🔗 API Reference#
Memory Models#
class BaseMemory#
Description: Basic class for memory methods.
__init__(config)#
Description: Constructor for BaseMemory, initializes the configuration.
Parameters:
config(object): Configuration for the memory methods.
Return: None
__reset_objects__(objects)#
Description: Resets the provided memory-related objects.
Parameters:
objects(list): List of objects to reset.
Return: None
reset()#
Description: Resets the state of the memory.
Parameters: None
Return: None
store(obj)#
Description: Stores an object in the memory.
Parameters:
obj(object): The object to store.
Return: None
recall(obj)#
Description: Recalls a stored object from the memory.
Parameters:
obj(object): The object or query to be recalled.
Return: object
manage(operation, kwargs)#
Description: Manages memory operations with specified operations and parameters.
Parameters:
operation(str): The operation to manage the memory.kwargs(dict): Additional keyword arguments for the operation.
Return: None
optimize(kwargs)#
Description: Optimizes memory usage based on parameters.
Parameters:
kwargs(dict): Additional keyword arguments for optimization.
Return: None
class ExplicitMemory#
Description: Explicit memory indicates the methods that represent memory contents with texts.
__init__(config)#
Description: Constructor for ExplicitMemory, initializes the configuration.
Parameters:
config(object): Configuration for explicit memory methods.
Return: None
reset()#
Description: Resets the state of explicit memory.
Parameters: None
Return: None
store(observation)#
Description: Stores an observation in the explicit memory.
Parameters:
observation(str): The observation to store.
Return: None
recall(query)#
Description: Recalls a stored observation from the explicit memory.
Parameters:
query(str): The query to recall related observation.
Return: str
display()#
Description: Displays the contents of the explicit memory.
Parameters: None
Return: None
manage(operation, kwargs)#
Description: Manages explicit memory operations with specified operations and parameters.
Parameters:
operation(str): The operation to manage the explicit memory.kwargs(dict): Additional keyword arguments for the operation.
Return: None
optimize(kwargs)#
Description: Optimizes explicit memory usage based on parameters.
Parameters:
kwargs(dict): Additional keyword arguments for optimization.
Return: None
class ImplicitMemory#
Description: Implicit memory indicates the methods that represent memory contents with parameters.
class FUMemory#
Description: FUMemory (Full Memory): Concatenate all information into one string, implementing long-context memory.
class GAMemory#
Description: GAMemory (Generative Agents): A pioneer memory model with weighted retrieval combination and self-reflection mechanism.
class LTMemory#
Description: LTMemory (Long-term Memory): Calculate semantic similarities with text embeddings to retrieve the most relevant information.
class MBMemory#
Description: MBMemory (MemoryBank): A multi-layered memory model with dynamic summarization and forgetting mechanism. It is an enhanced memory model for large language models, treating one step as one day.
class MGMemory#
Description: MGMemory (MemGPT): A hierarchical memory model that treats the memory system as an operating system. Based on the work by Packer et al., it is designed to manage and store information in a structured manner using multiple storage strategies.
class MTMemory#
Description: MTMemory (MemTree): A dynamic memory model with a tree-structured semantic representation to organize information.
class RFMemory#
Description: RFMemory (Reflexion): A famous memory method that can learn to memorize from previous trajectories by optimization.
class SCMemory#
Description: SCMemory (SCM): A self-controlled memory model that can recall minimum but necessary information for inference based on the framework described by Wang, Bing, et al.
class STMemory#
Description: STMemory (Short-term Memory) is a class that maintains the most recent information and concatenates them into a single string as context.
Memory Operations#
class BaseOptimize#
Description: An abstract base class for optimization strategies, which provides a template for resetting and calling optimization procedures.
__init__(config)#
Description: Constructor method for initializing the base class with the given configuration.
Parameters:
config(object): The configuration object containing settings for optimization.
Return: None
__reset_objects__(objects)#
Description: Resets the state of provided objects by calling their reset method.
Parameters:
objects(list): List of objects to be reset.
Return: None
reset()#
Description: Abstract method to reset the optimization process, to be implemented by subclasses.
Parameters: None
Return: None
__call__(kwargs)#
Description: Abstract method to execute the optimization process, to be implemented by subclasses.
Parameters:
kwargs(dict): Additional keyword arguments for the optimization call.
Return: None
class RFOptimize#
Description: A concrete implementation of the BaseOptimize class in Reflexion.
__init__(config, kwargs)#
Description: Constructor method for initializing the RFOptimize class with a configuration and additional keyword arguments.
Parameters:
config(object): The configuration object containing settings for the Reflector.kwargs(dict): Additional keyword arguments, including ‘insight’ dictionary.
Return: None
reset()#
Description: Resets the RFOptimize object state by resetting associated components like the reflector.
Parameters: None
Return: None
__call__(kwargs)#
Description: Executes the optimization procedure by generating new insights based on the current trial and updating the insight dictionary.
Parameters:
kwargs(dict): Additional keyword arguments, expected to include ‘new_trial’.
Return: None
class BaseRecall#
Description: An abstract base class for memory recall systems.
__init__(config)#
Description: Initializes the BaseRecall with a configuration.
Parameters:
config(object): Configuration object for the recall system.
Return: None
reset()#
Description: Abstract method to reset the state of the recall system.
Parameters: None
Return: None
__call__(query)#
Description: Abstract method for handling queries.
Parameters:
query(object): A query object to be processed.
Return: object
__reset_objects__(objects)#
Description: Resets a list of objects by calling their reset method.
Parameters:
objects(list): List of objects to reset.
Return: None
class FUMemoryRecall#
Description: Recall operation for FUMemory.
__init__(config, kwargs)#
Description: Initializes the FUMemoryRecall with configuration and additional arguments.
Parameters:
config(object): Configuration object for the recall system.kwargs(dict): Additional arguments including storage.
Return: None
reset()#
Description: Resets the truncation and utilization components.
Parameters: None
Return: None
__call__(query)#
Description: Fetches memory context from storage considering utilization and truncation.
Parameters:
query(object): A query to process.
Return: string
class STMemoryRecall#
Description: A class for short-term memory recall involving time-based retrieval.
__init__(config, kwargs)#
Description: Initializes the STMemoryRecall with configuration and additional arguments.
Parameters:
config(object): Configuration object for the recall system.kwargs(dict): Additional arguments including storage.
Return: None
reset()#
Description: Resets the components involved in truncation, utilization, and time retrieval.
Parameters: None
Return: None
__call__(query)#
Description: Processes the query to fetch recent memory using time-based ranking.
Parameters:
query(object): A query to process.
Return: string
class LTMemoryRecall#
Description: A class for long-term memory recall using text-based retrieval.
__init__(config, kwargs)#
Description: Initializes the LTMemoryRecall with configuration and additional arguments.
Parameters:
config(object): Configuration object for the recall system.kwargs(dict): Additional arguments including storage.
Return: None
reset()#
Description: Resets the components involved in truncation, utilization, and text retrieval.
Parameters: None
Return: None
__call__(query)#
Description: Retrieves long-term memory information based on text relevance.
Parameters:
query(object): A query to process.
Return: string
class GAMemoryRecall#
Description: Recall operation for GAMemory.
__init__(config, kwargs)#
Description: Initializes the GAMemoryRecall with specified configuration.
Parameters:
config(object): Configuration object for the recall system.kwargs(dict): Additional keyword arguments including storage options.
Return: None
reset()#
Description: Resets all involved memory retrieval and processing components.
Parameters: None
Return: None
__retention__(target_ids, timestamp)#
Description: Updates memory recency for specific memory IDs.
Parameters:
target_ids(list): List of memory IDs to update.timestamp(int/float): Current timestamp for recency update.
Return: None
__call__(query)#
Description: Processes the query to retrieve context using multiple retrieval criteria.
Parameters:
query(object): A query to be processed.
Return: string
class MBMemoryRecall#
Description: Recall operation for MBMemory.
__init__(config, kwargs)#
Description: Initializes the MBMemoryRecall with specified configuration and options.
Parameters:
config(object): Configuration object for the memory recall.kwargs(dict): Keyword arguments including storage and summary details.
Return: None
__retention__(mid, timestamp)#
Description: Updates recency of a given memory ID with current timestamp.
Parameters:
mid(int): Memory ID whose recency will be updated.timestamp(int/float): Current timestamp.
Return: None
reset()#
Description: Resets components and initializes the global summary.
Parameters: None
Return: None
__call__(query)#
Description: Retrieves memory context according to specified configurations and processes the query.
Parameters:
query(object): A query input to be processed for memory recall.
Return: string
class SCMemoryRecall#
Description: Recall operation for SCMemory.
__init__(config, kwargs)#
Description: Initializes the SCMemoryRecall with specified configuration.
Parameters:
config(object): Configuration object for the memory recall system.kwargs(dict): Keyword arguments including storage options and retrieval methods.
Return: None
reset()#
Description: Resets structured memory recall components.
Parameters: None
Return: None
__get_flash_memory__(timestamp)#
Description: Fetches the most recent memory items.
Parameters:
timestamp(int/float): Current timestamp used for recency adjustment.
Return: list
__check_require_activation__(text, flash_memory_list)#
Description: Determines if additional retrieval is required based on the given text and flash memory.
Parameters:
text(string): Query or input text.flash_memory_list(list): List of recent memory items.
Return: bool
__check_require_summary__(text, activation_summary_list, flash_memory_list)#
Description: Checks if summarized information suffices for the given query.
Parameters:
text(string): Query or observation text.activation_summary_list(list): List of retrieved information summaries.flash_memory_list(list): List of recent memory items.
Return: bool
__retention__(target_ids, timestamp)#
Description: Updates the recency of specific memory elements.
Parameters:
target_ids(list): List of memory IDs to be updated.timestamp(int/float): Timestamp used for the update.
Return: None
__call__(query)#
Description: Handles the query to retrieve and process the appropriate memory context.
Parameters:
query(object): A query that needs processing.
Return: string
class MGMemoryRecall#
Description: Recall operation for MGMemory.
__init__(config, kwargs)#
Description: Initializes the MGMemoryRecall with specified configuration and context data.
Parameters:
config(object): Configuration object for the memory systems.kwargs(dict): Additional arguments including context and storage options.
Return: None
reset()#
Description: Resets the memory system components.
Parameters: None
Return: None
__get_current_memory_prompt()#
Description: Generates a formatted prompt showing current memory state.
Parameters: None
Return: string
get_current_memory_context()#
Description: Composes the current memory context from the working and FIFO memory.
Parameters: None
Return: string
__trigger_function_call__(text)#
Description: Triggers function calls based on the provided text.
Parameters:
text(string): The input query or observation text.
Return: None
__memory_retrieval__(query)#
Description: Retrieves related information from archival storage and updates working memory.
Parameters:
query(string): The query string for information retrieval.
Return: None
__memory_recall__(query)#
Description: Retrieves related memory from recall storage and updates FIFO memory.
Parameters:
query(string): The query string for recalling memory.
Return: None
__memory_archive__(memory_list)#
Description: Stores specified FIFO memory items in archival storage.
Parameters:
memory_list(list): List of FIFO memory indices to archive.
Return: None
__memory_transfer__(memory_list)#
Description: Transfers specified FIFO memory entries to working memory.
Parameters:
memory_list(list): Indexes of FIFO memory to transfer.
Return: None
__memory_save__(memory_list)#
Description: Stores specified working memory items in archival storage.
Parameters:
memory_list(list): List of working memory indices to save.
Return: None
__call__(query)#
Description: Processes a query to manage and utilize current memory context.
Parameters:
query(object): The input query for memory processing.
Return: string
class RFMemoryRecall#
Description: Recall operation for RFMemory.
__init__(config, kwargs)#
Description: Initializes the RFMemoryRecall with key components and configuration.
Parameters:
config(object): The configuration object for recall functions.kwargs(dict): Dictionary of additional settings and storage.
Return: None
reset()#
Description: Resets components like truncation and utilization.
Parameters: None
Return: None
__call__(query)#
Description: Combines insights and memory to process the given query.
Parameters:
query(object): Query object or string to process.
Return: string
class BaseReflect#
Description: Base class for reflection, serving as an interface for derived classes.
__init__(config)#
Description: Initializes the BaseReflect class with a given configuration.
Parameters:
config(object): Configuration object containing various setup parameters.
Return: None
__reset_objects__(objects)#
Description: Resets the provided objects by invoking their reset method.
Parameters:
objects(list): List of objects that have a reset method to be called.
Return: None
reset()#
Description: Abstract method that must be implemented by derived classes to reset specific functionalities.
Parameters: None
Return: None
__call__(**kwargs)#
Description: Abstract method that must be implemented by derived classes to define specific calling behavior.
Parameters:
**kwargs(dict): Additional keyword arguments for flexibility in derived class implementations.
Return: None
class GAReflect#
Description: Derived class from BaseReflect, implementing reflection in Generative Agents.
__init__(config, **kwargs)#
Description: Initializes the GAReflect class with the given configuration and additional components such as storage and retrieval methods.
Parameters:
config(object): Configuration object containing setup parameters specific to GAReflector.**kwargs(dict): Additional components such as ‘storage’, ‘text_retrieval’, and ‘time_retrieval’.
Return: None
reset()#
Description: Resets the internal state of the GAReflector object.
Parameters: None
Return: None
__get_recursion_context__(mid)#
Description: Recursively builds a context string from memory identified by ‘mid’, aggregating related memory text.
Parameters:
mid(int): Memory identifier from which to collect and build the recursion context.
Return: Aggregated text from the recursion list.
__call__()#
Description: Executes the reflection process, generating insights when accumulated importance exceeds a threshold.
Parameters: None
Return: A list of generated insights, each with text, time, and source data.
class BaseStore#
Description: Base abstract class for different types of memory stores, providing the basic structure and abstract methods that need to be implemented by subclasses.
__init__(config)#
Description: Constructor that initializes the BaseStore with a configuration.
Parameters:
config(object): Configuration object for the memory store.
Return: None
__reset_objects__(objects)#
Description: Resets the provided objects by calling their reset method.
Parameters:
objects(list): List of objects to reset.
Return: None
reset()#
Description: Abstract method to reset the memory store.
Parameters: None
Return: None
__call__(observation)#
Description: Abstract method to handle observations.
Parameters:
observation(dict): Data to be processed by the memory store.
Return: None
class FUMemoryStore#
Description: Store operation for FUMemory.
__init__(config, kwargs)#
Description: Constructor that initializes the FUMemoryStore with a configuration and storage.
Parameters:
config(object): Configuration object for the memory store.kwargs(dict): Additional keyword arguments for initialization, including storage.
Return: None
reset()#
Description: Reset method for FUMemoryStore.
Parameters: None
Return: None
__call__(observation)#
Description: Adds an observation to the storage.
Parameters:
observation(dict or str): Data to be added to the store, can be a dictionary or string.
Return: None
class STMemoryStore#
Description: A memory store class that handles observations based on time, storing them along with a timestamp.
__init__(config, kwargs)#
Description: Constructor that initializes STMemoryStore with a configuration, storage, and time retrieval system.
Parameters:
config(object): Configuration object for the memory store.kwargs(dict): Additional keyword arguments for initialization, including storage and time retrieval.
Return: None
reset()#
Description: Reset method for STMemoryStore.
Parameters: None
Return: None
__call__(observation)#
Description: Stores an observation with a timestamp.
Parameters:
observation(dict or str): Data to be added to the store, containing a ‘time’ or a ‘text’ key.
Return: None
class LTMemoryStore#
Description: A memory store class that stores text observations for long-term retrieval.
__init__(config, kwargs)#
Description: Constructor that initializes LTMemoryStore with a configuration, storage, and text retrieval system.
Parameters:
config(object): Configuration object for the memory store.kwargs(dict): Additional keyword arguments for initialization, including storage and text retrieval.
Return: None
reset()#
Description: Reset method for LTMemoryStore.
Parameters: None
Return: None
__call__(observation)#
Description: Stores an observation and adds its text to the text retrieval system.
Parameters:
observation(dict or str): Data to be added to the store, containing a ‘text’ key.
Return: None
class GAMemoryStore#
Description: Store operation for GAMemory.
__init__(config, kwargs)#
Description: Constructor that initializes GAMemoryStore with configuration and various components for storage, retrieval, importance judgement, and reflection.
Parameters:
config(object): Configuration object for the memory store.kwargs(dict): Additional keyword arguments for initialization, including storage, text and time retrieval, importance judgement, and reflector.
Return: None
reset()#
Description: Reset method for GAMemoryStore.
Parameters: None
Return: None
__call__(observation)#
Description: Processes an observation by judging its importance, storing it, and updating based on its importance score.
Parameters:
observation(dict or str): Data to be added to the store, containing a ‘text’ key and optional ‘time’ and ‘source’ keys.
Return: None
class MBMemoryStore#
Description: Store operation for MBMemory.
__init__(config, kwargs)#
Description: Constructor that initializes the MBMemoryStore with configuration and components for storage, summarization, and text retrieval.
Parameters:
config(object): Configuration object for the memory store.kwargs(dict): Additional keyword arguments for initialization, including storage, summarization tools, and text retrieval.
Return: None
reset()#
Description: Reset method for MBMemoryStore.
Parameters: None
Return: None
__summarize_history__(time)#
Description: Summarizes information within a specified time block.
Parameters:
time(int or float): The time identifier for the block to summarize.
Return: None
__summarize_summary__()#
Description: Updates the global summary with local summaries.
Parameters: None
Return: None
__call__(observation)#
Description: Handles observations, stores them, and checks for summarization.
Parameters:
observation(dict or str): Data to be added to the store, if not provided with ‘time’, a new time is generated.
Return: None
class SCMemoryStore#
Description: Store operation for SCMemory.
__init__(config, kwargs)#
Description: Constructor that initializes SCMemoryStore with a configuration and components including storage, text retrieval, and summarization.
Parameters:
config(object): Configuration object for the memory store.kwargs(dict): Additional keyword arguments for initialization, including storage, text retrieval, and summarization components.
Return: None
reset()#
Description: Reset method for SCMemoryStore.
Parameters: None
Return: None
__call__(observation)#
Description: Processes and stores observations, creating a summary if one is not provided.
Parameters:
observation(dict or str): Data to be added to the store, containing a ‘text’ key and optionally a ‘time’.
Return: None
class MGMemoryStore#
Description: Store operation for MGMemory.
__init__(config, kwargs)#
Description: Constructor for MGMemoryStore, initializing with configuration and components for context management and summarization.
Parameters:
config(object): Configuration object for the memory store.kwargs(dict): Additional keyword arguments for initialization, including main context, recall storage, retrieval, and flushing components.
Return: None
reset()#
Description: Reset method for MGMemoryStore, resets the summarizer and flush checker.
Parameters: None
Return: None
__flush_queue__()#
Description: Flushes the FIFO queue and updates the recall storage and main context’s summary.
Parameters: None
Return: None
__call__(observation)#
Description: Processes observations, adding them to the FIFO queue and checking if a flush is needed.
Parameters:
observation(dict or str): Observation to be processed and stored in the queue.
Return: None
class MTMemoryStore#
Description: Store operation for MTMemory.
__init__(config, kwargs)#
Description: Constructor to initialize MTMemoryStore with configuration and components for storage, retrieval, and summarization.
Parameters:
config(object): Configuration object for the memory store.kwargs(dict): Additional keyword arguments for initialization, including storage, text retrieval, and summarizer.
Return: None
reset()#
Description: Reset method for MTMemoryStore, resets the summarizer.
Parameters: None
Return: None
__get_traverse_threshold__(d)#
Description: Calculates the threshold for node traversal based on depth.
Parameters:
d(int): Current depth in the tree.
Return: float
__recursive_insert_node__(new_node_id, text_embedding)#
Description: Recursively inserts a new node into the tree based on traversal criteria.
Parameters:
new_node_id(int): ID of the new node to be inserted.text_embedding(torch.Tensor): Embedding of the text for the new node.
Return: None
__call__(observation)#
Description: Processes observations, adding them as nodes in the tree and performing recursive inserts.
Parameters:
observation(dict or str): Data to be processed and added in tree structure.
Return: None
Memory Functions#
class BaseEncoder#
Description: Transfer textual messages into embeddings to represent in latent space by pre-trained models.
__init__(config)#
Description: Initializes the BaseEncoder with the given configuration.
Parameters:
config(object): The configuration used to initialize the encoder.
Return: None
reset()#
Description: Resets the encoder state if needed.
Parameters: None
Return: None
__call__(*args, **kwargs)#
Description: Abstract method to be implemented by subclasses for encoding text.
Parameters:
*args(list): Variable length argument list.**kwargs(dict): Arbitrary keyword arguments.
Return: Abstract
class LMEncoder#
Description: Embedding via LM transformers.
__init__(config)#
Description: Initializes the LMEncoder with the given configuration, tokenizer, and model.
Parameters:
config(object): The configuration used to initialize the LMEncoder.
Return: None
__call__(text, return_type)#
Description: Encodes the given text into embeddings using a language model.
Parameters:
text(str): The input text to encode.return_type(str): The type of data to return: ‘numpy’ for numpy array or ‘tensor’ for tensor.
Return: numpy.ndarray or torch.Tensor depending on the return_type.
class STEncoder#
Description: Embedding via Sentence Transformer.
__init__(config)#
Description: Initializes the STEncoder with the given configuration and model.
Parameters:
config(object): The configuration used to initialize the STEncoder.
Return: None
__call__(text, return_type)#
Description: Encodes the given text into embeddings using a Sentence Transformer model.
Parameters:
text(str): The input text to encode.return_type(str): The type of data to return: ‘numpy’ for numpy array or ‘tensor’ for tensor.
Return: numpy.ndarray or torch.Tensor depending on the return_type.
class BaseForget#
Description: Base abstract class for forget mechanisms used in simulation-oriented agents, particularly in role-playing and social simulations. It models the human cognitive psychology of forgetting.
__init__(config)#
Description: Initializes the BaseForget class with a configuration.
Parameters:
config(object): Configuration object for initializing the forget mechanism.
Return: None
reset()#
Description: Resets the forget mechanism to its initial state. This method can be overridden by subclasses if needed.
Parameters: None
Return: None
get_forget_prob(*args, **kwargs)#
Description: Abstract method to calculate the probability of forgetting. This method must be implemented by subclasses.
Parameters:
*args(tuple): Positional arguments required by the specific forgetting algorithm.**kwargs(dict): Keyword arguments required by the specific forgetting algorithm.
Return: float
class MBForget#
Description: Derived class implementing a forgetting mechanism according to an exponential curve.
__init__(config)#
Description: Initializes the MBForget class with a configuration by calling the superclass initializer.
Parameters:
config(object): Configuration object for initializing the forget mechanism.
Return: None
get_forget_prob(current_time, recency, strength)#
Description: Calculates the probability of forgetting using an exponential decay function based on the difference between current time and recency, modulated by the strength parameter.
Parameters:
current_time(float): The current time in the simulation.recency(float): The recency of the memory or information.strength(float): The strength of the memory or information.
Return: float
sample_forget(current_time, recency, strength)#
Description: Samples whether forgetting occurs based on a random draw and the calculated forgetting probability.
Parameters:
current_time(float): The current time in the simulation.recency(float): The recency of the memory or information.strength(float): The strength of the memory or information.
Return: bool
class BaseJudge#
Description: Assess given observations or intermediate messages on certain aspects.
__init__(config)#
Description: Initializes an instance of the BaseJudge class.
Parameters:
config(object): Configuration object for setting up the judge.
Return: None
reset()#
Description: Resets the state of the judge.
Parameters: None
Return: None
__call__()#
Description: Abstract method to be implemented by subclasses that performs the evaluation.
Parameters: None
Return: Abstract method with no implementation.
class LLMJudge#
Description: Judge via large language models.
__init__(config)#
Description: Initializes an instance of the LLMJudge class using a provided configuration.
Parameters:
config(object): Configuration object which includes LLM setup details.
Return: None
__post_scale__(res)#
Description: Processes the result from the LLM by scaling it.
Parameters:
res(string): Result obtained from the LLM which is expected to be evaluable to a float.
Return: Float resulting from scaling the evaluated result.
__post_bool__(res)#
Description: Processes the result from the LLM by converting it to a boolean.
Parameters:
res(string): Result obtained from the LLM which should be either ‘True’ or ‘False’.
Return: Boolean value derived from the result string or an error message if parsing fails.
__call__(input_dict, post_process)#
Description: Evaluates an input using the LLM and processes the output according to the specified post-processing method.
Parameters:
input_dict(dict): Dictionary containing the inputs required by the prompt template.post_process(string): Specifies the post-processing method to apply: ‘scale’ or ‘bool’.
Return: Processed result scaled to a float or converted to a boolean depending on the post-processing method.
class BaseLLM#
Description: Provides a convenient interface to utilize the powerful capability of different large language models.
__init__(config)#
Description: Initializes the BaseLLM instance with given configuration.
Parameters:
config(object): Configuration object for LLM.
Return: void
reset()#
Description: Resets the state of the language model interface.
Parameters: None
Return: void
fast_run(query)#
Description: Abstract method to be overridden for quickly running a query using the LLM.
Parameters:
query(str): The query to be processed by the LLM.
Return: str
class APILLM#
Description: Utilize LLM from APIs.
__init__(config)#
Description: Initializes the APILLM instance with given configuration and sets up the OpenAI client.
Parameters:
config(object): Configuration object containing API key and other settings.
Return: void
parse_response(response)#
Description: Parses the response from the API call and extracts relevant information.
Parameters:
response(object): Response object from the API call.
Return: dict
run(message_list)#
Description: Executes a conversation with the LLM API using a list of messages.
Parameters:
message_list(list): List of message dicts containing role and content.
Return: dict
fast_run(query)#
Description: Runs a single query quickly by wrapping it in user message format.
Parameters:
query(str): The query to be processed by the LLM.
Return: str
class LocalVLLM#
Description: Uses a local variant of LLM for processing.
__init__(config)#
Description: Initializes the LocalVLLM instance with model and sampling parameters.
Parameters:
config(object): Configuration object containing model name and temperature settings.
Return: void
run(message_list)#
Description: Executes a conversation with the local LLM using a list of messages.
Parameters:
message_list(list): List of message dicts containing role and content.
Return: object
fast_run(query)#
Description: Runs a single query quickly by wrapping it in user message format.
Parameters:
query(str): The query to be processed by the LLM.
Return: str
class BaseReflector#
Description: Draw new insights in higher level from existing information, commonly for reflection and optimization operations.
__init__(config)#
Description: Initialize the BaseReflector with a given configuration.
Parameters:
config(object): Configuration object for initializing the reflector.
Return: None
reset()#
Description: Reset the state of the reflector.
Parameters: None
Return: None
generate_insight(*args, **kwargs)#
Description: Abstract method for generating insights.
Parameters:
*args(tuple): Variable length argument list.**kwargs(dict): Arbitrary keyword arguments.
Return: None
class GAReflector#
Description: Implementation of Reflection in Generative Agents.
__init__(config)#
Description: Initialize the GAReflector with a given configuration and setup internal state.
Parameters:
config(object): Configuration object specific to GAReflector.
Return: None
reset()#
Description: Reset the state of the GAReflector, including accumulated importance and recursion list.
Parameters: None
Return: None
add_reflection(importance, recursion)#
Description: Add a new reflection.
Parameters:
importance(int): Numeric value indicating the importance of the reflection.recursion(object): Recursion object to be added to the list.
Return: None
delete_reflection(index)#
Description: Delete a reflection by its index. Not implemented.
Parameters:
index(int): Index of the reflection to be deleted.
Return: None
get_recursion(index)#
Description: Retrieve a recursion by its index.
Parameters:
index(int): Index of the recursion to retrieve.
Return: object
get_current_accmulated_importance()#
Description: Get the current accumulated importance.
Parameters: None
Return: int
get_reflection_threshold()#
Description: Get the reflection threshold from configuration.
Parameters: None
Return: int
self_ask(input_dict)#
Description: Generate a list of questions from input_dict using the LLM and configured prompt.
Parameters:
input_dict(dict): Input dictionary for formatting the prompt.
Return: list
generate_insight(input_dict)#
Description: Generate insights from input_dict using the LLM and configured prompt.
Parameters:
input_dict(dict): Input dictionary for formatting the prompt.
Return: list
class TrialReflector#
Description: A simplistic implementation of BaseReflector to generate reflections and insights.
__init__(config)#
Description: Initialize the TrialReflector with a given configuration.
Parameters:
config(object): Configuration object specific to TrialReflector.
Return: None
generate_insight(input_dict)#
Description: Generate insights from input_dict using the LLM and a specified prompt.
Parameters:
input_dict(dict): Input dictionary for formatting the prompt.
Return: str
class BaseRetrieval#
Description: Utilized to find most useful information for the current query or observation. Commonly by different aspects like semantic relevance, importance, recency and so on.
__init__(config)#
Description: Initializes the base retrieval class with a specified configuration.
Parameters:
config(dict): Configuration settings for retrieval.
Return: None
reset()#
Description: Resets the retrieval system to its initial state.
Parameters: None
Return: None
__call__()#
Description: Abstract method to be implemented by subclasses for executing retrieval with given parameters.
Parameters: None
Return: Undefined in abstract class.
add()#
Description: Abstract method to be implemented by subclasses for adding new data to the retrieval system.
Parameters: None
Return: Undefined in abstract class.
update()#
Description: Abstract method to be implemented by subclasses for updating existing data in the retrieval system.
Parameters: None
Return: Undefined in abstract class.
delete()#
Description: Abstract method to be implemented by subclasses for deleting data from the retrieval system.
Parameters: None
Return: Undefined in abstract class.
class LinearRetrieval#
Description: Retrieval in linear indexes.
__init__(config)#
Description: Initializes the linear retrieval class with a specified configuration.
Parameters:
config(dict): Configuration settings for retrieval.
Return: None
reset()#
Description: Resets the tensor store to None.
Parameters: None
Return: None
__call__(query, topk, with_score, sort)#
Description: Retrieves indices of the top matching items compared to a query based on specified criteria.
Parameters:
query(varied): Query to find relevant indices.topk(string or int or bool): Controls number of top matches returned.with_score(bool): Whether to return scores alongside indices.sort(bool): Option to sort the scores descending.
Return: Either indices or pairs of scores and indices.
delete(index)#
Description: Removes an item at the specified index from the tensor store.
Parameters:
index(int): Index of the item to be deleted.
Return: None
get_tensor_by_ids(id_list)#
Description: Retrieves tensors by their IDs from the tensor store.
Parameters:
id_list(list): List of IDs to retrieve tensors.
Return: Tensor(s) corresponding to the given IDs.
__calculate_scores__(query)#
Description: Abstract method to calculate similarity scores between a query and items in the tensor store.
Parameters:
query(varied): Query to calculate scores against.
Return: Tensor of scores.
class TextRetrieval#
Description: Retrieval of the most relevant texts according to the query.
__init__(config)#
Description: Initializes the text retrieval class with encoder configuration.
Parameters:
config(dict): Configuration including encoder settings.
Return: None
__normalize__(embedding)#
Description: Normalizes an embedding using L2 normalization.
Parameters:
embedding(tensor): Embedding to normalize.
Return: Normalized tensor.
__calculate_scores__(query)#
Description: Calculates similarity scores between the query and texts in the tensor store using configured method.
Parameters:
query(varied): Query text to calculate scores against.
Return: Tensor of similarity scores.
add(text)#
Description: Adds a new text’s embedding to the tensor store, with optional normalization.
Parameters:
text(string): Text to add to the retrieval store.
Return: The added text’s embedding as a tensor.
update(index, text)#
Description: Updates the tensor store at a specific index with a new text embedding.
Parameters:
index(int): Index to be updated.text(string): New text to update at the specified index.
Return: None
class ValueRetrieval#
Description: Retrieval of certain values with several algorithms.
__init__(config)#
Description: Initializes the value retrieval class with a specific configuration.
Parameters:
config(dict): Configuration settings for value retrieval.
Return: None
__calculate_scores__(query)#
Description: Calculates scores for the retrieval based on the given mode in the configuration.
Parameters:
query(int or float): Reference value to calculate scores.
Return: Tensor of scores.
add(value)#
Description: Adds a value to the tensor store.
Parameters:
value(int or float): Value to add to the store.
Return: None
update(index, value)#
Description: Updates the tensor store at a specific index with a new value.
Parameters:
index(int): Index to be updated.value(int or float): New value to update at the specified index.
Return: None
class TimeRetrieval#
Description: Retrieval according to timestamps with several algorithms.
__init__(config)#
Description: Initializes the time retrieval class with a specific configuration.
Parameters:
config(dict): Configuration settings for time-based retrieval.
Return: None
__call__(query, topk, with_score, sort)#
Description: Retrieves indices of the top matching timestamps based on the specified mode and sort criteria.
Parameters:
query(int or float): Query referring to the reference time.topk(string or int or bool): Controls number of top timestamps returned.with_score(bool): Whether to return scores alongside indices.sort(bool): Option to sort the scores descending.
Return: Either indices or pairs of scores and indices.
class BaseSummarizer#
Description: Summarize texts into a brief summary, which can decrease the lengths of texts and emphasize critical points.
__init__(config)#
Description: Initializes the BaseSummarizer with a given configuration.
Parameters:
config(object): Configuration for the summarizer containing necessary parameters.
Return: None
reset()#
Description: Resets any internal state of the summarizer if necessary.
Parameters: None
Return: None
__call__(*args, **kwargs)#
Description: Abstract method to be implemented in derived classes for summarization functionality.
Parameters:
*args(tuple): Variable length argument list.**kwargs(dict): Arbitrary keyword arguments.
Return: None
class LLMSummarizer#
Description: Summarize via large language models.
__init__(config)#
Description: Initializes the LLMSummarizer with a given configuration and prepares the LLM based on the specified method.
Parameters:
config(object): Configuration for the LLM summarizer containing necessary parameters including LLM configuration.
Return: None
__call__(input_dict)#
Description: Generates a summary using the specified large language model based on the provided input.
Parameters:
input_dict(dict): Dictionary containing input variables to be formatted into the prompt.
Return: A summary generated by the large language model.
class BaseTrigger#
Description: Designed to call functions or tools in extensible manners.
__init__(config)#
Description: Initialize the BaseTrigger with a configuration.
Parameters:
config(object): Configuration object for the trigger.
Return: None
reset()#
Description: Reset the trigger state, if applicable.
Parameters: None
Return: None
__call__(args, kwargs)#
Description: Abstract method to be implemented by subclasses to define the trigger’s behavior.
Parameters:
args(tuple): Positional arguments for the trigger call.kwargs(dict): Keyword arguments for the trigger call.
Return: None
class LLMTrigger#
Description: Utilizing LLMs to determine which function should be called with specific arguments.
__init__(config)#
Description: Initialize the LLMTrigger with a configuration and register functions.
Parameters:
config(object): Configuration object that includes LLM settings and a list of functions.
Return: None
register(func)#
Description: Register a function with the trigger, storing its name, arguments, argument types, and descriptions.
Parameters:
func(dict): Dictionary containing function details such as name, args, args_type, func_description, and args_description.
Return: None
__get_function_prompt__()#
Description: Generate a string that describes all registered functions.
Parameters: None
Return: str
__parse_excuate_function__(res)#
Description: Parse the result from the LLM to extract function calls and their arguments.
Parameters:
res(str): The result string from the LLM containing possible function calls.
Return: list
__call__(input_dict)#
Description: Execute the LLMTrigger by generating a prompt, running it against the LLM, and parsing the response.
Parameters:
input_dict(dict): Dictionary containing input data to format the prompt.
Return: list
class BaseTruncation#
Description: Abstract base class to manage truncation of text based on the limitations of token numbers by language models
__init__(config)#
Description: Initializes the BaseTruncation class with provided config.
Parameters:
config(object): Configuration object containing settings for truncation.
Return: None
reset()#
Description: Resets any internal state of the truncation process, if necessary.
Parameters: None
Return: None
__call__()#
Description: Abstract method that must be implemented in subclasses to perform truncation.
Parameters: None
Return: None
class LMTruncation#
Description: Implements BaseTruncation to handle text truncation using language models, capable of truncating by words or tokens.
__init__(config)#
Description: Initializes the LMTruncation class with provided config and sets up a tokenizer if required.
Parameters:
config(object): Configuration object containing settings for mode, path, and number of words or tokens.
Return: None
truncate_by_word(text)#
Description: Truncates the input text by a specified number of words.
Parameters:
text(str): The text to be truncated.
Return: str
truncate_by_token(text)#
Description: Truncates the input text by a specified number of tokens.
Parameters:
text(str): The text to be truncated.
Return: str
get_piece_number(text)#
Description: Gets the number of words or tokens in the given text based on the current mode.
Parameters:
text(str): The text for which the number of pieces (words or tokens) is to be calculated.
Return: int
check_truncation_needed(text)#
Description: Checks if truncation is needed for the given text.
Parameters:
text(str): The text to check for truncation necessity.
Return: bool
__call__(text)#
Description: Truncates the text based on the configured mode (word or token).
Parameters:
text(str): The text to be truncated.
Return: str
class BaseUtilization#
Description: An abstract base class designed to deal with different parts of memory contents, and formulate these into a unified output format.
__init__(config)#
Description: Initializes the BaseUtilization with the specified configuration.
Parameters:
config(object): Configuration object for the utilization instance.
Return: None
reset()#
Description: Resets any persistent state. This method does not perform any actions in the base class and is meant to be overridden in subclasses if needed.
Parameters: None
Return: None
__call__(*args, **kwargs)#
Description: Abstract method to be implemented by subclasses, allows the instance to be called as a function.
Parameters:
*args(tuple): Variable length argument list.**kwargs(dict): Arbitrary keyword arguments.
Return: The specific output defined by the subclass implementation.
class ConcateUtilization#
Description: A class for concatenating memory pieces into a single string suitable for application output. Inherits from BaseUtilization.
__init__(config)#
Description: Initializes the ConcateUtilization with the specified configuration, calling the parent class initializer.
Parameters:
config(object): Configuration object specific to concatenation operations.
Return: None
concate_list(l)#
Description: Concatenates a list of memory pieces into a single string, applying index formatting and separators from the configuration.
Parameters:
l(list): The list of memory pieces to concatenate.
Return: A single concatenated string of the list elements.
__call__(input_memory)#
Description: Concatenates input memory based on its type (list, dict, or string) into a formatted string using the instance’s configuration.
Parameters:
input_memory(Union[list, dict, str]): The memory content to be processed and concatenated into a string. Can be a list, dict, or string.
Return: A string formatted with the defined prefix, concatenated main body, and suffix.
Memory Configurations#
class AttributeDict#
Description: A utility class that converts a dictionary into an object with attribute-style access.
__init__(dictionary)#
Description: Initializes the AttributeDict with the provided dictionary.
Parameters:
dictionary(dict): The dictionary to be converted into an AttributeDict.
Return: None
class MemoryConfig#
Description: Handles memory configuration settings, including GPU environment settings.
__init__(arg_config)#
Description: Initializes the MemoryConfig with an optional argument config.
Parameters:
arg_config(Union[dict, ArgConfig, str, None]): The configuration for arguments, which can be a dictionary or file path.
Return: None
__load_arg_config__(arg_config)#
Description: Loads and returns the argument configuration.
Parameters:
arg_config(Union[ArgConfig, dict, str, None]): The configuration input, which can be an ArgConfig object or other types.
Return: ArgConfig
class ArgConfig#
Description: Represents and loads configuration arguments from a dictionary or YAML file.
__init__(obj)#
Description: Initializes the ArgConfig object, loading configurations from a dictionary or YAML file.
Parameters:
obj(Union[dict, str, None]): The object containing configuration data, which can be a dictionary or a file path.
Return: None
add_args(dictionary)#
Description: Adds arguments from a dictionary to the ArgConfig as attributes.
Parameters:
dictionary(dict): A dictionary containing arguments to be added.
Return: None
Memory Utilities#
def generate_candidate(adjust_dict)#
Description: Generates a list of configurations with various combinations based on the given adjustment dictionary, which is useful for hyper-parameter tuning.
Parameters:
adjust_dict(dict): A dictionary containing configuration details for generating candidate models.
Return: list: A list of dictionaries, each representing a configuration candidate with a specific model and adjusted hyper-parameters.
def automatic_select(reward_func, model_candidate)#
Description: Selects the best performing memory model based on the given reward function, which evaluates each candidate model and returns a sorted list of performances.
Parameters:
reward_func(function): A function that evaluates the performance of a memory model and returns a float score.model_candidate(list): A list of candidate models generated by generate_candidate.
Return: list: A sorted list of dictionaries, each containing the index, score, model name, and configuration of the candidate models, ordered by performance scores in descending order.
class Client#
Description: Client class for interacting with a server using HTTP requests. It allows initializing sessions, storing data, recalling data, and performing various operations on the server.
__init__(base_url)#
Description: Initializes the Client with the base URL and retrieves a session ID.
Parameters:
base_url(str): The base URL of the server.
Return: None
request(route_path, data)#
Description: Sends a POST request to the specified route path with the provided data.
Parameters:
route_path(str): The path of the route at the server to send the request.data(dict): The JSON data to be sent with the request.
Return: dict or bool - Returns the response JSON if successful, otherwise returns False.
get_session_id()#
Description: Initializes the session by requesting a session ID from the server.
Parameters: None
Return: str - The session ID.
initilize_memory(method, config_dict)#
Description: Initializes the memory on the server with the specified method and configuration.
Parameters:
method(str): The method to initialize memory with.config_dict(dict): Configuration options for the method.
Return: str or None - Info response from the server or None if unsuccessful.
reset()#
Description: Resets the server’s state.
Parameters: None
Return: str or None - Info response from the server or None if unsuccessful.
store(observation)#
Description: Stores an observation on the server. Converts the observation to the required format if necessary.
Parameters:
observation(str or dict): The observation to store, which can be a string or dictionary.
Return: str or None - Info response from the server or None if unsuccessful.
recall(query)#
Description: Recalls data from the server using a query. Converts the query to the required format if necessary.
Parameters:
query(str or dict): The recall query, which can be a string or dictionary.
Return: object - The recalled data or None if unsuccessful.
manage(operation, kwargs)#
Description: Manages the server operations with specific parameters.
Parameters:
operation(str): The manage operation to perform.kwargs(dict): Additional arguments for the operation.
Return: str or None - Info response from the server or None if unsuccessful.
display()#
Description: Displays the server’s current state.
Parameters: None
Return: str or None - Info response from the server or None if unsuccessful.
optimize(kwargs)#
Description: Optimizes server operations based on provided parameters.
Parameters:
kwargs(dict): Optimization parameters.
Return: str or None - Info response from the server or None if unsuccessful.
class BaseDisplay#
Description: An abstract base class for display mechanisms.
__init__(config, register_dict)#
Description: Constructor for the BaseDisplay class.
Parameters:
config(object): Configuration object for controlling display settings.register_dict(dict): A dictionary containing register values to display.
Return: None
reset()#
Description: Reset the display. This method currently does nothing and serves as a placeholder for subclasses to override if needed.
Parameters: None
Return: None
__call__()#
Description: Abstract method that subclasses must implement to define how to display content.
Parameters: None
Return: None
class TextDisplay#
Description: Display memory contents in text format.
__init__(config, register_dict)#
Description: Constructor for the TextDisplay class.
Parameters:
config(object): Configuration object for controlling text display settings.register_dict(dict): A dictionary containing register values to display.
Return: None
__get_one_item__(name, obj)#
Description: Helper method to get a single item’s display string.
Parameters:
name(str): The name of the item.obj(object): The object to display.
Return: str
__get_display_memory__(tag)#
Description: Format the entire memory content for display.
Parameters:
tag(str): A tag to prefix the display content.
Return: str
class ScreenDisplay#
Description: Display memory contents on the console.
__init__(config, register_dict)#
Description: Constructor for the ScreenDisplay class.
Parameters:
config(object): Configuration object for controlling display settings.register_dict(dict): A dictionary containing register values to display.
Return: None
__call__(tag)#
Description: Print the display memory content to the console.
Parameters:
tag(str): A tag to prefix the display content when printing.
Return: None
class FileDisplay#
Description: Display memory contents in a file.
__init__(config, register_dict)#
Description: Constructor for the FileDisplay class.
Parameters:
config(object): Configuration object for controlling file display settings.register_dict(dict): A dictionary containing register values to display.
Return: None
__call__(tag)#
Description: Append the display memory content to a file.
Parameters:
tag(str): A tag to prefix the display content when writing to the file.
Return: None
class BaseStorage#
Description: Abstract base class for storage structures.
__init__(config)#
Description: Initializes the storage with a given configuration.
Parameters:
config(dict): Configuration for the storage.
Return: None
reset()#
Description: Resets the storage to its initial state.
Parameters: None
Return: None
display()#
Description: Displays the content of the storage.
Parameters: None
Return: str
is_empty()#
Description: Checks if the storage is empty.
Parameters: None
Return: bool
class LinearStorage#
Description: Memory storage in linear structure.
__init__(config)#
Description: Initializes the linear storage with a given configuration.
Parameters:
config(dict): Configuration for the linear storage.
Return: None
reset()#
Description: Clears the memory and resets the counter.
Parameters: None
Return: None
display()#
Description: Displays each memory entity in the linear storage.
Parameters: None
Return: str
clear_memory(start, end)#
Description: Clears a portion of the memory based on start and end indices.
Parameters:
start(int or None): Start index of the memory to be cleared.end(int or None): End index of the memory to be cleared.
Return: None
get_element_number()#
Description: Returns the number of elements in memory.
Parameters: None
Return: int
is_empty()#
Description: Checks if the memory list is empty.
Parameters: None
Return: bool
get_memory_element_by_mid(mid)#
Description: Gets a memory element by its ID.
Parameters:
mid(int): The memory ID to look up.
Return: dict
get_memory_attribute_by_mid(mid, attr)#
Description: Gets a specific attribute of a memory element by its ID.
Parameters:
mid(int): The memory ID to look up.attr(str): The attribute of the memory to retrieve.
Return: Any
get_memory_text_by_mid(mid)#
Description: Gets the text attribute of a memory element by its ID.
Parameters:
mid(int): The memory ID to look up.
Return: str
get_mids_by_attribute(attr, value)#
Description: Finds memory IDs by checking for a matching attribute value.
Parameters:
attr(str): The attribute to check.value(Any): The value to compare against.
Return: list[int]
update_memory_attribute_by_mid(mid, attr, value)#
Description: Updates an attribute of a memory element by its ID.
Parameters:
mid(int): The memory ID to update.attr(str): The attribute to update.value(Any): The new value for the attribute.
Return: None
add(obj)#
Description: Adds a new object to the memory list.
Parameters:
obj(dict): The object to add, must contain ‘text’.
Return: None
delete_by_mid(mid)#
Description: Deletes a memory element by its ID.
Parameters:
mid(int): The memory ID to delete.
Return: None
delete_by_mid_list(mids)#
Description: Deletes multiple memory elements by their IDs.
Parameters:
mids(list[int]): A list of memory IDs to delete.
Return: None
get_all_memory_in_order()#
Description: Returns all memory elements in order.
Parameters: None
Return: list[dict]
class GraphStorage#
Description: Memory storage in graph structure.
__init__(config)#
Description: Initializes the graph storage with a given configuration.
Parameters:
config(dict): Configuration for the graph storage.
Return: None
reset()#
Description: Resets the graph, clearing nodes, edges and counters.
Parameters: None
Return: None
display()#
Description: Displays nodes and edges in the graph storage.
Parameters: None
Return: str
get_element_number()#
Description: Returns the number of nodes in the graph.
Parameters: None
Return: int
is_empty()#
Description: Checks if the graph is empty.
Parameters: None
Return: bool
get_node_id_by_mid(mid)#
Description: Gets the node ID by memory ID.
Parameters:
mid(int): The memory ID to look up.
Return: int
get_mid_by_node_id(node_id)#
Description: Gets the memory ID by node ID.
Parameters:
node_id(int): The node ID to look up.
Return: int
get_memory_element_by_node_id(node_id)#
Description: Gets a memory element by its node ID.
Parameters:
node_id(int): The node ID to look up.
Return: dict
get_memory_element_by_mid(mid)#
Description: Gets a memory element by its memory ID.
Parameters:
mid(int): The memory ID to look up.
Return: dict
get_memory_text_by_node_id(node_id)#
Description: Gets the text attribute of a memory element by its node ID.
Parameters:
node_id(int): The node ID to look up.
Return: str
get_memory_text_by_mid(mid)#
Description: Gets the text attribute of a memory element by its memory ID.
Parameters:
mid(int): The memory ID to look up.
Return: str
update_memory_attribute_by_node_id(node_id, attr, value)#
Description: Updates an attribute of a memory node by its node ID.
Parameters:
node_id(int): The node ID to update.attr(str): The attribute to update.value(Any): The new value for the attribute.
Return: None
update_memory_attribute_by_mid(mid, attr, value)#
Description: Updates an attribute of a memory node by its memory ID.
Parameters:
mid(int): The memory ID to update.attr(str): The attribute to update.value(Any): The new value for the attribute.
Return: None
__update_memory_order_map__()#
Description: Updates the memory order map to align node IDs and memory IDs.
Parameters: None
Return: None
add_node(obj)#
Description: Adds a new node object to the graph.
Parameters:
obj(dict): The node object to add, must contain ‘text’.
Return: int
add_edge(s, t, obj)#
Description: Adds an edge connecting two nodes in the graph.
Parameters:
s(int): The source node ID of the edge.t(int): The target node ID of the edge.obj(dict): The edge object containing additional properties.
Return: int
This API reference is organized with the help of LLMs. If you have some questions, please feel free to contact us.