SummarizationMiddleware

OSS Product Help LangGraph
1

Screenshot from 2025-12-10 01-45-09
Screenshot from 2025-12-10 01-45-091240Ă—861 142 KB

I am using SummarizationMiddleware ,

        SummarizationMiddleware(
            model = model,
            trigger=[("tokens", 200), ("messages", 2)],
            keep = ("messages", 3)
        ),

my total token 8598 but SummarizationMiddleware are not calling

2

Hi @mohiuddin1221

could you share the full agent code, so that I can reproduce the issue?

3

async def agent(state: InputState, runtime: Runtime[Context]):
client = MultiServerMCPClient(
{
“RedshopInternalMcpSercer”: {
“url”: “http://127.0.0.1:8000/mcp/”,
“transport”: “streamable_http”,
}
}
)
tools = await client.get_tools()
all_tools = tools + [send_email_tool]

agent_instance = create_agent(
    model=model,
    tools=all_tools,
    state_schema=InputState,
    system_prompt=create_Aget_prompt(state.get("user_input"), state.get("rephased_question")),
    context_schema=Context,
    checkpointer=InMemorySaver(),
    middleware=[
        HumanInTheLoopMiddleware(
            interrupt_on={
                "send_email_tool": {
                    "allowed_decisions": ["approve", "edit", "reject"],
                },
            }
        ),
        SummarizationMiddleware(
             model = model,
             trigger=[("tokens", 200), ("messages", 2)],
             keep = ("messages", 3)

        ),
        ToolCallLimitMiddleware(thread_limit=5, run_limit=5),
        handle_tool_errors,
    ],
)

rephased_question = state.get("rephased_question")
print("rephased_question", rephased_question)
result = await agent_instance.ainvoke(
        {"rephased_question": rephased_question}
)
print("Result message..................", result)
return result

If the code quality needs improvement, then please give me suggestions.

4

@mohiuddin1221

I am looking at it today

5

hi @mohiuddin1221

could you also share InputState and Context? Use code block for code snippet please.

6

@dataclass
class Context:
user_id: str
user_name: str
user_preference: str

class InputState(TypedDict):
user_input: str
rephased_question: str
answer: str

7

Ok, I see the issue.
This middleware operates only on the messages key in the agent state. It does not summarize (or otherwise transform) the whole state. There is no messages on your state.

1 Like
8

class InputState(TypedDict):
user_input: str
rephased_question: str
answer: str
thread_name: str
messages: Annotated[list[AnyMessage], add_messages]

#############
SummarizationMiddleware(
model = model,
trigger=[(“tokens”, 200)],
keep = (“messages”, 3)

    ),

I am pushing the state fields like user_input, rephased_question, and answer into every turn of messages. All messages are present, but still, the SummarizationMiddleware is not being called.

9

Can you share you entire agent code?