Skip to main content

openhands.memory.condenser.condenser

CONDENSER_METADATA_KEY​

Key identifying where metadata is stored in a State object's extra_data field.

get_condensation_metadata​

def get_condensation_metadata(state: State) -> list[dict[str, Any]]

Utility function to retrieve a list of metadata batches from a State.

Arguments:

  • state - The state to retrieve metadata from.

Returns:

list[dict[str, Any]]: A list of metadata batches, each representing a condensation.

CONDENSER_REGISTRY​

Registry of condenser configurations to their corresponding condenser classes.

Condenser Objects​

class Condenser(ABC)

Abstract condenser interface.

Condensers take a list of Event objects and reduce them into a potentially smaller list.

Agents can use condensers to reduce the amount of events they need to consider when deciding which action to take. To use a condenser, agents can call the condensed_history method on the current State being considered and use the results instead of the full history.

Example usage::

condenser = Condenser.from_config(condenser_config) events = condenser.condensed_history(state)

add_metadata​

def add_metadata(key: str, value: Any) -> None

Add information to the current metadata batch.

Any key/value pairs added to the metadata batch will be recorded in the State at the end of the current condensation.

Arguments:

  • key - The key to store the metadata under.

  • value - The metadata to store.

write_metadata​

def write_metadata(state: State) -> None

Write the current batch of metadata to the State.

Resets the current metadata batch: any metadata added after this call will be stored in a new batch and written to the State at the end of the next condensation.

metadata_batch​

@contextmanager
def metadata_batch(state: State)

Context manager to ensure batched metadata is always written to the State.

condense​

@abstractmethod
def condense(events: list[Event]) -> list[Event]

Condense a sequence of events into a potentially smaller list.

New condenser strategies should override this method to implement their own condensation logic. Call self.add_metadata in the implementation to record any relevant per-condensation diagnostic information.

Arguments:

  • events - A list of events representing the entire history of the agent.

Returns:

  • list[Event] - An event sequence representing a condensed history of the agent.

condensed_history​

def condensed_history(state: State) -> list[Event]

Condense the state's history.

register_config​

@classmethod
def register_config(cls, configuration_type: type[CondenserConfig]) -> None

Register a new condenser configuration type.

Instances of registered configuration types can be passed to from_config to create instances of the corresponding condenser.

Arguments:

  • configuration_type - The type of configuration used to create instances of the condenser.

Raises:

  • ValueError - If the configuration type is already registered.

from_config​

@classmethod
def from_config(cls, config: CondenserConfig) -> Condenser

Create a condenser from a configuration object.

Arguments:

  • config - Configuration for the condenser.

Returns:

  • Condenser - A condenser instance.

Raises:

  • ValueError - If the condenser type is not recognized.

RollingCondenser Objects​

class RollingCondenser(Condenser, ABC)

Base class for a specialized condenser strategy that applies condensation to a rolling history.

The rolling history is computed by appending new events to the most recent condensation. For example, the sequence of calls::

assert state.history == [event1, event2, event3] condensation = condenser.condensed_history(state)

...new events are added to the state...

assert state.history == [event1, event2, event3, event4, event5] condenser.condensed_history(state)

will result in second call to condensed_history passing condensation + [event4, event5] to the condense method.