How do I add memory when using createToolCallingAgent?

[canbaoSama]

Checked other resources

• I added a very descriptive title to this question.
• I searched the LangChain documentation with the integrated search.
• I used the GitHub search to find a similar question and didn't find it.

Commit to Help

• I commit to help with one of those options 👆

Example Code

// Definition
const memory = new BufferMemory({
    chatHistory: new ChatMessageHistory(),
    memoryKey: "chat_history",
    returnMessages: true,
    outputKey: 'output'
});
const prompt = ChatPromptTemplate.fromMessages(
    [
        ['system', `你是一个高效的工具助手，可以调用工具来完成用户的请求,也可以储存对话记录，用来回答用户问题`],
        ['system', '可查询的会话记录：{{chat_history}}'],
        ['system', '可用的工具：{{tools}}'],
        ['human', '{input}'],
        ['system', '{agent_scratchpad}'],
    ],
);
let executor: AgentExecutor

// call
if (!executor) {
                const agent = await createToolCallingAgent({ llm: createChatLLM(LLM_TYPE[this.params.chatType || 'QWEN_3']), tools: this.tools, prompt, streamRunnable: true });
                executor = new AgentExecutor({ agent, tools: this.tools, maxIterations: 50, memory });
            }
            const langChainStream = await executor.stream({
                input: this.params.text,
            });

            for await (const step of langChainStream) {
                this.ctx.state.ws.send(JSON.stringify({ code: 0, topic: 'aiAssistant', data: step }));
            }

Description

I have tried using a local model, and chat_history can record the conversation content, but the agent cannot correctly answer questions based on the historical conversation records.

System Info

"@langchain/core": "0.3.53",
"@langchain/mcp-adapters": "0.4.2",
"@langchain/ollama": "0.2.0",
"langchain": "0.3.24",

[dosubot[bot]]

Hi @canbaoSama! I'm Dosu and I’m helping the langchainjs team.

To add memory when using createToolCallingAgent, you need to ensure that your prompt includes a placeholder for the memory variable (e.g., chat_history), and that your memory instance (such as BufferMemory) is configured to use the same key. Your example code is mostly correct in this regard.

However, for the agent to correctly answer questions based on historical conversation records, you must ensure that the memory is properly injected into the agent's execution flow and that the chat history is updated after each interaction. The agent itself does not manage memory; this is handled by the executor or by explicitly passing the chat history as input.

Here are the key points for effective memory integration:

• Your prompt must have a placeholder for chat_history (which you have).
• The BufferMemory instance should use the same memoryKey as the prompt (which you have: "chat_history").
• After each agent run, you need to update the chat history with both the user's input and the agent's output. This ensures that the next invocation has the full context. If you use AgentExecutor, it will handle this for you if configured correctly.
• If you are streaming or managing the agent manually, you may need to append new HumanMessage and AIMessage objects to your ChatMessageHistory after each turn, so that the memory reflects the latest conversation state.

For reference, the LangChain docs show that memory integration requires you to track the inputs and outputs as chat history, and to pass this history into the agent on each call. See the official example for adding memory to a custom agent.

If your agent is not using the historical conversation records, double-check that:

• The chat_history variable is being populated and passed into the agent on each call.
• The memory is not being reset or recreated between turns.
• The prompt is referencing chat_history in a way that the model can use (e.g., not buried in a system message that the model ignores).

There are no known issues with memory integration in the versions you listed, but effective use of memory with createToolCallingAgent requires careful orchestration at the prompt and executor level, as the agent itself is stateless and expects the memory context to be provided externally. For more details on how BufferMemory works, see the memory module documentation.

If you are still having trouble, consider logging the contents of your chat_history before each agent call to ensure it contains the expected messages. If you are using tool calls, make sure each tool call is followed by a corresponding ToolMessage in the chat history, as discussed in this thread.

Let me know if you need a more concrete code example or have questions about a specific part of your setup!

To reply, just mention my name @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}