langgraph/how-tos/memory/manage-conversation-history/

[giscus[bot]]

langgraph/how-tos/memory/manage-conversation-history/

Build language agents as graphs

https://langchain-ai.github.io/langgraph/how-tos/memory/manage-conversation-history/

[jojohannsen]

Thanks for making all these great How-to guides.

In this one, I'm seeing a problem with that very last "what's my name?" -- the error is

anthropic.BadRequestError: Error code: 400 - {'type': 'error', 'error': {'type': 'invalid_request_error', 'message': 'messages: first message must use the "user" role'}}

https://smith.langchain.com/o/89f6cd25-e2fd-498b-82ab-058e41888172/projects/p/e52b5fcf-4a6b-4663-8034-e4f7f4c20ed7?timeModel=%7B%22duration%22%3A%227d%22%7D

It doesn't make sense, since the first message actually is a HumanMessage. Also, if I add a second HumanMessage, then everything is fine (that's the successful one in the langsmith project)

[hwchase17]

can you share a public version of your langsmith trace? dont think i can see that

[jojohannsen]

https://smith.langchain.com/public/866fa557-063c-4e56-94f2-9712d7a0caba/r is the one with the error

https://smith.langchain.com/public/866fa557-063c-4e56-94f2-9712d7a0caba/r is the exact same one, but with one additional human message, no error on this one

[jojohannsen]

https://smith.langchain.com/public/91ca3b3e-ffc6-4812-8883-7c0c9becfadd/r this is the no-error one (above was error link twice...)

[vbarda]

ah there was a mistake in the docs, updating! it should be messages[-1:] in the filter_messages

[archit-spec]

not able to access these traces

[antoremin]

Consider the following use case:

I'm working on a research graph. The graph:

1. takes in a request, plans a series of research actions (search, etc),
2. runs these actions (usually 5-15 at the time)
3. reviews results, compiles a search results document
4. Reviews the request, the plan, results, checks if done, goes to the plan node once more.

What would be the best way to manage trimming messages for history? I'm dealing with two issues:

1. Tool calls get trimmed and # of tool calls and tool results don't match so model calls fall over.
2. Message history overflows, but summarization doesn't deliver enough context granularity.
   Different nodes would benefit from certain parts of previous messages (e.g search results compiler should see all ToolMessage items from the previous run, whereas the planner should probably see previous HumanMessages and maybe some other select AIMessages)

The first issue seems like something that should have a solution on Langchain side.
The second is something that I think could be added for the benefit of many real world use cases: how do you support longer-running langgraphs that need to deal with granular information.

[eyurtsev]

For (1), it really depends on what you're doing with the chat history downstream. If you want to feed it directly into a chat model, then the chat history must meet these criteria to be considered valid by most chat model providers:

1. Start on system message or human message
2. Ends on human message or tool message
3. Any tool message only appears after a corresponding ai message

To achieve this you can specify the following with trim_messages:

• start_on="human",
• end_on=("human", "tool"),

Note that you will probably also want to include the system message (include_system=True).

This configuration will handle the simplest manipulation of the conversation history. If you need something more complex, you can use trim_messages as a building block.

For (2), we added this tutorial: https://langchain-ai.github.io/langgraph/tutorials/memory/long_term_memory_agent/. Things can get complex for more complex use cases.

[edwinow]

Thanks Eugene for the response, helpful to know about trim_messages!

For others, here's updated link to the latest docs for long term memory and there are a couple of really helpful new videos that the team have put out too on their youtube channel on the topic.

[ghost]

is there a way to implement this in the pre-built react agent? do we modify the code of the agent?

[Maheshbabu9199]

Hello,

I am trying to create an agent using create_react_agent. I am creating a memory using MemorySaver() class. But I am not able to find out how can I add custom messages to the agent's checkpointer. I've tried using aput_writes of MemorySaver class, but it seems the agent is not using the memory. When I include these custom messages in the prompt, the agent is giving me expected answer around the messages but from checkpointer it is not retrieving that information.

Can anyone help me in this ??

Thanks!!