Tool messages are duplicated when they immediately follow AI message with tool calls

[yash025]

Description

I'm encountering an issue where tool messages get duplicated during summarization when they immediately follow an AI message with tool calls. This duplication causes a validation error from the LLM provider because the duplicated tool messages don't have a corresponding AI message with tool calls immediately before them.

Error Message

{'error_message': 'Chat completion failed', 'cause': "BadRequestError: AzureException BadRequestError - Invalid parameter: messages with role 'tool' must be a response to a preceeding message with 'tool_calls'."}

Expected Behavior

Tool messages should only appear once in the summarized output, maintaining the natural conversation flow and proper message sequence validation.

Reproduction

Here's a minimal example of the message structure that triggers this:

messages = [
    HumanMessage(content="Analyze metrics for asset XYZ"),
    AIMessage(content="", tool_calls=[
        {"id": "call_1", "name": "query_errors", "args": {...}},
        {"id": "call_2", "name": "query_latency", "args": {...}},
    ]),
    ToolMessage(content="error_results", tool_call_id="call_1"),
    ToolMessage(content="latency_results", tool_call_id="call_2"),
]

After summarization, this becomes:

[
    HumanMessage(...),
    AIMessage(tool_calls=[...]),           # ← AI message with tool calls
    ToolMessage(tool_call_id="call_1"),    # ← Valid tool response  
    ToolMessage(tool_call_id="call_2"),    # ← Valid tool response
    ToolMessage(tool_call_id="call_1"),    # ← DUPLICATE - no preceding AI message!
    ToolMessage(tool_call_id="call_2"),    # ← DUPLICATE - no preceding AI message!
]

The duplicated tool messages at the end violate the chat completion API requirements because they don't immediately follow an AI message with the corresponding tool calls.

Likely Problem

Looking at the _preprocess_messages function around line 197-214, the issue is in this logic:

# If the last message is an AI message with tool calls,
# include subsequent corresponding tool messages in the summary as well,
# to avoid issues w/ the LLM provider
if (
    messages_to_summarize
    and isinstance(messages_to_summarize[-1], AIMessage)
    and (tool_calls := messages_to_summarize[-1].tool_calls)
):
    # Add any matching tool messages from our dictionary
    for tool_call in tool_calls:
        if tool_call["id"] in tool_call_id_to_tool_message:
            tool_message = tool_call_id_to_tool_message[tool_call["id"]]
            messages_to_summarize.append(tool_message)

The function builds tool_call_id_to_tool_message from ALL messages, then when it finds an AI message with tool calls at the boundary, it appends the corresponding tool messages - even if they're already part of messages_to_summarize. This creates invalid message sequences that LLM providers reject.

Suggested Fix

Check if tool messages are already included before adding them:

if (
    messages_to_summarize
    and isinstance(messages_to_summarize[-1], AIMessage)
    and (tool_calls := messages_to_summarize[-1].tool_calls)
):
    # Get IDs of tool messages already in messages_to_summarize
    existing_tool_call_ids = {
        msg.tool_call_id for msg in messages_to_summarize 
        if isinstance(msg, ToolMessage) and msg.tool_call_id
    }
    
    # Only add tool messages that aren't already included
    for tool_call in tool_calls:
        tool_call_id = tool_call["id"]
        if (tool_call_id in tool_call_id_to_tool_message and 
            tool_call_id not in existing_tool_call_ids):
            tool_message = tool_call_id_to_tool_message[tool_call_id]
            messages_to_summarize.append(tool_message)

Environment

• LangMem version: 0.0.27
• Using with LangGraph ReAct agents
• Azure OpenAI as the LLM provider
• Standard tool call/response pattern

---

Please let me know if this is a valid issue - I'd be happy to contribute.

[yash025]

@hinthornw @vbarda can you please take a look at this issue?

[dominikriedermbw]

There is another side effect to this bug. When the last message in the messages_to_summarize is a ToolMessage, it does not enter this section:

    if (
        messages_to_summarize
        and isinstance(messages_to_summarize[-1], AIMessage)
        and (tool_calls := messages_to_summarize[-1].tool_calls)
    ):
        # Add any matching tool messages from our dictionary
        for tool_call in tool_calls:
            if tool_call["id"] in tool_call_id_to_tool_message:
                tool_message = tool_call_id_to_tool_message[tool_call["id"]]
                n_tokens_to_summarize += token_counter([tool_message])
                messages_to_summarize.append(tool_message)

But you can have cases, where you have actually two or more ToolMessages (caused by one AI Message) and you only end up with one of the ToolMessages in the messages_to_summarize.

This also leads to the following error:

from None\nopenai.BadRequestError: Error code: 400 - {\'error\': {\'message\': "An assistant message with \'tool_calls\' must be followed by tool messages responding to each \'tool_call_id\'. The following tool_call_ids did not have response messages: call_Wf1PNyuxkzzmpgSzKzS0JfFM", \'type\': \'invalid_request_error\', \'param\': \'messages.[5].role\', \'code\': None}}\n'}]

In my opinion you should also check, whether the last message in messages_to_summarize is a ToolMessage and then check, whether you contain all the other ToolMessages (if the respective AI message has multiple tool calls)

[yash025]

@hinthornw @vbarda @dominikriedermbw please review this PR: #100

[yash025]

@hinthornw @vbarda can you please assist here?

[yash025]

hi @hinthornw @vbarda can you guys at least reply? is this project not maintained anymore?

[yash025]

have provided the fix already, just review it and see if it makes sense or not.