langchain-ai/langgraph
1
0
Fork 0
mirror of https://github.com/langchain-ai/langgraph.git synced 2026-07-15 12:35:43 -04:00
Code Issues 314 Packages Projects Releases 17 Wiki Activity

HELP IN ADDING MEMORY TO REACT AGENT #794

New Issue
Closed
opened 2026-02-20 17:41:47 -05:00 by yindo · 4 comments
yindo commented 2026-02-20 17:41:47 -05:00
Owner
Copy Link
Copy Source

Originally created by @yadgire-saurabh on GitHub (Jul 11, 2025).

I have built a react agent with a pre_model_hook [code blocks attached]

`
def pre_model_hook(state):
"""Trims messages to fit in the context window."""
trimmed_messages = trim_messages(
state["messages"],
strategy="last",
token_counter=count_tokens_approximately,
max_tokens=10000,
start_on="human",
end_on=("human", "tool"),
allow_partial = True,
)
return {"llm_input_messages": trimmed_messages}

agent_executor = create_react_agent(
llm,
tools=mcp_server_tools,
state_schema=CustomAgentState,
prompt=prompt_func,
pre_model_hook=pre_model_hook,
checkpointer=checkpointer
)
`

I have a few questions:

How can I see the contents of y memory/pre_model_hook without streaming output (I am using invoke)
Is the pre_model_hook used as memory/context for subsequent calls
How is post_model_hook different?
If I pass tools as ToolNode([mcp_tools]), it does not work
Can you help me with any of these issues I am facing?

Image

EDIT: The trimmed messages only shows the last HumanMessage.
Here is my agent graph image

Thanks!

Originally created by @yadgire-saurabh on GitHub (Jul 11, 2025). I have built a react agent with a pre_model_hook [code blocks attached] ` def pre_model_hook(state): """Trims messages to fit in the context window.""" trimmed_messages = trim_messages( state["messages"], strategy="last", token_counter=count_tokens_approximately, max_tokens=10000, start_on="human", end_on=("human", "tool"), allow_partial = True, ) return {"llm_input_messages": trimmed_messages} agent_executor = create_react_agent( llm, tools=mcp_server_tools, state_schema=CustomAgentState, prompt=prompt_func, pre_model_hook=pre_model_hook, checkpointer=checkpointer ) ` I have a few questions: How can I see the contents of y memory/pre_model_hook without streaming output (I am using invoke) Is the pre_model_hook used as memory/context for subsequent calls How is post_model_hook different? If I pass tools as ToolNode([mcp_tools]), it does not work Can you help me with any of these issues I am facing? <img width="1098" height="1216" alt="Image" src="https://github.com/user-attachments/assets/e29ef294-5099-41b8-9cc6-4ff82dfa0fcb" /> EDIT: The trimmed messages only shows the last HumanMessage. Here is my agent graph image Thanks!
yindo added the invalidawaiting author revision labels 2026-02-20 17:41:47 -05:00
yindo closed this issue 2026-02-20 17:41:47 -05:00
yindo commented 2026-02-20 17:41:48 -05:00
Author
Owner
Copy Link
Copy Source

@yoadsn commented on GitHub (Jul 11, 2025):

@yadgire-saurabh I hope I understood your question and might be able to help with any of them.

If you want to see the state after the pre_model_hook you will need to use a checkpointer, and inspect the "state history" of the agent. (the CompiledStateGraph).
You can also stream the updates as you mentioned - I think for debugging that more convenient.

The pre_model_hook is called before the "agent" node any time the node is about to be invoked (after "start" or after "tools" are done executing).

The post_model_hook is called after the "agent" node any time it's done producing a result (with or without a tool call).

I am not sure why the tool passing does not work for you, I suggest debugging the Library code itself - I do that often when things do not behave as they should.

Reg "message trimming" - I am not using that functionality myself, and so cannot comment on that.

hth

@yoadsn commented on GitHub (Jul 11, 2025): @yadgire-saurabh I hope I understood your question and might be able to help with any of them. If you want to see the state after the `pre_model_hook` you will need to use a checkpointer, and inspect the "state history" of the agent. (the CompiledStateGraph). You can also stream the updates as you mentioned - I think for debugging that more convenient. The `pre_model_hook` is called before the "agent" node any time the node is about to be invoked (after "start" or after "tools" are done executing). The `post_model_hook` is called after the "agent" node any time it's done producing a result (with or without a tool call). I am not sure why the tool passing does not work for you, I suggest debugging the Library code itself - I do that often when things do not behave as they should. Reg "message trimming" - I am not using that functionality myself, and so cannot comment on that. hth
yindo commented 2026-02-20 17:41:48 -05:00
Author
Owner
Copy Link
Copy Source

@yadgire-saurabh commented on GitHub (Jul 11, 2025):

Thanks for the detailed response. I already have a checkpointer added at the time of invoking the react agent (a chat assistant). The thread_id is the user ID which remains unchanged throughout the chat instance. I think the issue is - I compiled the graph for every user input in the chat.

Will debug!

@yadgire-saurabh commented on GitHub (Jul 11, 2025): Thanks for the detailed response. I already have a checkpointer added at the time of invoking the react agent (a chat assistant). The thread_id is the user ID which remains unchanged throughout the chat instance. I think the issue is - I compiled the graph for every user input in the chat. Will debug!
yindo commented 2026-02-20 17:41:48 -05:00
Author
Owner
Copy Link
Copy Source

@dlrklc commented on GitHub (Jul 13, 2025):

Hi @yadgire-saurabh, for tool passing this works for me when I use create_react_agent
For example, if there is a calculator tool

from langchain.tools import Tool
calculator_tool = Tool(`
    name="calculator",
    func=calculate,
    description="Use this tool to perform basic arithmetic calculations. Input should be a math expression like '2 + 2'."
)

or if I created with @tool decorator

from langchain_core.tools import tool

@tool
def calculator_tool(expression: str) -> str:
    return str(eval(expression))

And then,

agent = create_react_agent(llm=llm, tools=[calculator_tool], prompt=prompt)

Maybe the reason is not having square brackets around mcp_server_tools in your code.

Hope this helps!

@dlrklc commented on GitHub (Jul 13, 2025): Hi @yadgire-saurabh, for tool passing this works for me when I use `create_react_agent` For example, if there is a calculator tool ```python from langchain.tools import Tool calculator_tool = Tool(` name="calculator", func=calculate, description="Use this tool to perform basic arithmetic calculations. Input should be a math expression like '2 + 2'." ) ``` or if I created with `@tool` decorator ```python from langchain_core.tools import tool @tool def calculator_tool(expression: str) -> str: return str(eval(expression)) ``` And then, ```python agent = create_react_agent(llm=llm, tools=[calculator_tool], prompt=prompt) ``` Maybe the reason is not having square brackets around mcp_server_tools in your code. Hope this helps!
yindo commented 2026-02-20 17:41:48 -05:00
Author
Owner
Copy Link
Copy Source

@eyurtsev commented on GitHub (Jul 14, 2025):

@yadgire-saurabh could you please provide an MRE or check whether @dlrklc identified your problem correctly

@eyurtsev commented on GitHub (Jul 14, 2025): @yadgire-saurabh could you please provide an MRE or check whether @dlrklc identified your problem correctly
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
agents
awaiting author revision
bug
change
chat models
checkpointer
ci
code quality
code quality
dependencies
documentation
drawing
duplicate
enhancement
good first issue
help wanted
invalid
maintainer
meta
missing information
not maintained by langchain
open-swe-auto
open-swe-max
open-swe-max-auto
pending
performance
pull-request
Mirrored from GitHub Pull Request
 question
requires design
sdk
server
stale
studio
tests
type checking
usage question
v0.5
v1.1
No Label awaiting author revision invalid
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
yindo
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: langchain-ai/langgraph#794
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?