strands.memory.extraction.types

Primitive types for the memory extraction subsystem.

ExtractionResult

@dataclass
class ExtractionResult()

Defined in: src/strands/memory/extraction/types.py:37

A discrete entry produced by an :class:Extractor, ready to write via add.

ExtractorContext

@dataclass
class ExtractorContext()

Defined in: src/strands/memory/extraction/types.py:45

Context passed to :meth:Extractor.extract.

Attributes:

• default_model - The agent’s model, supplied so an extractor can default to it.

Extractor

class Extractor(Protocol)

Defined in: src/strands/memory/extraction/types.py:56

Transforms conversation messages into discrete, searchable entries.

Optional on a store’s :class:ExtractionConfig: when absent, the manager passes messages straight to the store’s add_messages (the no-extractor passthrough), which is the right path for backends that extract server-side.

extract

async def extract(
        messages: list[Message],
        context: ExtractorContext | None = None) -> list[ExtractionResult]

Defined in: src/strands/memory/extraction/types.py:64

Extract entries from a batch of messages.

MemoryMessageFilter

@dataclass
class MemoryMessageFilter()

Defined in: src/strands/memory/extraction/types.py:70

Filters content blocks out of messages before extraction.

Blocks whose kind is in :attr:exclude are stripped; a message left with no content is dropped. Defaults to excluding tool traffic (toolUse / toolResult).

ExtractionTriggerContext

@dataclass
class ExtractionTriggerContext()

Defined in: src/strands/memory/extraction/types.py:86

Context handed to :meth:ExtractionTrigger.attach.

Attributes:

• agent - The agent the trigger attaches its hooks to.
• fire - Save this store’s unsaved messages now. Runs in the background and returns immediately. To await completion, see MemoryManager.flush.

ExtractionTrigger

class ExtractionTrigger(ABC)

Defined in: src/strands/memory/extraction/types.py:99

Controls when a store’s :class:ExtractionConfig runs.

A trigger is a self-attaching value object: :meth:attach wires the agent hooks it needs and calls :attr:ExtractionTriggerContext.fire when extraction should happen. Subclass for custom triggering logic. A trigger that never fires never extracts; for a guaranteed final write, use MemoryManager.flush.

Attributes:

• name - Stable identifier for this trigger kind, used in logging.

attach

@abstractmethod
def attach(context: ExtractionTriggerContext) -> None

Defined in: src/strands/memory/extraction/types.py:115

Wire this trigger into the agent lifecycle.

Called once per store during MemoryManager initialization. Register hooks on context.agent and call context.fire() when extraction should run.

ExtractionConfig

@dataclass
class ExtractionConfig()

Defined in: src/strands/memory/extraction/types.py:126

Per-store automatic-extraction configuration.

Attributes:

• trigger - When to run extraction. A single trigger or a non-empty list; multiple triggers compose (extraction runs whenever any fires).
• extractor - How to turn messages into entries. When set, the store must implement add. When omitted, the manager hands the filtered messages straight to the store’s add_messages (for backends that extract server-side).
• filter - Content blocks to strip before extraction. Defaults to :data:DEFAULT_MEMORY_MESSAGE_FILTER (excludes toolUse / toolResult). Pass MemoryMessageFilter(exclude=[]) to keep tool blocks.