initialize_agent not saving and returning messages in memory

System Info

langchain version == 0.0.271

Who can help?

@hwchase17

Information

[X] The official example notebooks/scripts
[X] My own modified scripts

Related Components

[ ] LLMs/Chat Models
[ ] Embedding Models
[ ] Prompts / Prompt Templates / Prompt Selectors
[ ] Output Parsers
[ ] Document Loaders
[ ] Vector Stores / Retrievers
[X] Memory
[X] Agents / Agent Executors
[ ] Tools / Toolkits
[ ] Chains
[ ] Callbacks/Tracing
[ ] Async

Reproduction

chat_history = MessagesPlaceholder(variable_name="chat_history") memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True)

llm = ChatOpenAI(temperature=0, model_name='gpt-3.5-turbo') agent_chain = initialize_agent( tools=tools, llm=llm, agent=AgentType.STRUCTURED_CHAT_ZERO_SHOT_REACT_DESCRIPTION, verbose=True, memory=memory, max_iterations=1, early_stopping_method='generate', handle_parsing_errors=True, system_prompt="You are an assistant named Mikey", agent_kwargs={ "memory_prompts": [chat_history], "input_variables": ["input", "agent_scratchpad", "chat_history"], } ) agent_chain.run(input="My name is Dev")

agent_chain.run(input="What is my name")

Expected behavior

Expected behavior final answer: Your name is Dev

Output behavior: Final answer: I don't have access to personal information like your name. Is there anything else I can help you with?

Note: What happens is that the memory is not being passed along, so the model doesn't know my name

🤖

Hello,

Thank you for bringing this issue to our attention. It seems like the problem you're experiencing is related to the memory handling in the LangChain framework. Specifically, the run method in the ConversationChain class is not implemented to save the context and return the memory, which is why the model doesn't recognize previously mentioned information like a user's name.

Here are a few similar issues that have been solved in the past:

Initialize_agent not storing messages when memory is present: The issue was related to how the save_context method in the ConversationBufferMemory class was functioning. The solution was to ensure that the keys used in self._get_input_output(inputs, outputs) correctly correspond to the keys used to store the input and output strings in the inputs and outputs dictionaries. Additionally, the add_user_message and add_ai_message methods needed to be correctly adding the messages to the agent's memory.
Does the new Structured Chat Agent support ConversationMemory?: The solution was to add chat history to the STRUCTURED_CHAT_ZERO_SHOT_REACT_DESCRIPTION agent with the following code:

chat_history = MessagesPlaceholder(variable_name="chat_history")
memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True)

agent = initialize_agent(
    tools,
    llm,
    agent=AgentType.STRUCTURED_CHAT_ZERO_SHOT_REACT_DESCRIPTION,
    verbose=True,
    agent_kwargs={
        "memory_prompts": [chat_history],
        "input_variables": ["input", "agent_scratchpad", "chat_history"]
    },
    memory=memory,
)

Agent + Memory causing issues with tool interaction: The issue was resolved by adding more clarity to the prompt to distinguish between new input and chat history.

Based on these solutions, I would recommend checking the keys used in your inputs and outputs dictionaries and ensuring that the add_user_message and add_ai_message methods are correctly adding the messages to the agent's memory. Additionally, you might want to add more clarity to your prompt to distinguish between new input and chat history.

I hope this helps! If you have any other questions or need further clarification, please don't hesitate to ask.

Best, Dosu

Sources

This response is meant to be useful and save you time. It is not meant to be a precise solution, but rather a starting point for your own research.

Help me be more useful! Please leave a 👍 if this is helpful and 👎 if it is irrelevant.

If you want to continue the conversation, start your reply with @dosu-bot.

langchain-ai / langchain