Memory#

1. Concept#

The Memory module in CAMEL provides a flexible and powerful system for storing, retrieving, and managing information for AI agents. It enables agents to maintain context across conversations and retrieve relevant information from past interactions, enhancing the coherence and relevance of AI responses.

2. Get Started#

2.1 Basic Usage#

Here’s a quick example of how to use the LongtermAgentMemory:

from camel.memories import (
    ChatHistoryBlock,
    LongtermAgentMemory,
    MemoryRecord,
    ScoreBasedContextCreator,
    VectorDBBlock,
)
from camel.messages import BaseMessage
from camel.types import ModelType, OpenAIBackendRole
from camel.utils import OpenAITokenCounter

# Initialize the memory
memory = LongtermAgentMemory(
    context_creator=ScoreBasedContextCreator(
        token_counter=OpenAITokenCounter(ModelType.GPT_4O_MINI),
        token_limit=1024,
    ),
    chat_history_block=ChatHistoryBlock(),
    vector_db_block=VectorDBBlock(),
)

# Create and write new records
records = [
    MemoryRecord(
        message=BaseMessage.make_user_message(
            role_name="User",
            meta_dict=None,
            content="What is CAMEL AI?",
        ),
        role_at_backend=OpenAIBackendRole.USER,
    ),
    MemoryRecord(
        message=BaseMessage.make_assistant_message(
            role_name="Agent",
            meta_dict=None,
            content="CAMEL-AI.org is the 1st LLM multi-agent framework and "
            "an open-source community dedicated to finding the scaling law "
            "of agents.",
        ),
        role_at_backend=OpenAIBackendRole.ASSISTANT,
    ),
]
memory.write_records(records)

# Get context for the agent
context, token_count = memory.get_context()

print(context)
print(f"Retrieved context (token count: {token_count}):")
for message in context:
    print(f"{message}")

>>> Retrieved context (token count: 49):
{'role': 'user', 'content': 'What is AI?'}
{'role': 'assistant', 'content': 'AI refers to systems that mimic human intelligence.'}

Adding LongtermAgentMemory to your ChatAgent:

from camel.agents import ChatAgent

# Define system message for the agent
sys_msg = BaseMessage.make_assistant_message(
    role_name='Agent',
    content='You are a curious agent wondering about the universe.',
)

# Initialize agent
agent = ChatAgent(system_message=sys_msg)

# Set memory to the agent
agent.memory = memory


# Define a user message
usr_msg = BaseMessage.make_user_message(
    role_name='User',
    content="Tell me which is the 1st LLM multi-agent framework based on what we have discussed",
)

# Sending the message to the agent
response = agent.step(usr_msg)

# Check the response (just for illustrative purpose)
print(response.msgs[0].content)

>>> CAMEL AI is recognized as the first LLM (Large Language Model) multi-agent framework. It is an open-source community initiative focused on exploring the scaling laws of agents, enabling the development and interaction of multiple AI agents in a collaborative environment. This framework allows researchers and developers to experiment with various configurations and interactions among agents, facilitating advancements in AI capabilities and understanding.

3. Core Components#

3.1 MemoryRecord#

The basic data unit in the CAMEL memory system.

Attributes:

message: The main content of the record (BaseMessage)
role_at_backend: The role this message played at the OpenAI backend (OpenAIBackendRole)
uuid: A unique identifier for the record
extra_info: Additional key-value pairs for extra information

Methods:

from_dict(): Static method to construct a MemoryRecord from a dictionary
to_dict(): Convert the MemoryRecord to a dictionary for serialization
to_openai_message(): Convert the record to an OpenAIMessage object

3.2 ContextRecord#

The result of memory retrieval from AgentMemory, which is used by ContextCreator to select records for creating the context for the agent.

Attributes:

memory_record: A MemoryRecord
score: A float value representing the relevance or importance of the record

3.3 MemoryBlock (Abstract Base Class)#

Serves as the fundamental component within the agent memory system. MemoryBlock class follows the “Composite Design pattern” so that you can compose objects into tree structures and then work with these structures as if they were individual objects.

write_records(): Write multiple records to memory
write_record(): Write a single record to memory
clear(): Remove all records from memory

3.4 BaseContextCreator (Abstract Base Class)#

Defines the context creation strategies when retrieval messages exceed the model’s context length.

token_counter: For counting tokens in a message
token_limit: The maximum number of tokens allowed in the generated context
create_context(): Create conversational context from chat history

3.5 AgentMemory (Abstract Base Class)#

A specialized form of MemoryBlock designed for direct integration with an agent.

retrieve(): Get a list of ContextRecords from memory
get_context_creator(): Get a context creator
get_context(): Get chat context with a proper size for the agent

4. Memory Block Implementations#

4.1 ChatHistoryBlock#

Stores all the chat histories and retrieves with optional window size.

Initialization:

storage: Optional BaseKeyValueStorage (default: InMemoryKeyValueStorage)
keep_rate: Float value for scoring historical messages (default: 0.9)

Methods:

retrieve(): Get recent chat history with optional window size
write_records(): Write new records to chat history
clear(): Remove all chat messages

Use Case: Ideal for maintaining recent conversation context and flow.

4.2 VectorDBBlock#

Uses vector embeddings for maintaining and retrieving information.

Initialization:

storage: Optional BaseVectorStorage (default: QdrantStorage)
embedding: Optional BaseEmbedding (default: OpenAIEmbedding)

Methods:

retrieve(): Get similar records based on a keyword
write_records(): Convert and write new records to the vector database
clear(): Remove all records from the vector database

Use Case: Better for retrieving semantically relevant information across a large set of messages.

Key Differences:

Storage Mechanism: ChatHistoryBlock uses key-value storage, VectorDBBlock uses vector database storage.
Retrieval Method: ChatHistoryBlock retrieves based on recency, VectorDBBlock retrieves based on semantic similarity.
Data Representation: ChatHistoryBlock stores messages in original form, VectorDBBlock converts to vector embeddings.

5. Agent Memory Implementations#

5.1 ChatHistoryMemory#

An AgentMemory implementation that wraps ChatHistoryBlock.

Initialization:

context_creator: BaseContextCreator
storage: Optional BaseKeyValueStorage
window_size: Optional int for retrieval window

Methods:

retrieve(): Get recent chat messages
write_records(): Write new records to chat history
get_context_creator(): Get the context creator
clear(): Remove all chat messages

5.2 VectorDBMemory#

An AgentMemory implementation that wraps VectorDBBlock.

Initialization:

context_creator: BaseContextCreator
storage: Optional BaseVectorStorage
retrieve_limit: int for maximum number of retrieved messages (default:3)

Methods:

retrieve(): Get relevant messages from the vector database
write_records(): Write new records and update current topic
get_context_creator(): Get the context creator

5.3 LongtermAgentMemory#

Combines ChatHistoryMemory and VectorDBMemory for comprehensive memory management.

Initialization:

context_creator: BaseContextCreator
chat_history_block: Optional ChatHistoryBlock
vector_db_block: Optional VectorDBBlock
retrieve_limit: int for maximum number of retrieved messages (default: 3)

Methods:

retrieve(): Get context from both chat history and vector database
write_records(): Write new records to both chat history and vector database
get_context_creator(): Get the context creator
clear(): Remove all records from both memory blocks

6. Advanced Topics#

6.1 Customizing Context Creator#

You can create custom context creators by subclassing BaseContextCreator:

from camel.memories import BaseContextCreator

class MyCustomContextCreator(BaseContextCreator):
    @property
    def token_counter(self):
        # Implement your token counting logic
        return 

    @property
    def token_limit(self):
        return 1000  # Or any other limit

    def create_context(self, records):
        # Implement your context creation logic
        pass

6.2 Customizing Vector Database Block#

For VectorDBBlock, you can customize it by adjusting the embedding models or vector storages:

from camel.embeddings import OpenAIEmbedding
from camel.memories import VectorDBBlock
from camel.storages import QdrantStorage

vector_db = VectorDBBlock(
    embedding=OpenAIEmbedding(),
    storage=QdrantStorage(vector_dim=OpenAIEmbedding().get_output_dim()),
)

6.3 Performance Considerations#

For large-scale applications, consider using persistent storage backends instead of in-memory storage.
Optimize your context creator to balance between context relevance and token limits.
When using VectorDBMemory, be mindful of the trade-off between retrieval accuracy and speed as the database grows.

Memory

Contents

Memory#

1. Concept#

2. Get Started#

2.1 Basic Usage#

3. Core Components#

3.1 MemoryRecord#

3.2 ContextRecord#

3.3 MemoryBlock (Abstract Base Class)#

3.4 BaseContextCreator (Abstract Base Class)#

3.5 AgentMemory (Abstract Base Class)#

4. Memory Block Implementations#

4.1 ChatHistoryBlock#

4.2 VectorDBBlock#

5. Agent Memory Implementations#

5.1 ChatHistoryMemory#

5.2 VectorDBMemory#

5.3 LongtermAgentMemory#

6. Advanced Topics#

6.1 Customizing Context Creator#

6.2 Customizing Vector Database Block#

6.3 Performance Considerations#