: agent_id: Unique identifier for the agent contents: List of dicts with keys: - "content" (required): Text content to store - "context" (optional): Context about the memory - "event_date" (optional): When the event occurred document_id: Optional document ID for tracking and upsert document_metadata: Optional metadata about the document upsert: If True and document_id exists, delete old units and create new ones fact_type_override: Override fact type for all facts ('world', 'agent', 'opinion') confidence_score: Confidence score for opinions (0.0 to 1.0) Returns: List of lists of unit IDs (one list per content item) Example: unit_ids = await memory.put_batch_async( agent_id="user123", contents=[ {"content": "Alice works at Google", "context": "conversation"}, {"content": "Bob loves Python", "context": "conversation"}, ], document_id="meeting-2024-01-15", upsert=True ) # Returns: [["unit-id-1"], ["unit-id-2"]] """ start_time = time.time() logger.info(f"\n{'='*60}") logger.info(f"PUT_BATCH_ASYNC START: {agent_id}") logger.info(f"Batch size: {len(contents)} content items") logger.info(f"{'='*60}") if not contents: return [] # Step 1: Extract facts from ALL contents in parallel step_start = time.time() # Create tasks for parallel fact extraction fact_extraction_tasks = [] for item in contents: content = item["content"] context = item.get("context", "") event_date = item.get("event_date") or utcnow() task = extract_facts(content, event_date, context) fact_extraction_tasks.append((task, event_date, context)) # Wait for all fact extractions to complete all_fact_results = await asyncio.gather(*[task for task, _, _ in fact_extraction_tasks]) logger.info(f"[1] Extract facts (parallel): {len(fact_extraction_tasks)} contents in {time.time() - step_start:.3f}s") # Flatten and track which facts belong to which content all_fact_texts = [] all_fact_dates = [] all_contexts = [] all_fact_entities = [] # NEW: Store LLM-extracted entities per fact all_fact_types = [] # Store fact type (world or agent) content_boundaries = [] # [(start_idx, end_idx), ...] current_idx = 0 for i, ((_, event_date, context), fact_dicts) in enumerate(zip(fact_extraction_tasks, all_fact_results)): start_idx = current_idx for fact_dict in fact_dicts: all_fact_texts.append(fact_dict['fact']) try: from dateutil import parser as date_parser fact_date = date_parser.isoparse(fact_dict['date']) all_fact_dates.append(fact_date) except Exception: all_fact_dates.append(event_date) all_contexts.append(context) # Extract entities from fact (default to empty list if not present) all_fact_entities.append(fact_dict.get('entities', [])) # Extract fact type (use override if provided, else use extracted type or default to 'world') if fact_type_override: all_fact_types.append(fact_type_override) else: all_fact_types.append(fact_dict.get('fact_type', 'world')) end_idx = current_idx + len(fact_dicts) content_boundaries.append((start_idx, end_idx)) current_idx = end_idx total_facts = len(all_fact_texts) if total_facts == 0: return [[] for _ in contents] # Step 2: Generate ALL embeddings in ONE batch (HUGE speedup!) step_start = time.time() all_embeddings = await self._generate_embeddings_batch(all_fact_texts) logger.info(f"[2] Generate embeddings (parallel): {len(all_embeddings)} embeddings in {time.time() - step_start:.3f}s") # Step 3: Process everything in ONE database transaction logger.debug("Getting connection pool") pool = await self._get_pool() logger.debug("Acquiring connection from pool") async with pool.acquire() as conn: logger.debug("Starting transaction") async with conn.transaction(): logger.debug("Inside transaction") try: # Handle document tracking and upsert if document_id: logger.debug(f"Handling document tracking for {document_id}") import hashlib import json # Calculate content hash from all content items combined_content = "\n".join([c.get("content", "") for c in contents]) content_hash = hashlib.sha256(combined_content.encode()).hexdigest() # If upsert, delete old document first (cascades to units and links) if upsert: deleted = await conn.fetchval( "DELETE FROM documents WHERE id = $1 AND agent_id = $2 RETURNING id", document_id, agent_id ) if deleted: logger.debug(f"[3.1] Upsert: Deleted existing document '{document_id}' and all its units") # Insert or update document # Always use ON CONFLICT for idempotent behavior await conn.execute( """ INSERT INTO documents (id, agent_id, original_text, content_hash, metadata) VALUES ($1, $2, $3, $4, $5) ON CONFLICT (id, agent_id) DO UPDATE SET original_text = EXCLUDED.original_text, content_hash = EXCLUDED.content_hash, metadata = EXCLUDED.metadata, updated_at = NOW() """, document_id, agent_id, combined_content, content_hash, json.dumps(document_metadata or {}) ) logger.debug(f"[3.2] Document '{document_id}' stored/updated") # Deduplication check for all facts (batched by time window) logger.debug("Starting deduplication check") step_start = time.time() # Group facts by event_date (rounded to 12-hour buckets) for batching from collections import defaultdict time_buckets = defaultdict(list) for idx, (sentence, embedding, fact_date) in enumerate(zip(all_fact_texts, all_embeddings, all_fact_dates)): # Round to 12-hour bucket to group similar times bucket_key = fact_date.replace(hour=(fact_date.hour // 12) * 12, minute=0, second=0, microsecond=0) time_buckets[bucket_key].append((idx, sentence, embedding, fact_date)) # Process each bucket in batch all_is_duplicate = [False] * total_facts # Initialize all as not duplicate for bucket_date, bucket_items in time_buckets.items(): indices = [item[0] for item in bucket_items] sentences = [item[1] for item in bucket_items] embeddings = [item[2] for item in bucket_items] # Use bucket_date as representative for time window dup_flags = await self._find_duplicate_facts_batch( conn, agent_id, sentences, embeddings, bucket_date, time_window_hours=24 ) # Map results back to original indices for idx, is_dup in zip(indices, dup_flags): all_is_duplicate[idx] = is_dup duplicates_filtered = sum(all_is_duplicate) new_facts = total_facts - duplicates_filtered logger.debug(f"Deduplication complete: {duplicates_filtered} duplicates filtered, {new_facts} new facts ({len(time_buckets)} time buckets)") logger.info(f"[3] Deduplication check: {duplicates_filtered} duplicates filtered, {new_facts} new facts in {time.time() - step_start:.3f}s") # Filter out duplicates filtered_sentences = [s for s, is_dup in zip(all_fact_texts, all_is_duplicate) if not is_dup] filtered_embeddings = [e for e, is_dup in zip(all_embeddings, all_is_duplicate) if not is_dup] filtered_dates = [d for d, is_dup in zip(all_fact_dates, all_is_duplicate) if not is_dup] filtered_contexts = [c for c, is_dup in zip(all_contexts, all_is_duplicate) if not is_dup] filtered_entities = [ents for ents, is_dup in zip(all_fact_entities, all_is_duplicate) if not is_dup] filtered_fact_types = [ft for ft, is_dup in zip(all_fact_types, all_is_duplicate) if not is_dup] if not filtered_sentences: logger.debug(f"[PUT_BATCH_ASYNC] All facts were duplicates, returning empty") return [[] for _ in contents] # Batch insert ALL units step_start = time.time() # Convert embeddings to strings for asyncpg vector type filtered_embeddings_str = [str(emb) for emb in filtered_embeddings] # Prepare confidence scores (only for opinions) # If fact_type is 'opinion' and no confidence_score provided, use default of 1.0 confidence_scores = [ confidence_score if confidence_score is not None else 1.0 if ft == 'opinion' else None for ft in filtered_fact_types ] results = await conn.fetch( """ INSERT INTO memory_units (agent_id, document_id, text, context, embedding, event_date, fact_type, confidence_score, access_count) SELECT * FROM unnest($1::text[], $2::text[], $3::text[], $4::text[], $5::vector[], $6::timestamptz[], $7::text[], $8::float[], $9::integer[]) RETURNING id """, [agent_id] * len(filtered_sentences), [document_id] * len(filtered_sentences) if document_id else [None] * len(filtered_sentences), filtered_sentences, filtered_contexts, filtered_embeddings_str, filtered_dates, filtered_fact_types, confidence_scores, [0] * len(filtered_sentences) ) created_unit_ids = [str(row['id']) for row in results] logger.debug(f"Batch insert complete: {len(created_unit_ids)} units created") logger.info(f"[5] Batch insert units: {len(created_unit_ids)} units in {time.time() - step_start:.3f}s") # Process entities for ALL units logger.debug("Processing entities") step_start = time.time() all_entity_links = await self._extract_entities_batch_optimized( conn, agent_id, created_unit_ids, filtered_sentences, "", filtered_dates, filtered_entities ) logger.debug(f"Entity processing complete: {len(all_entity_links)} links") logger.info(f"[6] Process entities (batched): {time.time() - step_start:.3f}s") # Create temporal links logger.debug("Creating temporal links") step_start = time.time() await self._create_temporal_links_batch_per_fact(conn, agent_id, created_unit_ids) logger.debug("Temporal links complete") logger.info(f"[7] Batch create temporal links: {time.time() - step_start:.3f}s") # Create semantic links logger.debug("Creating semantic links") step_start = time.time() await self._create_semantic_links_batch(conn, agent_id, created_unit_ids, filtered_embeddings) logger.debug("Semantic links complete") logger.info(f"[8] Batch create semantic links: {time.time() - step_start:.3f}s") # Insert entity links logger.debug("Inserting entity links") step_start = time.time() if all_entity_links: await self._insert_entity_links_batch(conn, all_entity_links) logger.debug("Entity links inserted") logger.info(f"[9] Batch insert entity links: {time.time() - step_start:.3f}s") # Transaction auto-commits on success commit_start = time.time() logger.debug(f"[10] Commit: {time.time() - commit_start:.3f}s") # Map created unit IDs back to original content items # Account for duplicates when mapping back result_unit_ids = [] filtered_idx = 0 for start_idx, end_idx in content_boundaries: content_unit_ids = [] for i in range(start_idx, end_idx): if not all_is_duplicate[i]: content_unit_ids.append(created_unit_ids[filtered_idx]) filtered_idx += 1 result_unit_ids.append(content_unit_ids) total_time = time.time() - start_time logger.info(f"\n{'='*60}") logger.info(f"PUT_BATCH_ASYNC COMPLETE: {len(created_unit_ids)} units from {len(contents)} contents in {total_time:.3f}s") logger.info(f"{'='*60}\n") # Trigger opinion reinforcement in background (non-blocking) # Only trigger if there are entities in the new units if any(filtered_entities): asyncio.create_task( self._reinforce_opinions_async( agent_id=agent_id, created_unit_ids=created_unit_ids, unit_texts=filtered_sentences, unit_entities=filtered_entities ) ) logger.debug("[PUT_BATCH_ASYNC] Opinion reinforcement task queued in background") return result_unit_ids except Exception as e: # Transaction auto-rolls back on exception import traceback traceback.print_exc() raise Exception(f"Failed to store batch memory: {str(e)}") def search( self, agent_id: str, query: str, thinking_budget: int = 50, top_k: int = 10, enable_trace: bool = False, weight_activation: float = 0.30, weight_semantic: float = 0.30, weight_recency: float = 0.25, weight_frequency: float = 0.15, mmr_lambda: float = 0.5, fact_type: Optional[str] = None, ) -> tuple[List[Dict[str, Any]], Optional[Any]]: """ Search memories using spreading activation (synchronous wrapper). This is a synchronous wrapper around search_async() for convenience. For best performance, use search_async() directly. Args: agent_id: Agent ID to search for query: Search query thinking_budget: How many units to explore (computational budget) top_k: Number of results to return enable_trace: If True, returns detailed SearchTrace object weight_activation: Weight for activation component (default: 0.30) weight_semantic: Weight for semantic similarity component (default: 0.30) weight_recency: Weight for recency component (default: 0.25) weight_frequency: Weight for frequency component (default: 0.15) mmr_lambda: Lambda for MMR diversification (0=max diversity, 1=no diversity, default: 0.5) fact_type: Optional filter for fact type ('world' or 'agent') Returns: Tuple of (results, trace) """ # Run async version synchronously return asyncio.run(self.search_async( agent_id, query, thinking_budget, top_k, enable_trace, weight_activation, weight_semantic, weight_recency, weight_frequency, mmr_lambda, fact_type )) async def search_async( self, agent_id: str, query: str, thinking_budget: int = 50, top_k: int = 10, enable_trace: bool = False, weight_activation: float = 0.30, weight_semantic: float = 0.30, weight_recency: float = 0.25, weight_frequency: float = 0.15, mmr_lambda: float = 0.5, fact_type: Optional[str] = None, max_neighbors_per_node: int = 20, ) -> tuple[List[Dict[str, Any]], Optional[Any]]: """ Search memories using spreading activation (ASYNC version). This implements the core SEARCH operation: 1. Find entry points (most relevant units via vector search) 2. Spread activation through the graph 3. Weight results by activation + recency + frequency 4. Return top results Args: agent_id: Agent ID to search for query: Search query thinking_budget: How many units to explore (computational budget) top_k: Number of results to return live_tracer: Optional LiveSearchTracer for visualization Returns: List of memory units with their weights, sorted by relevance """ # Backpressure: limit concurrent searches to prevent overwhelming the database async with self._search_semaphore: # Initialize tracer if requested from .search_tracer import SearchTracer tracer = SearchTracer(query, thinking_budget, top_k) if enable_trace else None if tracer: tracer.start() pool = await self._get_pool() search_start = time.time() # Buffer logs for clean output in concurrent scenarios search_id = f"{agent_id[:8]}-{int(time.time() * 1000) % 100000}" log_buffer = [] log_buffer.append(f"[SEARCH {search_id}] Query: '{query[:50]}...' (budget={thinking_budget}, top_k={top_k})") try: # Step 1: Generate query embedding (CPU-bound, no DB needed) step_start = time.time() query_embedding = self._generate_embedding(query) step_duration = time.time() - step_start log_buffer.append(f" [1] Generate query embedding: {step_duration:.3f}s") if tracer: tracer.record_query_embedding(query_embedding) tracer.add_phase_metric("generate_query_embedding", step_duration) # Step 2: Find entry points (acquire connection only for this query) step_start = time.time() query_embedding_str = str(query_embedding) # Log connection acquisition conn_acquire_start = time.time() async with pool.acquire() as conn: conn_acquire_time = time.time() - conn_acquire_start if conn_acquire_time > 0.1: # Log if waiting > 100ms log_buffer.append(f" [2.1] Waited {conn_acquire_time:.3f}s for connection (pool busy)") # Build entry point query with optional fact_type filter if fact_type: entry_points = await conn.fetch( """ SELECT id, text, context, event_date, access_count, embedding, 1 - (embedding <=> $1::vector) AS similarity FROM memory_units WHERE agent_id = $2 AND embedding IS NOT NULL AND fact_type = $3 AND (1 - (embedding <=> $1::vector)) >= 0.5 ORDER BY embedding <=> $1::vector LIMIT 3 """, query_embedding_str, agent_id, fact_type ) else: entry_points = await conn.fetch( """ SELECT id, text, context, event_date, access_count, embedding, 1 - (embedding <=> $1::vector) AS similarity FROM memory_units WHERE agent_id = $2 AND embedding IS NOT NULL AND (1 - (embedding <=> $1::vector)) >= 0.5 ORDER BY embedding <=> $1::vector LIMIT 3 """, query_embedding_str, agent_id ) step_duration = time.time() - step_start log_buffer.append(f" [2] Find entry points: {len(entry_points)} found in {step_duration:.3f}s") if tracer: tracer.add_phase_metric("find_entry_points", step_duration, {"count": len(entry_points)}) for rank, ep in enumerate(entry_points, 1): tracer.add_entry_point( node_id=str(ep["id"]), text=ep["text"], similarity=ep["similarity"], rank=rank ) if not entry_points: logger.debug(f"[SEARCH] Complete: 0 results in {time.time() - search_start:.3f}s") if tracer: trace = tracer.finalize([]) return [], trace return [], None # Step 3: Spreading activation with budget (in-memory processing) step_start = time.time() visited = set() results = [] budget_remaining = thinking_budget # Initialize entry points with their actual similarity scores instead of 1.0 # Format: (unit, activation, is_entry, parent_node_id, link_type, link_weight) queue = [(dict(unit), unit["similarity"], True, None, None, None) for unit in entry_points] # Track substep timings calculate_weight_time = 0 query_neighbors_time = 0 process_neighbors_time = 0 # Track which nodes were visited for deferred access count update visited_node_ids = [] # Process nodes in batches for efficient neighbor querying # OPTIMIZATION: Increased from 50 to 100 since we now limit neighbors per node # This reduces round trips while keeping result set manageable BATCH_SIZE = 100 nodes_to_process = [] # (unit, activation, is_entry_point, parent_node_id, link_type, link_weight) while queue and budget_remaining > 0: # Collect a batch of nodes to process (in-memory, no DB) while queue and len(nodes_to_process) < BATCH_SIZE and budget_remaining > 0: current_unit, activation, is_entry_point, parent_node_id, link_type, link_weight = queue.pop(0) unit_id = str(current_unit["id"]) if unit_id not in visited: visited.add(unit_id) budget_remaining -= 1 nodes_to_process.append((current_unit, activation, is_entry_point, parent_node_id, link_type, link_weight)) visited_node_ids.append(unit_id) # Track for deferred update elif tracer: # Node already visited - prune tracer.prune_node(unit_id, "already_visited", activation) if not nodes_to_process: break # Acquire connection ONLY for neighbor queries (defer access count updates) node_ids = [str(node[0]["id"]) for node in nodes_to_process] # Log connection acquisition for batch queries batch_conn_start = time.time() async with pool.acquire() as conn: batch_conn_acquire = time.time() - batch_conn_start if batch_conn_acquire > 0.1: # Log if waiting > 100ms log_buffer.append(f" [3.3.1] Waited {batch_conn_acquire:.3f}s for connection (pool busy) - batch size: {len(node_ids)}") # Query neighbors for ALL nodes in batch at once (without embeddings for speed) # Convert string UUIDs to UUID type for faster matching substep_start = time.time() uuid_array = [uuid.UUID(nid) for nid in node_ids] # Build neighbor query with optional fact_type filter # OPTIMIZATION: Limit neighbors per node to reduce data transfer # Dense graphs can have 100+ neighbors per node, but spreading activation # only needs top-weighted neighbors. This reduces query from 9000→1000 rows. # Configurable via max_neighbors_per_node parameter (default: 20) if fact_type: all_neighbors = await conn.fetch( """ SELECT * FROM ( SELECT ml.from_unit_id, ml.to_unit_id, ml.weight, ml.link_type, ml.entity_id, mu.text, mu.context, mu.event_date, mu.access_count, mu.id as neighbor_id, ROW_NUMBER() OVER (PARTITION BY ml.from_unit_id ORDER BY ml.weight DESC) as rn FROM memory_links ml JOIN memory_units mu ON ml.to_unit_id = mu.id WHERE ml.from_unit_id = ANY($1::uuid[]) AND ml.weight >= 0.1 AND mu.fact_type = $2 ) sub WHERE rn <= $3 ORDER BY from_unit_id, weight DESC """, uuid_array, fact_type, max_neighbors_per_node ) else: all_neighbors = await conn.fetch( """ SELECT * FROM ( SELECT ml.from_unit_id, ml.to_unit_id, ml.weight, ml.link_type, ml.entity_id, mu.text, mu.context, mu.event_date, mu.access_count, mu.id as neighbor_id, ROW_NUMBER() OVER (PARTITION BY ml.from_unit_id ORDER BY ml.weight DESC) as rn FROM memory_links ml JOIN memory_units mu ON ml.to_unit_id = mu.id WHERE ml.from_unit_id = ANY($1::uuid[]) AND ml.weight >= 0.1 ) sub WHERE rn <= $2 ORDER BY from_unit_id, weight DESC """, uuid_array, max_neighbors_per_node ) neighbor_query_time = time.time() - substep_start if neighbor_query_time > 1.0: # Log slow neighbor queries log_buffer.append(f" [3.3.3] Slow NEIGHBOR query: {neighbor_query_time:.3f}s for {len(node_ids)} nodes → {len(all_neighbors)} neighbors") query_neighbors_time += neighbor_query_time # Fetch embeddings for current batch nodes (needed for weight calculation) substep_start = time.time() embeddings = await conn.fetch( "SELECT id, embedding FROM memory_units WHERE id = ANY($1::uuid[])", uuid_array ) embedding_map = {str(row["id"]): row["embedding"] for row in embeddings} fetch_embeddings_time = time.time() - substep_start if fetch_embeddings_time > 0.5: log_buffer.append(f" [3.3.4] Slow EMBEDDING fetch: {fetch_embeddings_time:.3f}s for {len(node_ids)} nodes") query_neighbors_time += fetch_embeddings_time # Group neighbors by from_unit_id (in-memory, no DB) substep_start = time.time() neighbors_by_node = {} for neighbor in all_neighbors: from_id = str(neighbor["from_unit_id"]) if from_id not in neighbors_by_node: neighbors_by_node[from_id] = [] neighbors_by_node[from_id].append(neighbor) # Process each node in the batch (CPU-bound, no DB) for current_unit, activation, is_entry_point, parent_node_id, parent_link_type, parent_link_weight in nodes_to_process: unit_id = str(current_unit["id"]) # Calculate combined weight event_date = current_unit["event_date"] days_since = (utcnow() - event_date).total_seconds() / 86400 recency_weight = calculate_recency_weight(days_since) frequency_weight = calculate_frequency_weight(current_unit.get("access_count", 0)) # Normalize frequency to [0, 1] range frequency_normalized = (frequency_weight - 1.0) / 1.0 # Calculate semantic similarity between query and this memory # Get embedding from the map we fetched memory_embedding = embedding_map.get(unit_id) if memory_embedding is not None: # Convert embedding to list of floats if it's a string or other type if isinstance(memory_embedding, str): import json memory_embedding = json.loads(memory_embedding) elif not isinstance(memory_embedding, (list, np.ndarray)): # If it's some other type, try to convert it memory_embedding = list(memory_embedding) # Cosine similarity = 1 - cosine distance query_vec = np.array(query_embedding, dtype=np.float64) memory_vec = np.array(memory_embedding, dtype=np.float64) # Cosine similarity dot_product = np.dot(query_vec, memory_vec) norm_query = np.linalg.norm(query_vec) norm_memory = np.linalg.norm(memory_vec) semantic_similarity = dot_product / (norm_query * norm_memory) if norm_query > 0 and norm_memory > 0 else 0.0 else: semantic_similarity = 0.0 # Combined weight using configurable parameters final_weight = ( weight_activation * activation + weight_semantic * semantic_similarity + weight_recency * recency_weight + weight_frequency * frequency_normalized ) # Notify tracer if tracer: tracer.visit_node( node_id=unit_id, text=current_unit["text"], context=current_unit.get("context", ""), event_date=event_date, access_count=current_unit.get("access_count", 0), is_entry_point=is_entry_point, parent_node_id=parent_node_id, link_type=parent_link_type, link_weight=parent_link_weight, activation=activation, semantic_similarity=semantic_similarity, recency=recency_weight, frequency=frequency_normalized, final_weight=final_weight, ) results.append({ "id": unit_id, "text": current_unit["text"], "context": current_unit.get("context", ""), "event_date": event_date.isoformat(), "weight": final_weight, "activation": activation, "semantic_similarity": semantic_similarity, "recency": recency_weight, "frequency": frequency_weight, "embedding": memory_embedding, # Store for MMR }) # Spread to neighbors (from batch query results) neighbors = neighbors_by_node.get(unit_id, []) # Group neighbors by to_unit_id to handle multiple connections neighbors_grouped = {} for neighbor in neighbors: neighbor_id = str(neighbor["to_unit_id"]) if neighbor_id not in neighbors_grouped: neighbors_grouped[neighbor_id] = [] neighbors_grouped[neighbor_id].append(neighbor) # Process each unique neighbor (aggregating multiple links) for neighbor_id, neighbor_links in neighbors_grouped.items(): if neighbor_id in visited: continue # Sort links by weight descending to identify primary link neighbor_links_sorted = sorted(neighbor_links, key=lambda x: x["weight"], reverse=True) primary_link = neighbor_links_sorted[0] # Aggregate link weights: max + 30% bonus for additional links max_weight = primary_link["weight"] bonus_weight = sum(link["weight"] for link in neighbor_links_sorted[1:]) * 0.3 combined_weight = max_weight + bonus_weight # Calculate new activation using combined weight new_activation = activation * combined_weight * 0.8 # 0.8 = decay factor # Use primary link metadata for queue and trace primary_link_type = primary_link["link_type"] primary_entity_id = str(primary_link["entity_id"]) if primary_link["entity_id"] else None if new_activation > 0.1: queue.append(({ "id": primary_link["to_unit_id"], "text": primary_link["text"], "context": primary_link.get("context", ""), "event_date": primary_link["event_date"], "access_count": primary_link["access_count"], }, new_activation, False, unit_id, primary_link_type, combined_weight)) # parent_id, link_type, combined_weight # Record all links in trace (primary + additional) if tracer: # Add primary link with combined activation tracer.add_neighbor_link( from_node_id=unit_id, to_node_id=neighbor_id, link_type=primary_link_type, link_weight=combined_weight, entity_id=primary_entity_id, new_activation=new_activation, followed=True ) # Add additional links as supplementary (if multiple connections exist) for additional_link in neighbor_links_sorted[1:]: additional_link_type = additional_link["link_type"] additional_entity_id = str(additional_link["entity_id"]) if additional_link["entity_id"] else None tracer.add_neighbor_link( from_node_id=unit_id, to_node_id=neighbor_id, link_type=additional_link_type, link_weight=additional_link["weight"], entity_id=additional_entity_id, new_activation=None, # Don't show activation for supplementary links followed=True, is_supplementary=True # Mark as supplementary link ) elif tracer: # Record pruned link tracer.add_neighbor_link( from_node_id=unit_id, to_node_id=neighbor_id, link_type=primary_link_type, link_weight=combined_weight, entity_id=primary_entity_id, new_activation=new_activation, followed=False, prune_reason="activation_too_low" ) calculate_weight_time += time.time() - substep_start process_neighbors_time += time.time() - substep_start # Clear batch for next iteration nodes_to_process = [] spreading_activation_time = time.time() - step_start num_batches = (len(visited) + BATCH_SIZE - 1) // BATCH_SIZE # Ceiling division log_buffer.append(f" [3] Spreading activation: {len(visited)} nodes visited in {spreading_activation_time:.3f}s") log_buffer.append(f" [3.1] Calculate weights: {calculate_weight_time:.3f}s") log_buffer.append(f" [3.2] Query neighbors: {query_neighbors_time:.3f}s ({num_batches} batched queries)") log_buffer.append(f" [3.3] Process neighbors: {process_neighbors_time:.3f}s") if tracer: tracer.add_phase_metric("spreading_activation", spreading_activation_time, { "nodes_visited": len(visited), "num_batches": num_batches }) # Step 4: Queue access count updates (background worker will process them) if visited_node_ids: await self._access_count_queue.put(visited_node_ids) log_buffer.append(f" [4] Queued access count updates for {len(visited_node_ids)} nodes") # Step 5: Sort by final weight and apply MMR for diversity step_start = time.time() results.sort(key=lambda x: x["weight"], reverse=True) # Apply MMR (Maximal Marginal Relevance) for diversity if lambda < 1.0 if mmr_lambda < 1.0 and len(results) > top_k: top_results = self._apply_mmr(results, top_k, mmr_lambda, log_buffer) log_buffer.append(f" [5] MMR diversification (λ={mmr_lambda}): {time.time() - step_start:.3f}s") else: top_results = results[:top_k] # Add original rank and remove embeddings from results for idx, result in enumerate(top_results): result["original_rank"] = idx + 1 result["mmr_score"] = None result["mmr_relevance"] = None result["mmr_max_similarity"] = None result["mmr_diversified"] = False result.pop("embedding", None) log_buffer.append(f" [5] Sort and return top {top_k} (no MMR): {time.time() - step_start:.3f}s") total_time = time.time() - search_start log_buffer.append(f"[SEARCH {search_id}] Complete: {len(top_results)} results in {total_time:.3f}s") # Log all buffered logs at once logger.info("\n" + "\n".join(log_buffer)) # Finalize trace if enabled if tracer: trace = tracer.finalize(top_results) return top_results, trace return top_results, None except Exception as e: log_buffer.append(f"[SEARCH {search_id}] ERROR after {time.time() - search_start:.3f}s: {str(e)}") logger.error("\n" + "\n".join(log_buffer)) raise Exception(f"Failed to search memories: {str(e)}") def _apply_mmr( self, results: List[Dict[str, Any]], top_k: int, mmr_lambda: float, log_buffer: List[str] ) -> List[Dict[str, Any]]: """ Apply Maximal Marginal Relevance (MMR) to diversify search results. MMR balances relevance with diversity by selecting results that are: 1. Relevant to the query (high score) 2. Different from already selected results (low similarity) Formula: MMR = λ * relevance - (1-λ) * max_similarity_to_selected Args: results: Sorted list of all results with embeddings top_k: Number of results to select mmr_lambda: Balance parameter (0=max diversity, 1=max relevance) log_buffer: Logging buffer Returns: Diversified list of top_k results """ if not results or top_k <= 0: return [] # Normalize weights to [0, 1] for fair comparison with similarity max_weight = max(r["weight"] for r in results) min_weight = min(r["weight"] for r in results) weight_range = max_weight - min_weight if max_weight > min_weight else 1.0 # Pre-compute normalized relevance scores for all results for idx, result in enumerate(results): result["original_rank"] = idx + 1 result["normalized_relevance"] = (result["weight"] - min_weight) / weight_range # Extract embeddings as a numpy array for vectorized operations # Shape: (num_results, embedding_dim) embeddings_list = [] valid_indices = [] for idx, result in enumerate(results): if result.get("embedding") is not None: embeddings_list.append(result["embedding"]) valid_indices.append(idx) if not embeddings_list: # No embeddings available, just return top-k by relevance return results[:top_k] # Stack embeddings into a matrix (num_results, embedding_dim) embeddings_matrix = np.array(embeddings_list, dtype=np.float32) # Normalize embeddings for faster cosine similarity (just dot product after normalization) norms = np.linalg.norm(embeddings_matrix, axis=1, keepdims=True) norms[norms == 0] = 1.0 # Avoid division by zero embeddings_matrix = embeddings_matrix / norms selected_indices = [] remaining_indices = list(range(len(results))) diversified_count = 0 for selection_round in range(min(top_k, len(results))): if not remaining_indices: break best_mmr_score = float('-inf') best_remaining_idx = 0 # Vectorized computation for all remaining candidates for remaining_idx, candidate_idx in enumerate(remaining_indices): candidate = results[candidate_idx] normalized_relevance = candidate["normalized_relevance"] # Calculate max similarity to selected results max_similarity = 0.0 if selected_indices and candidate_idx in valid_indices: # Find position in embeddings_matrix embedding_idx = valid_indices.index(candidate_idx) candidate_embedding = embeddings_matrix[embedding_idx] # Vectorized similarity calculation with all selected embeddings if selected_indices: selected_embedding_indices = [valid_indices.index(idx) for idx in selected_indices if idx in valid_indices] if selected_embedding_indices: selected_embeddings = embeddings_matrix[selected_embedding_indices] # Compute cosine similarities in one operation (already normalized, so just dot product) similarities = np.dot(selected_embeddings, candidate_embedding) max_similarity = float(np.max(similarities)) # MMR score: balance relevance and diversity mmr_score = mmr_lambda * normalized_relevance - (1 - mmr_lambda) * max_similarity if mmr_score > best_mmr_score: best_mmr_score = mmr_score best_remaining_idx = remaining_idx best_max_similarity = max_similarity # Select the best candidate best_candidate_idx = remaining_indices.pop(best_remaining_idx) best_candidate = results[best_candidate_idx] # Store MMR metadata best_candidate["mmr_score"] = best_mmr_score best_candidate["mmr_relevance"] = best_candidate["normalized_relevance"] best_candidate["mmr_max_similarity"] = best_max_similarity best_candidate["mmr_diversified"] = best_remaining_idx > 0 selected_indices.append(best_candidate_idx) if best_remaining_idx > 0: diversified_count += 1 log_buffer.append(f" MMR: Selected {len(selected_indices)} results, {diversified_count} diversified picks") # Return selected results in order selected_results = [results[idx] for idx in selected_indices] # Remove embeddings from final results (not needed in response) for result in selected_results: result.pop("embedding", None) result.pop("normalized_relevance", None) # Clean up temp field return selected_results async def get_document(self, document_id: str, agent_id: str) -> Optional[Dict[str, Any]]: """ Retrieve document metadata and statistics. Args: document_id: Document ID to retrieve agent_id: Agent ID that owns the document Returns: Dictionary with document info or None if not found """ pool = await self._get_pool() async with pool.acquire() as conn: doc = await conn.fetchrow( """ SELECT d.id, d.agent_id, d.original_text, d.content_hash, d.metadata, d.created_at, d.updated_at, COUNT(mu.id) as unit_count FROM documents d LEFT JOIN memory_units mu ON mu.document_id = d.id WHERE d.id = $1 AND d.agent_id = $2 GROUP BY d.id, d.agent_id, d.original_text, d.content_hash, d.metadata, d.created_at, d.updated_at """, document_id, agent_id ) if not doc: return None import json return { "id": doc["id"], "agent_id": doc["agent_id"], "original_text": doc["original_text"], "content_hash": doc["content_hash"], "metadata": json.loads(doc["metadata"]) if doc["metadata"] else {}, "unit_count": doc["unit_count"], "created_at": doc["created_at"], "updated_at": doc["updated_at"] } async def delete_document(self, document_id: str, agent_id: str) -> Dict[str, int]: """ Delete a document and all its associated memory units and links. Args: document_id: Document ID to delete agent_id: Agent ID that owns the document Returns: Dictionary with counts of deleted items """ pool = await self._get_pool() async with pool.acquire() as conn: async with conn.transaction(): # Count units before deletion units_count = await conn.fetchval( "SELECT COUNT(*) FROM memory_units WHERE document_id = $1", document_id ) # Delete document (cascades to memory_units and all their links) deleted = await conn.fetchval( "DELETE FROM documents WHERE id = $1 AND agent_id = $2 RETURNING id", document_id, agent_id ) return { "document_deleted": 1 if deleted else 0, "memory_units_deleted": units_count if deleted else 0 } async def delete_agent(self, agent_id: str) -> Dict[str, int]: """ Delete all data for a specific agent (multi-tenant cleanup). This is much more efficient than dropping all tables and allows multiple agents to coexist in the same database. Deletes (with CASCADE): - All memory units for this agent - All entities for this agent - All associated links, unit-entity associations, and co-occurrences Args: agent_id: Agent ID to delete Returns: Dictionary with counts of deleted items """ pool = await self._get_pool() async with pool.acquire() as conn: async with conn.transaction(): try: # Count before deletion for reporting units_count = await conn.fetchval("SELECT COUNT(*) FROM memory_units WHERE agent_id = $1", agent_id) entities_count = await conn.fetchval("SELECT COUNT(*) FROM entities WHERE agent_id = $1", agent_id) # Delete memory units (cascades to unit_entities, memory_links) await conn.execute("DELETE FROM memory_units WHERE agent_id = $1", agent_id) # Delete entities (cascades to unit_entities, entity_cooccurrences, memory_links with entity_id) await conn.execute("DELETE FROM entities WHERE agent_id = $1", agent_id) return { "memory_units_deleted": units_count, "entities_deleted": entities_count } except Exception as e: raise Exception(f"Failed to delete agent data: {str(e)}") async def list_agents(self) -> List[str]: """ Get list of all agent IDs in the database. Returns: List of agent IDs """ pool = await self._get_pool() async with pool.acquire() as conn: # Get distinct agent IDs from memory_units agents = await conn.fetch(""" SELECT DISTINCT agent_id FROM memory_units WHERE agent_id IS NOT NULL ORDER BY agent_id """) return [row['agent_id'] for row in agents] async def get_graph_data(self, agent_id: Optional[str] = None, fact_type: Optional[str] = None): """ Get graph data for visualization. Args: agent_id: Filter by agent ID fact_type: Filter by fact type (world, agent, opinion) Returns: Dict with nodes, edges, and table_rows """ pool = await self._get_pool() async with pool.acquire() as conn: # Get memory units, optionally filtered by agent_id and fact_type query_conditions = [] query_params = [] param_count = 0 if agent_id: param_count += 1 query_conditions.append(f"agent_id = ${param_count}") query_params.append(agent_id) if fact_type: param_count += 1 query_conditions.append(f"fact_type = ${param_count}") query_params.append(fact_type) where_clause = "WHERE " + " AND ".join(query_conditions) if query_conditions else "" units = await conn.fetch(f""" SELECT id, text, event_date, context FROM memory_units {where_clause} ORDER BY event_date """, *query_params) # Get links, filtering to only include links between units of the selected agent unit_ids = [row['id'] for row in units] if unit_ids: links = await conn.fetch(""" SELECT ml.from_unit_id, ml.to_unit_id, ml.link_type, ml.weight, e.canonical_name as entity_name FROM memory_links ml LEFT JOIN entities e ON ml.entity_id = e.id WHERE ml.from_unit_id = ANY($1::uuid[]) AND ml.to_unit_id = ANY($1::uuid[]) ORDER BY ml.link_type, ml.weight DESC """, unit_ids) else: links = [] # Get entity information unit_entities = await conn.fetch(""" SELECT ue.unit_id, e.canonical_name, e.entity_type FROM unit_entities ue JOIN entities e ON ue.entity_id = e.id ORDER BY ue.unit_id """) # Build entity mapping entity_map = {} for row in unit_entities: unit_id = row['unit_id'] entity_name = row['canonical_name'] entity_type = row['entity_type'] if unit_id not in entity_map: entity_map[unit_id] = [] entity_map[unit_id].append(f"{entity_name} ({entity_type})") # Build nodes nodes = [] for row in units: unit_id = row['id'] text = row['text'] event_date = row['event_date'] context = row['context'] entities = entity_map.get(unit_id, []) entity_count = len(entities) # Color by entity count if entity_count == 0: color = "#e0e0e0" elif entity_count == 1: color = "#90caf9" else: color = "#42a5f5" nodes.append({ "data": { "id": str(unit_id), "label": f"{text[:30]}..." if len(text) > 30 else text, "text": text, "date": event_date.isoformat() if event_date else "", "context": context if context else "", "entities": ", ".join(entities) if entities else "None", "color": color } }) # Build edges edges = [] for row in links: from_id = str(row['from_unit_id']) to_id = str(row['to_unit_id']) link_type = row['link_type'] weight = row['weight'] entity_name = row['entity_name'] # Color by link type if link_type == 'temporal': color = "#00bcd4" line_style = "dashed" elif link_type == 'semantic': color = "#ff69b4" line_style = "solid" elif link_type == 'entity': color = "#ffd700" line_style = "solid" else: color = "#999999" line_style = "solid" edges.append({ "data": { "id": f"{from_id}-{to_id}-{link_type}", "source": from_id, "target": to_id, "linkType": link_type, "weight": weight, "entityName": entity_name if entity_name else "", "color": color, "lineStyle": line_style } }) # Build table rows table_rows = [] for row in units: unit_id = row['id'] entities = entity_map.get(unit_id, []) table_rows.append({ "id": str(unit_id)[:8] + "...", "text": row['text'], "context": row['context'] if row['context'] else "N/A", "date": row['event_date'].strftime("%Y-%m-%d %H:%M") if row['event_date'] else "N/A", "entities": ", ".join(entities) if entities else "None" }) return { "nodes": nodes, "edges": edges, "table_rows": table_rows, "total_units": len(units) } async def _evaluate_opinion_update_async( self, client, opinion_text: str, opinion_confidence: float, new_event_text: str, entity_name: str, model: str = "openai/gpt-oss-120b", ) -> Optional[Dict[str, Any]]: """ Evaluate if an opinion should be updated based on a new event. Args: client: OpenAI client opinion_text: Current opinion text (includes reasons) opinion_confidence: Current confidence score (0.0-1.0) new_event_text: Text of the new event entity_name: Name of the entity this opinion is about model: LLM model to use Returns: Dict with 'action' ('keep'|'update'), 'new_confidence', 'new_text' (if action=='update') or None if no changes needed """ from pydantic import BaseModel, Field class OpinionEvaluation(BaseModel): """Evaluation of whether an opinion should be updated.""" action: str = Field(description="Action to take: 'keep' (no change) or 'update' (modify opinion)") reasoning: str = Field(description="Brief explanation of why this action was chosen") new_confidence: float = Field(description="New confidence score (0.0-1.0). Can be higher, lower, or same as before.") new_opinion_text: Optional[str] = Field( default=None, description="If action is 'update', the revised opinion text that acknowledges the previous view. Otherwise None." ) evaluation_prompt = f"""You are evaluating whether an existing opinion should be updated based on new information. ENTITY: {entity_name} EXISTING OPINION: {opinion_text} Current confidence: {opinion_confidence:.2f} NEW EVENT: {new_event_text} Evaluate whether this new event: 1. REINFORCES the opinion (increase confidence, keep text) 2. WEAKENS the opinion (decrease confidence, keep text) 3. CHANGES the opinion (update both text and confidence, noting "Previously I thought X, but now Y...") 4. IRRELEVANT (keep everything as is) Guidelines: - Only suggest 'update' action if the new event genuinely contradicts or significantly modifies the opinion - If updating the text, acknowledge the previous opinion and explain the change - Confidence should reflect accumulated evidence (0.0 = no confidence, 1.0 = very confident) - Small changes in confidence are normal; large jumps should be rare""" try: response = await client.beta.chat.completions.parse( model=model, messages=[ {"role": "system", "content": "You evaluate and update opinions based on new information."}, {"role": "user", "content": evaluation_prompt} ], response_format=OpinionEvaluation, temperature=0.3 # Lower temperature for more consistent evaluation ) result = response.choices[0].message.parsed # Only return updates if something actually changed if result.action == 'keep' and abs(result.new_confidence - opinion_confidence) < 0.01: return None return { 'action': result.action, 'reasoning': result.reasoning, 'new_confidence': result.new_confidence, 'new_text': result.new_opinion_text if result.action == 'update' else None } except Exception as e: logger.warning(f"Failed to evaluate opinion update: {str(e)}") return None async def _reinforce_opinions_async( self, agent_id: str, created_unit_ids: List[str], unit_texts: List[str], unit_entities: List[List[Dict[str, str]]], model: str = "openai/gpt-oss-120b", ): """ Background task to reinforce opinions based on newly ingested events. This runs asynchronously and does not block the put operation. Args: agent_id: Agent ID created_unit_ids: List of newly created memory unit IDs unit_texts: Texts of the newly created units unit_entities: Entities extracted from each unit model: LLM model to use for evaluation """ try: # Extract all unique entity names from the new units entity_names = set() for entities_list in unit_entities: for entity in entities_list: entity_names.add(entity['text']) if not entity_names: logger.debug("[REINFORCE] No entities found in new units, skipping opinion reinforcement") return logger.debug(f"[REINFORCE] Starting opinion reinforcement for {len(entity_names)} entities") pool = await self._get_pool() async with pool.acquire() as conn: # Find all opinions related to these entities opinions = await conn.fetch( """ SELECT DISTINCT mu.id, mu.text, mu.confidence_score, e.canonical_name FROM memory_units mu JOIN unit_entities ue ON mu.id = ue.unit_id JOIN entities e ON ue.entity_id = e.id WHERE mu.agent_id = $1 AND mu.fact_type = 'opinion' AND e.canonical_name = ANY($2::text[]) """, agent_id, list(entity_names) ) if not opinions: logger.debug("[REINFORCE] No existing opinions found for these entities") return logger.debug(f"[REINFORCE] Found {len(opinions)} opinions to potentially reinforce") # Use cached LLM client if self._llm_client is None: logger.error("[REINFORCE] LLM client not available, skipping opinion reinforcement") return client = self._llm_client # Evaluate each opinion against the new events updates_to_apply = [] for opinion in opinions: opinion_id = str(opinion['id']) opinion_text = opinion['text'] opinion_confidence = opinion['confidence_score'] entity_name = opinion['canonical_name'] # Find all new events mentioning this entity relevant_events = [] for unit_text, entities_list in zip(unit_texts, unit_entities): if any(e['text'] == entity_name for e in entities_list): relevant_events.append(unit_text) if not relevant_events: continue # Combine all relevant events combined_events = "\n".join(relevant_events) # Evaluate if opinion should be updated evaluation = await self._evaluate_opinion_update_async( client, opinion_text, opinion_confidence, combined_events, entity_name, model ) if evaluation: updates_to_apply.append({ 'opinion_id': opinion_id, 'evaluation': evaluation }) # Apply all updates in a single transaction if updates_to_apply: async with conn.transaction(): for update in updates_to_apply: opinion_id = update['opinion_id'] evaluation = update['evaluation'] if evaluation['action'] == 'update' and evaluation['new_text']: # Update both text and confidence await conn.execute( """ UPDATE memory_units SET text = $1, confidence_score = $2, updated_at = NOW() WHERE id = $3 """, evaluation['new_text'], evaluation['new_confidence'], uuid.UUID(opinion_id) ) logger.debug(f"[REINFORCE] Updated opinion {opinion_id[:8]}... (action: {evaluation['action']}, confidence: {evaluation['new_confidence']:.2f})") else: # Only update confidence await conn.execute( """ UPDATE memory_units SET confidence_score = $1, updated_at = NOW() WHERE id = $2 """, evaluation['new_confidence'], uuid.UUID(opinion_id) ) logger.debug(f"[REINFORCE] Updated confidence for opinion {opinion_id[:8]}... (confidence: {evaluation['new_confidence']:.2f})") logger.debug(f"[REINFORCE] Applied {len(updates_to_apply)} opinion updates") else: logger.debug("[REINFORCE] No opinion updates needed") except Exception as e: logger.error(f"[REINFORCE] Error during opinion reinforcement: {str(e)}") import traceback traceback.print_exc()