hindsight-api 0.4.1__py3-none-any.whl → 0.4.3__py3-none-any.whl

hindsight_api/__init__.py

@@ -46,4 +46,4 @@ __all__ = [
     "RemoteTEICrossEncoder",
     "LLMConfig",
 ]
-__version__ = "0.4.1"
+__version__ = "0.4.3"

hindsight_api/api/http.py

@@ -92,8 +92,7 @@ class RecallRequest(BaseModel):
     query: str
     types: list[str] | None = Field(
         default=None,
-        description="List of fact types to recall: 'world', 'experience', 'observation'. Defaults to world and experience if not specified. "
-        "Note: 'opinion' is accepted but ignored (opinions are excluded from recall).",
+        description="List of fact types to recall: 'world', 'experience', 'observation'. Defaults to world and experience if not specified.",
     )
     budget: Budget = Budget.MID
     max_tokens: int = 4096
@@ -504,13 +503,6 @@ class ReflectRequest(BaseModel):
     )
 
 
-class OpinionItem(BaseModel):
-    """Model for an opinion with confidence score."""
-
-    text: str
-    confidence: float
-
-
 class ReflectFact(BaseModel):
     """A fact used in think response."""
 
@@ -529,7 +521,7 @@ class ReflectFact(BaseModel):
 
     id: str | None = None
     text: str
-    type: str | None = None  # fact type: world, experience, opinion
+    type: str | None = None  # fact type: world, experience, observation
     context: str | None = None
     occurred_start: str | None = None
     occurred_end: str | None = None
@@ -1412,9 +1404,10 @@ def create_app(
                 worker_id=worker_id,
                 executor=memory.execute_task,
                 poll_interval_ms=config.worker_poll_interval_ms,
-                batch_size=config.worker_batch_size,
                 max_retries=config.worker_max_retries,
                 tenant_extension=getattr(memory, "_tenant_extension", None),
+                max_slots=config.worker_max_slots,
+                consolidation_max_slots=config.worker_consolidation_max_slots,
             )
             poller_task = asyncio.create_task(poller.run())
             logging.info(f"Worker poller started (worker_id={worker_id})")
@@ -1707,9 +1700,7 @@ def _register_routes(app: FastAPI):
         description="Recall memory using semantic similarity and spreading activation.\n\n"
         "The type parameter is optional and must be one of:\n"
         "- `world`: General knowledge about people, places, events, and things that happen\n"
-        "- `experience`: Memories about experience, conversations, actions taken, and tasks performed\n"
-        "- `opinion`: The bank's formed beliefs, perspectives, and viewpoints\n\n"
-        "Set `include_entities=true` to get entity observations alongside recall results.",
+        "- `experience`: Memories about experience, conversations, actions taken, and tasks performed",
         operation_id="recall_memories",
         tags=["Memory"],
     )
@@ -1723,10 +1714,8 @@ def _register_routes(app: FastAPI):
         metrics = get_metrics_collector()
 
         try:
-            # Default to world and experience if not specified (exclude observation and opinion)
-            # Filter out 'opinion' even if requested - opinions are excluded from recall
+            # Default to world and experience if not specified (exclude observation)
             fact_types = request.types if request.types else list(VALID_RECALL_FACT_TYPES)
-            fact_types = [ft for ft in fact_types if ft != "opinion"]
 
             # Parse query_timestamp if provided
             question_date = None
@@ -1858,8 +1847,7 @@ def _register_routes(app: FastAPI):
         "2. Retrieves world facts relevant to the query\n"
         "3. Retrieves existing opinions (bank's perspectives)\n"
         "4. Uses LLM to formulate a contextual answer\n"
-        "5. Extracts and stores any new opinions formed\n"
-        "6. Returns plain text answer, the facts used, and new opinions",
+        "5. Returns plain text answer and the facts used",
         operation_id="reflect",
         tags=["Memory"],
     )

hindsight_api/api/mcp.py

@@ -29,15 +29,26 @@ logger = logging.getLogger(__name__)
 # Default bank_id from environment variable
 DEFAULT_BANK_ID = os.environ.get("HINDSIGHT_MCP_BANK_ID", "default")
 
+# MCP authentication token (optional - if set, Bearer token auth is required)
+MCP_AUTH_TOKEN = os.environ.get("HINDSIGHT_API_MCP_AUTH_TOKEN")
+
 # Context variable to hold the current bank_id
 _current_bank_id: ContextVar[str | None] = ContextVar("current_bank_id", default=None)
 
+# Context variable to hold the current API key (for tenant auth propagation)
+_current_api_key: ContextVar[str | None] = ContextVar("current_api_key", default=None)
+
 
 def get_current_bank_id() -> str | None:
     """Get the current bank_id from context."""
     return _current_bank_id.get()
 
 
+def get_current_api_key() -> str | None:
+    """Get the current API key from context."""
+    return _current_api_key.get()
+
+
 def create_mcp_server(memory: MemoryEngine) -> FastMCP:
     """
     Create and configure the Hindsight MCP server.
@@ -54,6 +65,7 @@ def create_mcp_server(memory: MemoryEngine) -> FastMCP:
     # Configure and register tools using shared module
     config = MCPToolsConfig(
         bank_id_resolver=get_current_bank_id,
+        api_key_resolver=get_current_api_key,  # Propagate API key for tenant auth
         include_bank_id_param=True,  # HTTP MCP supports multi-bank via parameter
         tools=None,  # All tools
         retain_fire_and_forget=False,  # HTTP MCP supports sync/async modes
@@ -65,7 +77,11 @@ def create_mcp_server(memory: MemoryEngine) -> FastMCP:
 
 
 class MCPMiddleware:
-    """ASGI middleware that extracts bank_id from header or path and sets context.
+    """ASGI middleware that handles authentication and extracts bank_id from header or path.
+
+    Authentication:
+        If HINDSIGHT_API_MCP_AUTH_TOKEN is set, all requests must include a valid
+        Authorization header with Bearer token or direct token matching the configured value.
 
     Bank ID can be provided via:
     1. X-Bank-Id header (recommended for Claude Code)
@@ -74,7 +90,7 @@ class MCPMiddleware:
 
     For Claude Code, configure with:
         claude mcp add --transport http hindsight http://localhost:8888/mcp \\
-            --header "X-Bank-Id: my-bank"
+            --header "X-Bank-Id: my-bank" --header "Authorization: Bearer <token>"
     """
 
     def __init__(self, app, memory: MemoryEngine):
@@ -98,6 +114,22 @@ class MCPMiddleware:
             await self.mcp_app(scope, receive, send)
             return
 
+        # Extract auth token from header (for tenant auth propagation)
+        auth_header = self._get_header(scope, "Authorization")
+        auth_token: str | None = None
+        if auth_header:
+            # Support both "Bearer <token>" and direct token
+            auth_token = auth_header[7:].strip() if auth_header.startswith("Bearer ") else auth_header.strip()
+
+        # Authenticate if MCP_AUTH_TOKEN is configured
+        if MCP_AUTH_TOKEN:
+            if not auth_token:
+                await self._send_error(send, 401, "Authorization header required")
+                return
+            if auth_token != MCP_AUTH_TOKEN:
+                await self._send_error(send, 401, "Invalid authentication token")
+                return
+
         path = scope.get("path", "")
 
         # Strip any mount prefix (e.g., /mcp) that FastAPI might not have stripped
@@ -132,8 +164,10 @@ class MCPMiddleware:
             bank_id = DEFAULT_BANK_ID
             logger.debug(f"Using default bank_id: {bank_id}")
 
-        # Set bank_id context
-        token = _current_bank_id.set(bank_id)
+        # Set bank_id and api_key context
+        bank_id_token = _current_bank_id.set(bank_id)
+        # Store the auth token for tenant extension to validate
+        api_key_token = _current_api_key.set(auth_token) if auth_token else None
         try:
             new_scope = scope.copy()
             new_scope["path"] = new_path
@@ -152,7 +186,9 @@ class MCPMiddleware:
 
             await self.mcp_app(new_scope, receive, send_wrapper)
         finally:
-            _current_bank_id.reset(token)
+            _current_bank_id.reset(bank_id_token)
+            if api_key_token is not None:
+                _current_api_key.reset(api_key_token)
 
     async def _send_error(self, send, status: int, message: str):
         """Send an error response."""
@@ -176,6 +212,10 @@ def create_mcp_app(memory: MemoryEngine):
     """
     Create an ASGI app that handles MCP requests.
 
+    Authentication:
+        Set HINDSIGHT_API_MCP_AUTH_TOKEN to require Bearer token authentication.
+        If not set, MCP endpoint is open (for local development).
+
     Bank ID can be provided via:
     1. X-Bank-Id header: claude mcp add --transport http hindsight http://localhost:8888/mcp --header "X-Bank-Id: my-bank"
     2. URL path: /mcp/{bank_id}/

hindsight_api/config.py

@@ -26,6 +26,9 @@ ENV_LLM_API_KEY = "HINDSIGHT_API_LLM_API_KEY"
 ENV_LLM_MODEL = "HINDSIGHT_API_LLM_MODEL"
 ENV_LLM_BASE_URL = "HINDSIGHT_API_LLM_BASE_URL"
 ENV_LLM_MAX_CONCURRENT = "HINDSIGHT_API_LLM_MAX_CONCURRENT"
+ENV_LLM_MAX_RETRIES = "HINDSIGHT_API_LLM_MAX_RETRIES"
+ENV_LLM_INITIAL_BACKOFF = "HINDSIGHT_API_LLM_INITIAL_BACKOFF"
+ENV_LLM_MAX_BACKOFF = "HINDSIGHT_API_LLM_MAX_BACKOFF"
 ENV_LLM_TIMEOUT = "HINDSIGHT_API_LLM_TIMEOUT"
 ENV_LLM_GROQ_SERVICE_TIER = "HINDSIGHT_API_LLM_GROQ_SERVICE_TIER"
 
@@ -34,16 +37,31 @@ ENV_RETAIN_LLM_PROVIDER = "HINDSIGHT_API_RETAIN_LLM_PROVIDER"
 ENV_RETAIN_LLM_API_KEY = "HINDSIGHT_API_RETAIN_LLM_API_KEY"
 ENV_RETAIN_LLM_MODEL = "HINDSIGHT_API_RETAIN_LLM_MODEL"
 ENV_RETAIN_LLM_BASE_URL = "HINDSIGHT_API_RETAIN_LLM_BASE_URL"
+ENV_RETAIN_LLM_MAX_CONCURRENT = "HINDSIGHT_API_RETAIN_LLM_MAX_CONCURRENT"
+ENV_RETAIN_LLM_MAX_RETRIES = "HINDSIGHT_API_RETAIN_LLM_MAX_RETRIES"
+ENV_RETAIN_LLM_INITIAL_BACKOFF = "HINDSIGHT_API_RETAIN_LLM_INITIAL_BACKOFF"
+ENV_RETAIN_LLM_MAX_BACKOFF = "HINDSIGHT_API_RETAIN_LLM_MAX_BACKOFF"
+ENV_RETAIN_LLM_TIMEOUT = "HINDSIGHT_API_RETAIN_LLM_TIMEOUT"
 
 ENV_REFLECT_LLM_PROVIDER = "HINDSIGHT_API_REFLECT_LLM_PROVIDER"
 ENV_REFLECT_LLM_API_KEY = "HINDSIGHT_API_REFLECT_LLM_API_KEY"
 ENV_REFLECT_LLM_MODEL = "HINDSIGHT_API_REFLECT_LLM_MODEL"
 ENV_REFLECT_LLM_BASE_URL = "HINDSIGHT_API_REFLECT_LLM_BASE_URL"
+ENV_REFLECT_LLM_MAX_CONCURRENT = "HINDSIGHT_API_REFLECT_LLM_MAX_CONCURRENT"
+ENV_REFLECT_LLM_MAX_RETRIES = "HINDSIGHT_API_REFLECT_LLM_MAX_RETRIES"
+ENV_REFLECT_LLM_INITIAL_BACKOFF = "HINDSIGHT_API_REFLECT_LLM_INITIAL_BACKOFF"
+ENV_REFLECT_LLM_MAX_BACKOFF = "HINDSIGHT_API_REFLECT_LLM_MAX_BACKOFF"
+ENV_REFLECT_LLM_TIMEOUT = "HINDSIGHT_API_REFLECT_LLM_TIMEOUT"
 
 ENV_CONSOLIDATION_LLM_PROVIDER = "HINDSIGHT_API_CONSOLIDATION_LLM_PROVIDER"
 ENV_CONSOLIDATION_LLM_API_KEY = "HINDSIGHT_API_CONSOLIDATION_LLM_API_KEY"
 ENV_CONSOLIDATION_LLM_MODEL = "HINDSIGHT_API_CONSOLIDATION_LLM_MODEL"
 ENV_CONSOLIDATION_LLM_BASE_URL = "HINDSIGHT_API_CONSOLIDATION_LLM_BASE_URL"
+ENV_CONSOLIDATION_LLM_MAX_CONCURRENT = "HINDSIGHT_API_CONSOLIDATION_LLM_MAX_CONCURRENT"
+ENV_CONSOLIDATION_LLM_MAX_RETRIES = "HINDSIGHT_API_CONSOLIDATION_LLM_MAX_RETRIES"
+ENV_CONSOLIDATION_LLM_INITIAL_BACKOFF = "HINDSIGHT_API_CONSOLIDATION_LLM_INITIAL_BACKOFF"
+ENV_CONSOLIDATION_LLM_MAX_BACKOFF = "HINDSIGHT_API_CONSOLIDATION_LLM_MAX_BACKOFF"
+ENV_CONSOLIDATION_LLM_TIMEOUT = "HINDSIGHT_API_CONSOLIDATION_LLM_TIMEOUT"
 
 ENV_EMBEDDINGS_PROVIDER = "HINDSIGHT_API_EMBEDDINGS_PROVIDER"
 ENV_EMBEDDINGS_LOCAL_MODEL = "HINDSIGHT_API_EMBEDDINGS_LOCAL_MODEL"
@@ -90,13 +108,17 @@ ENV_MCP_LOCAL_BANK_ID = "HINDSIGHT_API_MCP_LOCAL_BANK_ID"
 ENV_MCP_INSTRUCTIONS = "HINDSIGHT_API_MCP_INSTRUCTIONS"
 ENV_MENTAL_MODEL_REFRESH_CONCURRENCY = "HINDSIGHT_API_MENTAL_MODEL_REFRESH_CONCURRENCY"
 
+# Vertex AI configuration
+ENV_LLM_VERTEXAI_PROJECT_ID = "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID"
+ENV_LLM_VERTEXAI_REGION = "HINDSIGHT_API_LLM_VERTEXAI_REGION"
+ENV_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY = "HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY"
+
 # Retain settings
 ENV_RETAIN_MAX_COMPLETION_TOKENS = "HINDSIGHT_API_RETAIN_MAX_COMPLETION_TOKENS"
 ENV_RETAIN_CHUNK_SIZE = "HINDSIGHT_API_RETAIN_CHUNK_SIZE"
 ENV_RETAIN_EXTRACT_CAUSAL_LINKS = "HINDSIGHT_API_RETAIN_EXTRACT_CAUSAL_LINKS"
 ENV_RETAIN_EXTRACTION_MODE = "HINDSIGHT_API_RETAIN_EXTRACTION_MODE"
 ENV_RETAIN_CUSTOM_INSTRUCTIONS = "HINDSIGHT_API_RETAIN_CUSTOM_INSTRUCTIONS"
-ENV_RETAIN_OBSERVATIONS_ASYNC = "HINDSIGHT_API_RETAIN_OBSERVATIONS_ASYNC"
 
 # Observations settings (consolidated knowledge from facts)
 ENV_ENABLE_OBSERVATIONS = "HINDSIGHT_API_ENABLE_OBSERVATIONS"
@@ -121,8 +143,9 @@ ENV_WORKER_ENABLED = "HINDSIGHT_API_WORKER_ENABLED"
 ENV_WORKER_ID = "HINDSIGHT_API_WORKER_ID"
 ENV_WORKER_POLL_INTERVAL_MS = "HINDSIGHT_API_WORKER_POLL_INTERVAL_MS"
 ENV_WORKER_MAX_RETRIES = "HINDSIGHT_API_WORKER_MAX_RETRIES"
-ENV_WORKER_BATCH_SIZE = "HINDSIGHT_API_WORKER_BATCH_SIZE"
 ENV_WORKER_HTTP_PORT = "HINDSIGHT_API_WORKER_HTTP_PORT"
+ENV_WORKER_MAX_SLOTS = "HINDSIGHT_API_WORKER_MAX_SLOTS"
+ENV_WORKER_CONSOLIDATION_MAX_SLOTS = "HINDSIGHT_API_WORKER_CONSOLIDATION_MAX_SLOTS"
 
 # Reflect agent settings
 ENV_REFLECT_MAX_ITERATIONS = "HINDSIGHT_API_REFLECT_MAX_ITERATIONS"
@@ -133,8 +156,16 @@ DEFAULT_DATABASE_SCHEMA = "public"
 DEFAULT_LLM_PROVIDER = "openai"
 DEFAULT_LLM_MODEL = "gpt-5-mini"
 DEFAULT_LLM_MAX_CONCURRENT = 32
+DEFAULT_LLM_MAX_RETRIES = 10  # Max retry attempts for LLM API calls
+DEFAULT_LLM_INITIAL_BACKOFF = 1.0  # Initial backoff in seconds for retry exponential backoff
+DEFAULT_LLM_MAX_BACKOFF = 60.0  # Max backoff cap in seconds for retry exponential backoff
 DEFAULT_LLM_TIMEOUT = 120.0  # seconds
 
+# Vertex AI defaults
+DEFAULT_LLM_VERTEXAI_PROJECT_ID = None  # Required for Vertex AI
+DEFAULT_LLM_VERTEXAI_REGION = "us-central1"
+DEFAULT_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY = None  # Optional, uses ADC if not set
+
 DEFAULT_EMBEDDINGS_PROVIDER = "local"
 DEFAULT_EMBEDDINGS_LOCAL_MODEL = "BAAI/bge-small-en-v1.5"
 DEFAULT_EMBEDDINGS_LOCAL_FORCE_CPU = False  # Force CPU mode for local embeddings (avoids MPS/XPC issues on macOS)
@@ -179,7 +210,6 @@ DEFAULT_RETAIN_EXTRACT_CAUSAL_LINKS = True  # Extract causal links between facts
 DEFAULT_RETAIN_EXTRACTION_MODE = "concise"  # Extraction mode: "concise", "verbose", or "custom"
 RETAIN_EXTRACTION_MODES = ("concise", "verbose", "custom")  # Allowed extraction modes
 DEFAULT_RETAIN_CUSTOM_INSTRUCTIONS = None  # Custom extraction guidelines (only used when mode="custom")
-DEFAULT_RETAIN_OBSERVATIONS_ASYNC = False  # Run observation generation async (after retain completes)
 
 # Observations defaults (consolidated knowledge from facts)
 DEFAULT_ENABLE_OBSERVATIONS = True  # Observations enabled by default
@@ -200,8 +230,9 @@ DEFAULT_WORKER_ENABLED = True  # API runs worker by default (standalone mode)
 DEFAULT_WORKER_ID = None  # Will use hostname if not specified
 DEFAULT_WORKER_POLL_INTERVAL_MS = 500  # Poll database every 500ms
 DEFAULT_WORKER_MAX_RETRIES = 3  # Max retries before marking task failed
-DEFAULT_WORKER_BATCH_SIZE = 10  # Tasks to claim per poll cycle
 DEFAULT_WORKER_HTTP_PORT = 8889  # HTTP port for worker metrics/health
+DEFAULT_WORKER_MAX_SLOTS = 10  # Total concurrent tasks per worker
+DEFAULT_WORKER_CONSOLIDATION_MAX_SLOTS = 2  # Max concurrent consolidation tasks per worker
 
 # Reflect agent settings
 DEFAULT_REFLECT_MAX_ITERATIONS = 10  # Max tool call iterations before forcing response
@@ -286,23 +317,46 @@ class HindsightConfig:
     llm_model: str
     llm_base_url: str | None
     llm_max_concurrent: int
+    llm_max_retries: int
+    llm_initial_backoff: float
+    llm_max_backoff: float
     llm_timeout: float
 
+    # Vertex AI configuration
+    llm_vertexai_project_id: str | None
+    llm_vertexai_region: str
+    llm_vertexai_service_account_key: str | None
+
     # Per-operation LLM configuration (None = use default LLM config)
     retain_llm_provider: str | None
     retain_llm_api_key: str | None
     retain_llm_model: str | None
     retain_llm_base_url: str | None
+    retain_llm_max_concurrent: int | None
+    retain_llm_max_retries: int | None
+    retain_llm_initial_backoff: float | None
+    retain_llm_max_backoff: float | None
+    retain_llm_timeout: float | None
 
     reflect_llm_provider: str | None
     reflect_llm_api_key: str | None
     reflect_llm_model: str | None
     reflect_llm_base_url: str | None
+    reflect_llm_max_concurrent: int | None
+    reflect_llm_max_retries: int | None
+    reflect_llm_initial_backoff: float | None
+    reflect_llm_max_backoff: float | None
+    reflect_llm_timeout: float | None
 
     consolidation_llm_provider: str | None
     consolidation_llm_api_key: str | None
     consolidation_llm_model: str | None
     consolidation_llm_base_url: str | None
+    consolidation_llm_max_concurrent: int | None
+    consolidation_llm_max_retries: int | None
+    consolidation_llm_initial_backoff: float | None
+    consolidation_llm_max_backoff: float | None
+    consolidation_llm_timeout: float | None
 
     # Embeddings
     embeddings_provider: str
@@ -343,7 +397,6 @@ class HindsightConfig:
     retain_extract_causal_links: bool
     retain_extraction_mode: str
     retain_custom_instructions: str | None
-    retain_observations_async: bool
 
     # Observations settings (consolidated knowledge from facts)
     enable_observations: bool
@@ -368,8 +421,9 @@ class HindsightConfig:
     worker_id: str | None
     worker_poll_interval_ms: int
     worker_max_retries: int
-    worker_batch_size: int
     worker_http_port: int
+    worker_max_slots: int
+    worker_consolidation_max_slots: int
 
     # Reflect agent settings
     reflect_max_iterations: int
@@ -387,20 +441,71 @@ class HindsightConfig:
             llm_model=os.getenv(ENV_LLM_MODEL, DEFAULT_LLM_MODEL),
             llm_base_url=os.getenv(ENV_LLM_BASE_URL) or None,
             llm_max_concurrent=int(os.getenv(ENV_LLM_MAX_CONCURRENT, str(DEFAULT_LLM_MAX_CONCURRENT))),
+            llm_max_retries=int(os.getenv(ENV_LLM_MAX_RETRIES, str(DEFAULT_LLM_MAX_RETRIES))),
+            llm_initial_backoff=float(os.getenv(ENV_LLM_INITIAL_BACKOFF, str(DEFAULT_LLM_INITIAL_BACKOFF))),
+            llm_max_backoff=float(os.getenv(ENV_LLM_MAX_BACKOFF, str(DEFAULT_LLM_MAX_BACKOFF))),
             llm_timeout=float(os.getenv(ENV_LLM_TIMEOUT, str(DEFAULT_LLM_TIMEOUT))),
+            # Vertex AI
+            llm_vertexai_project_id=os.getenv(ENV_LLM_VERTEXAI_PROJECT_ID) or DEFAULT_LLM_VERTEXAI_PROJECT_ID,
+            llm_vertexai_region=os.getenv(ENV_LLM_VERTEXAI_REGION, DEFAULT_LLM_VERTEXAI_REGION),
+            llm_vertexai_service_account_key=os.getenv(ENV_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY)
+            or DEFAULT_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY,
             # Per-operation LLM config (None = use default)
             retain_llm_provider=os.getenv(ENV_RETAIN_LLM_PROVIDER) or None,
             retain_llm_api_key=os.getenv(ENV_RETAIN_LLM_API_KEY) or None,
             retain_llm_model=os.getenv(ENV_RETAIN_LLM_MODEL) or None,
             retain_llm_base_url=os.getenv(ENV_RETAIN_LLM_BASE_URL) or None,
+            retain_llm_max_concurrent=int(os.getenv(ENV_RETAIN_LLM_MAX_CONCURRENT))
+            if os.getenv(ENV_RETAIN_LLM_MAX_CONCURRENT)
+            else None,
+            retain_llm_max_retries=int(os.getenv(ENV_RETAIN_LLM_MAX_RETRIES))
+            if os.getenv(ENV_RETAIN_LLM_MAX_RETRIES)
+            else None,
+            retain_llm_initial_backoff=float(os.getenv(ENV_RETAIN_LLM_INITIAL_BACKOFF))
+            if os.getenv(ENV_RETAIN_LLM_INITIAL_BACKOFF)
+            else None,
+            retain_llm_max_backoff=float(os.getenv(ENV_RETAIN_LLM_MAX_BACKOFF))
+            if os.getenv(ENV_RETAIN_LLM_MAX_BACKOFF)
+            else None,
+            retain_llm_timeout=float(os.getenv(ENV_RETAIN_LLM_TIMEOUT)) if os.getenv(ENV_RETAIN_LLM_TIMEOUT) else None,
             reflect_llm_provider=os.getenv(ENV_REFLECT_LLM_PROVIDER) or None,
             reflect_llm_api_key=os.getenv(ENV_REFLECT_LLM_API_KEY) or None,
             reflect_llm_model=os.getenv(ENV_REFLECT_LLM_MODEL) or None,
             reflect_llm_base_url=os.getenv(ENV_REFLECT_LLM_BASE_URL) or None,
+            reflect_llm_max_concurrent=int(os.getenv(ENV_REFLECT_LLM_MAX_CONCURRENT))
+            if os.getenv(ENV_REFLECT_LLM_MAX_CONCURRENT)
+            else None,
+            reflect_llm_max_retries=int(os.getenv(ENV_REFLECT_LLM_MAX_RETRIES))
+            if os.getenv(ENV_REFLECT_LLM_MAX_RETRIES)
+            else None,
+            reflect_llm_initial_backoff=float(os.getenv(ENV_REFLECT_LLM_INITIAL_BACKOFF))
+            if os.getenv(ENV_REFLECT_LLM_INITIAL_BACKOFF)
+            else None,
+            reflect_llm_max_backoff=float(os.getenv(ENV_REFLECT_LLM_MAX_BACKOFF))
+            if os.getenv(ENV_REFLECT_LLM_MAX_BACKOFF)
+            else None,
+            reflect_llm_timeout=float(os.getenv(ENV_REFLECT_LLM_TIMEOUT))
+            if os.getenv(ENV_REFLECT_LLM_TIMEOUT)
+            else None,
             consolidation_llm_provider=os.getenv(ENV_CONSOLIDATION_LLM_PROVIDER) or None,
             consolidation_llm_api_key=os.getenv(ENV_CONSOLIDATION_LLM_API_KEY) or None,
             consolidation_llm_model=os.getenv(ENV_CONSOLIDATION_LLM_MODEL) or None,
             consolidation_llm_base_url=os.getenv(ENV_CONSOLIDATION_LLM_BASE_URL) or None,
+            consolidation_llm_max_concurrent=int(os.getenv(ENV_CONSOLIDATION_LLM_MAX_CONCURRENT))
+            if os.getenv(ENV_CONSOLIDATION_LLM_MAX_CONCURRENT)
+            else None,
+            consolidation_llm_max_retries=int(os.getenv(ENV_CONSOLIDATION_LLM_MAX_RETRIES))
+            if os.getenv(ENV_CONSOLIDATION_LLM_MAX_RETRIES)
+            else None,
+            consolidation_llm_initial_backoff=float(os.getenv(ENV_CONSOLIDATION_LLM_INITIAL_BACKOFF))
+            if os.getenv(ENV_CONSOLIDATION_LLM_INITIAL_BACKOFF)
+            else None,
+            consolidation_llm_max_backoff=float(os.getenv(ENV_CONSOLIDATION_LLM_MAX_BACKOFF))
+            if os.getenv(ENV_CONSOLIDATION_LLM_MAX_BACKOFF)
+            else None,
+            consolidation_llm_timeout=float(os.getenv(ENV_CONSOLIDATION_LLM_TIMEOUT))
+            if os.getenv(ENV_CONSOLIDATION_LLM_TIMEOUT)
+            else None,
             # Embeddings
             embeddings_provider=os.getenv(ENV_EMBEDDINGS_PROVIDER, DEFAULT_EMBEDDINGS_PROVIDER),
             embeddings_local_model=os.getenv(ENV_EMBEDDINGS_LOCAL_MODEL, DEFAULT_EMBEDDINGS_LOCAL_MODEL),
@@ -460,10 +565,6 @@ class HindsightConfig:
                 os.getenv(ENV_RETAIN_EXTRACTION_MODE, DEFAULT_RETAIN_EXTRACTION_MODE)
             ),
             retain_custom_instructions=os.getenv(ENV_RETAIN_CUSTOM_INSTRUCTIONS) or DEFAULT_RETAIN_CUSTOM_INSTRUCTIONS,
-            retain_observations_async=os.getenv(
-                ENV_RETAIN_OBSERVATIONS_ASYNC, str(DEFAULT_RETAIN_OBSERVATIONS_ASYNC)
-            ).lower()
-            == "true",
             # Observations settings (consolidated knowledge from facts)
             enable_observations=os.getenv(ENV_ENABLE_OBSERVATIONS, str(DEFAULT_ENABLE_OBSERVATIONS)).lower() == "true",
             consolidation_batch_size=int(
@@ -484,8 +585,11 @@ class HindsightConfig:
             worker_id=os.getenv(ENV_WORKER_ID) or DEFAULT_WORKER_ID,
             worker_poll_interval_ms=int(os.getenv(ENV_WORKER_POLL_INTERVAL_MS, str(DEFAULT_WORKER_POLL_INTERVAL_MS))),
             worker_max_retries=int(os.getenv(ENV_WORKER_MAX_RETRIES, str(DEFAULT_WORKER_MAX_RETRIES))),
-            worker_batch_size=int(os.getenv(ENV_WORKER_BATCH_SIZE, str(DEFAULT_WORKER_BATCH_SIZE))),
             worker_http_port=int(os.getenv(ENV_WORKER_HTTP_PORT, str(DEFAULT_WORKER_HTTP_PORT))),
+            worker_max_slots=int(os.getenv(ENV_WORKER_MAX_SLOTS, str(DEFAULT_WORKER_MAX_SLOTS))),
+            worker_consolidation_max_slots=int(
+                os.getenv(ENV_WORKER_CONSOLIDATION_MAX_SLOTS, str(DEFAULT_WORKER_CONSOLIDATION_MAX_SLOTS))
+            ),
             # Reflect agent settings
             reflect_max_iterations=int(os.getenv(ENV_REFLECT_MAX_ITERATIONS, str(DEFAULT_REFLECT_MAX_ITERATIONS))),
         )

hindsight_api/daemon.py

@@ -52,7 +52,10 @@ class IdleTimeoutMiddleware:
                 logger.info(f"Idle timeout reached ({self.idle_timeout}s), shutting down daemon")
                 # Give a moment for any in-flight requests
                 await asyncio.sleep(1)
-                os._exit(0)
+                # Send SIGTERM to ourselves to trigger graceful shutdown
+                import signal
+
+                os.kill(os.getpid(), signal.SIGTERM)
 
 
 class DaemonLock:

hindsight_api/engine/consolidation/consolidator.py

@@ -144,10 +144,14 @@ async def run_consolidation_job(
     }
 
     batch_num = 0
+    last_progress_timings = {}  # Track timings at last progress log
     while True:
         batch_num += 1
         batch_start = time.time()
 
+        # Snapshot timings at batch start for per-batch calculation
+        batch_start_timings = perf.timings.copy()
+
         # Fetch next batch of unconsolidated memories
         async with pool.acquire() as conn:
             t0 = time.time()
@@ -217,19 +221,44 @@ async def run_consolidation_job(
             elif action == "skipped":
                 stats["skipped"] += 1
 
-            # Log progress periodically
+            # Log progress periodically with timing breakdown
             if stats["memories_processed"] % 10 == 0:
+                # Calculate timing deltas since last progress log
+                timing_parts = []
+                for key in ["recall", "llm", "embedding", "db_write"]:
+                    if key in perf.timings:
+                        delta = perf.timings[key] - last_progress_timings.get(key, 0)
+                        timing_parts.append(f"{key}={delta:.2f}s")
+
+                timing_str = f" | {', '.join(timing_parts)}" if timing_parts else ""
                 logger.info(
                     f"[CONSOLIDATION] bank={bank_id} progress: "
-                    f"{stats['memories_processed']}/{total_count} memories processed"
+                    f"{stats['memories_processed']}/{total_count} memories processed{timing_str}"
                 )
 
+                # Update last progress snapshot
+                last_progress_timings = perf.timings.copy()
+
         batch_time = time.time() - batch_start
         perf.log(
             f"[2] Batch {batch_num}: {len(memories)} memories in {batch_time:.3f}s "
             f"(avg {batch_time / len(memories):.3f}s/memory)"
         )
 
+        # Log timing breakdown after each batch (delta from batch start)
+        timing_parts = []
+        for key in ["recall", "llm", "embedding", "db_write"]:
+            if key in perf.timings:
+                delta = perf.timings[key] - batch_start_timings.get(key, 0)
+                timing_parts.append(f"{key}={delta:.3f}s")
+
+        if timing_parts:
+            avg_per_memory = batch_time / len(memories) if memories else 0
+            logger.info(
+                f"[CONSOLIDATION] bank={bank_id} batch {batch_num}/{len(memories)} memories: "
+                f"{', '.join(timing_parts)} | avg={avg_per_memory:.3f}s/memory"
+            )
+
     # Build summary
     perf.log(
         f"[3] Results: {stats['memories_processed']} memories -> "
@@ -836,7 +865,14 @@ Focus on DURABLE knowledge that serves this mission, not ephemeral state.
         )
         # Parse JSON response - should be an array
         if isinstance(result, str):
-            result = json.loads(result)
+            # Strip markdown code fences (some models wrap JSON in ```json ... ```)
+            clean = result.strip()
+            if clean.startswith("```"):
+                clean = clean.split("\n", 1)[1] if "\n" in clean else clean[3:]
+                if clean.endswith("```"):
+                    clean = clean[:-3]
+                clean = clean.strip()
+            result = json.loads(clean)
         # Ensure result is a list
         if isinstance(result, list):
             return result