Agent with Checkpointer Reusing Tool Results in Continuing Conversations

OSS Product Help LangChain
1

Hi, I’m using LangChain v1 (TypeScript) with createAgent and MongoDBSaver checkpointer for a RAG application. I’m experiencing behaviour where tool results seem to be reused across conversation turns instead of making fresh tool calls.

Setup:

  • LangChain v1.x (TypeScript)

  • createAgent with MongoDBSaver checkpointer

  • Custom retrieval tool with parameters captured in closure (e.g., groupIds)

  • Multi-turn conversations

Observed Behaviour:

  • Request 1 (new conversation):

    • question: “When was Company1 formed?”
    • groupIds: [“681…”, “684…”]
    • //Works - retrieval tool called, correct results

  • Request 2 (continuing conversation):

    • question: “What is Company2 fiscal year?”
    • conversationId: “69123…” // same conversation
    • groupIds: [“681…”, “684…”] // SAME groupIds

    // Returns wrong results - citations show Company1 docs from Request 1

What I Expected:

  • Agent loads conversation history (Q&A about Company1)

  • Agent sees NEW question about Company2

  • Agent calls the retrieval tool with a fresh query

  • Tool performs a new vector search (same groupIds, different semantic query)

  • Returns Company2 documents

What Actually Happens:

  • Agent returns “not found” or Company1 information

  • Citations show the SAME documents from Request 1

  • Tool is either not called or returns cached results

What I’ve Tried:

  1. Middleware to filter ToolMessages:

    export const filterToolMessagesMiddleware = createMiddleware({

    name: 'FilterToolMessages',

    beforeModel: async (state: { messages: BaseMessage[] }) => {

        const filtered = state.messages.filter(msg => 

            !(msg instanceof ToolMessage) && 

            !(msg instanceof AIMessage && msg.tool_calls?.length > 0)

        )

        return { messages: filtered }

    }

})

const agent = createAgent({

    model: llm,

    tools: [retrievalTool],

    checkpointer,

    middleware: [filterToolMessagesMiddleware]

})
  • Result: Still reuses old results

  1. Manual checkpoint management (current workaround):

    // Load checkpoint manually

const existingState = await checkpointer.get(threadConfig)

const existingMessages = existingState?.channel_values?.messages || []

// Filter out tool messages

const conversationHistory = existingMessages.filter(msg => 

    !(msg instanceof ToolMessage) && 

    !(msg instanceof AIMessage && msg.tool_calls?.length > 0)

)

// Create agent WITHOUT checkpointer

const agent = createAgent({

    model: llm,

    tools: [retrievalTool]

    // NO checkpointer

})

// Invoke with filtered history

const result = await agent.invoke({

    messages: [...conversationHistory, new HumanMessage(question)]

})

// Save manually

await checkpointer.put(threadConfig, {

    channel_values: { messages: result.messages }

})

Result: Works - fresh tool calls every time

Questions:

  1. Is this expected behaviour when using createAgent with a checkpointer?

  2. Should middleware be sufficient to prevent tool result reuse?

  3. Is manual checkpoint management the recommended approach for this use case?

  4. Am I missing something in my implementation?

Any guidance would be greatly appreciated!

Environment:

  • LangChain: v1.x (TypeScript)

  • Node.js: v24

  • Checkpointer: MongoDBSaver

2

hi @juancarlos_la

this is actually very weird imho. I would expect that either the LLM calls the tool per unique query or the framework reuses the response for the same query.

What LLM model are you using?

3

I’m using anthropic.claude-3-sonnet-20240229-v1:0.

I’ve created a bug for it here, with more detailed information: Agent with Checkpointer Reuses Tool Results in Multi-Turn Conversations · Issue #9384 · langchain-ai/langchainjs · GitHub

4

great, thanks!
btw this is for python, not javascript, repo Agent with Checkpointer Reuses Tool Results in Multi-Turn Conversations · Issue #33936 · langchain-ai/langchain · GitHub - it should not be there