llm.generate_str use_history is not working

[HuntZhaozq]

I use the example of strealit_mcp_basic_agent. And set the use_history=True in the llm.generate_str method like the above.

mcp-agent/examples/streamlit_mcp_basic_agent/main.py

Line 57 in af54d24

message=prompt, request_params=RequestParams(use_history=True)

But it is not working. When I chat with the agent in the second round. The chat history of the first round is gone. How to solve this?

[saqadri]

I use the example of strealit_mcp_basic_agent. And set the use_history=True in the llm.generate_str method like the above.

mcp-agent/examples/streamlit_mcp_basic_agent/main.py

Line 57 in af54d24

message=prompt, request_params=RequestParams(use_history=True)
But it is not working. When I chat with the agent in the second round. The chat history of the first round is gone. How to solve this?

@HuntZhaozq thanks for reporting this issue. This is an annoying thing about Streamlit, which reruns the script top-to-bottom, overwriting anything that is held in-memory.

The solution is to save the messages into Streamlit session state (st.session_state), and pass the full messages array each time.

So in that example application you linked, change:

mcp-agent/examples/streamlit_mcp_basic_agent/main.py

Line 57 in af54d24

message=prompt, request_params=RequestParams(use_history=True)

to

response = await llm.generate_str(
    message=st.session_state["messages"], request_params=RequestParams(use_history=True)
)

I am working on proper session state management for Streamlit so you don't have to think about these problems (e.g. when use_history is True, in streamlit mode the framework should automatically be saving things to streamlit session_state). So I'll leave this issue open until that's fixed.

But please let me know if my workaround above works for you.

[HuntZhaozq]

Yes, it works. Thank you!

[doncat99]

I use the example of strealit_mcp_basic_agent. And set the use_history=True in the llm.generate_str method like the above.
mcp-agent/examples/streamlit_mcp_basic_agent/main.py
Line 57 in af54d24
message=prompt, request_params=RequestParams(use_history=True)
But it is not working. When I chat with the agent in the second round. The chat history of the first round is gone. How to solve this?

@HuntZhaozq thanks for reporting this issue. This is an annoying thing about Streamlit, which reruns the script top-to-bottom, overwriting anything that is held in-memory.

The solution is to save the messages into Streamlit session state (st.session_state), and pass the full messages array each time.

So in that example application you linked, change:

mcp-agent/examples/streamlit_mcp_basic_agent/main.py

Line 57 in af54d24

message=prompt, request_params=RequestParams(use_history=True)
to

response = await llm.generate_str(
    message=st.session_state["messages"], request_params=RequestParams(use_history=True)
)

I am working on proper session state management for Streamlit so you don't have to think about these problems (e.g. when use_history is True, in streamlit mode the framework should automatically be saving things to streamlit session_state). So I'll leave this issue open until that's fixed.

But please let me know if my workaround above works for you.

@saqadri, Should mcp-agent implement session management in close loop, instead of relying on Streamlit?

And the readme file points out that "Memory -- adding support for long-term memory" is on the roadmap. I do believe Session management is part of the Memory.

Btw: Memory is the most valuable feature on the list.

Roadmap
We will be adding a detailed roadmap (ideally driven by your feedback). The current set of priorities include:

Durable Execution -- allow workflows to pause/resume and serialize state so they can be replayed or be paused indefinitely. We are working on integrating [Temporal](https://github.com/lastmile-ai/mcp-agent/blob/main/src/mcp_agent/executor/temporal.py) for this purpose.
Memory -- adding support for long-term memory
Streaming -- Support streaming listeners for iterative progress
Additional MCP capabilities -- Expand beyond tool calls to support:
Resources
Prompts
Notifications