Token usage metrics for LLM calls. Tracks input/output tokens for a single request to enable per-request cost tracking and monitoring.

Context for a bank write operation validation (pre-operation).

Raised when authentication fails.

Context for a bank read operation validation (pre-operation).

Metadata about a text chunk. Used to track which facts were extracted from which chunks.

Result from an LLM call that may include tool calls.

Context for request authentication and authorization. This dataclass carries authentication data from HTTP requests to the memory engine

A single memory fact returned by search or think operations. This represents a unit of information stored in the memory system, includin

Result from the reflect agent.

An entity extracted from text.

An entity extracted from text.

A tool call requested by the LLM.

Result of analyzing a natural language query. Contains extracted structured information like temporal constraints.

Raised when an extension fails to load.

Advanced memory system using temporal and semantic linking with PostgreSQL. This class provides: - Embedding generation for semantic sea

Response model for bank profile.

Response model for delete operations.

Response model for a directive.

Disposition traits that influence how memories are formed and interpreted.

Link between two memory units through a shared entity. Used for entity-based graph connections in the memory graph.

Timing breakdown for a single MPFP retrieval call.

Temporal constraint extracted from a query. Represents a time range with start and end dates.

Tenant context returned by authentication. Contains the PostgreSQL schema name for tenant isolation. All database queries will use fully

HTTP delivery configuration for a webhook.

Response model for bank configuration.

Disposition traits for a memory bank. All traits are scored 1-5 where: - skepticism: 1=trusting, 5=skeptical (how much to doubt or quest

Manages an embedded PostgreSQL server instance using pg0-embedded.

Response model for a mental model (stored reflect response).

Result from a single pattern traversal.

Result context for post-recall hook. Contains the operation parameters and the result.

Represents a tenant for worker discovery. Used by list_tenants() to return tenant information including the PostgreSQL schema name for d

Response model for a webhook.

Type for bank profile data.

Causal relationship from this fact to a previous fact (stored format).

Causal relationship from this fact to a previous fact (stored format).

Information about a chunk.

Query analyzer using dateparser library. Uses dateparser to extract temporal expressions from natural language queries. Supports 200+ la

Default implementation of ExtensionContext. Uses the system's database URL and migration infrastructure.

Default single-tenant extension with no authentication. This is the default extension used when no tenant extension is configured. It pr

A neighbor node with its edge weight.

Entity labels configuration for a bank (controlled vocabulary).

An observation about an entity.

Current mental model of an entity. Contains observations synthesized from facts about the entity.

Final fact model for storage - built from lenient parsing of LLM response. This is what fact_extraction returns and what the rest of the pip

Final fact model for storage - built from lenient parsing of LLM response. This is what fact_extraction returns and what the rest of the pip

Configuration container for Hindsight API.

Graph retrieval via direct link expansion from seeds. Runs three expansions through precomputed memory_links: entity co-occurrence, sema

No-op span that provides the same interface as OpenTelemetry Span but does nothing.

Raised when an operation fails validation.

State for a pattern traversal between hops.

Input content item to be retained as memories. Represents a single piece of content to extract facts from.

Context for a retain operation validation (pre-operation). Contains ALL user-provided parameters for the retain operation.

Raise from a task handler to schedule a retry at a specific time.

An entry point node with its initial score.

Total token usage across all LLM calls.

Polls PostgreSQL for pending tasks and executes them. Uses FOR UPDATE SKIP LOCKED for safe distributed claiming, allowing multiple worke

Stat accumulation entry for a resolved entity (post-transaction update).

An entity that needs to be inserted (no matching candidate found).

LLM provider using Anthropic's Claude models. Supports structured output, tool calling, and extended thinking mode. Handles format conve

Response model for submitting an async operation.

Azure Blob Storage backend. Uses obstore (Rust-backed) for high-throughput async access to Azure Blob Storage. Supports account key, SAS

Graph retrieval using BFS-style spreading activation. Starting from semantic entry points, spreads activation through the memory graph (

Response model for background update. Deprecated: use MissionResponse instead.

Context for filtering the bank list (post-query).

Response model for listing all banks.

Result of filtering the bank list.

Response model for bank statistics endpoint.

Namespace for bank-related operations.

Namespace for bank-related operations.

Metadata for child batch_retain operations (individual sub-batches).

Metadata for parent batch_retain operations (when split into sub-batches).

Task backend using PostgreSQL as broker. submit_task() stores task_payload in async_operations table. Actual polling and execution is ha

A single operation with connection budget. Provides methods to acquire connections within the budget.

A pool wrapper that limits concurrent connection acquisitions. This can be passed to functions expecting a pool, and acquire() calls wil

Response model for cancel operation endpoint.

Chunk data for a single chunk.

A task claimed from the database with its schema context.

LLM provider using Claude Code authentication. Authenticates using Claude Pro/Max credentials via `claude auth login` and makes API call

Response model for clearing observations for a specific memory.

LLM provider using OpenAI Codex OAuth authentication. Authenticates using ChatGPT Plus/Pro credentials stored in ~/.codex/auth.json and

Cohere cross-encoder implementation using the Cohere Rerank API. Supports rerank-english-v3.0 and rerank-multilingual-v3.0 models.

Cohere embeddings implementation using the Cohere API. Supports embed-english-v3.0 (1024 dims) and embed-multilingual-v3.0 (1024 dims).

Raised when trying to access a bank-configurable field from global config.

Resolves hierarchical configuration with tenant/bank overrides.

Thin wrapper around LLMProvider that applies bank-specific config to every call. Obtained via ``LLMProvider.with_config(resolved_config)``.

Manages per-operation connection budgets. Usage: manager = ConnectionBudgetManager(default_budget=4) # Start an operation

Performance logging for consolidation operations.

Performance logging for consolidation operations.

Response model for consolidation trigger endpoint.

Result of a successful file conversion.

Response model for mental model creation.

Neural reranking using a cross-encoder model. Configured via environment variables (see cross_encoder.py). Default local model is cross-

Response model for delete document endpoint.

Information about a directive that was applied during reflect.