Add checkpoint to MemorySaver

[cocoza4]

Hi all,

How can I manually add checkpoints into MemorySaver?
This is how i did it

from langgraph.checkpoint.base import empty_checkpoint
from langchain_core.messages import BaseMessage
from langgraph.checkpoint.memory import  CheckpointMetadata

memory = MemorySaver()


async def add_memory(history: List[BaseMessage], config):
    ckp = empty_checkpoint()
    ckp["channel_values"]["messages"] = history

    await memory.aput(config, checkpoint=ckp, metadata=CheckpointMetadata(source="input", step=-1), new_versions=None)

But LangGraph still didn't pick the memory correctly. Any ideas?

[hwchase17]

Will let @vbarda respond, but OOC why do you want to do this?

[vbarda]

hi! i would recommend against calling .put / .aput methods on the checkpointers -- they are meant only to be used by langgraph library internally

is your intention to include the history together with the new message you send to the graph? or do you just need to retain the history before the new message is sent? for the former, you can just invoke the graph with the same thread ID and the message history should be included automatically. for the latter, if you need to be able to reference the state at a previous point in time, you can just use graph.get_state_history(thread_config). not sure if that answers your question

also, i am not sure i understand why you're using MemorySaver if you want to save the results to Firebase

[cocoza4]

@vbarda thanks for ur response. My intention is the latter which is for the LangGraph to be able to refer the chat history
For example,

Question: show me campaign with highest impressions?
LangGraph answer: campaign xxx shows the highest impressions of 5K
New question: show me clicks of this campaign in the last 30 days?

In this scenario, LangGraph should be able to refer this campaign to be campaign xxx. Hope this clarify the question.

[vbarda]

this should already work as you described provided that you have defined a reducer for the messages key in your state correctly (provided the agent also has ability to do tool calling / some other way to search for the campaign information). you should then just invoke the graph with the same thread ID both for the original question and the new one. hope this helps!

see more in these docs:
https://langchain-ai.github.io/langgraph/how-tos/persistence/
https://langchain-ai.github.io/langgraph/concepts/low_level/#reducers

[cocoza4]

@vbarda, thanks for your response. I have come up with a workaround without using MemorySaver at all, here's a shortened code snippet

from langgraph.prebuilt.chat_agent_executor import AgentState, add_messages

class HistoryAwareAgentState(AgentState):
    chat_history: Annotated[list[BaseMessage], add_messages]


prompt = ChatPromptTemplate.from_messages(
    [
        ("system", system_message),
        MessagesPlaceholder("chat_history"),
        ("human", "{messages}"),
    ]
)

agent = create_react_agent(
    llm, tools=tool_node, state_schema=HistoryAwareAgentState, state_modifier=prompt
)

chat_history = query_from_firebase(xxx)

question = "show me clicks of this campaign in the last 30 days?"
message_str = f"""
Strictly follow the instructions below to answer the question.
If relevant, refer to the previous messages for additional context e.g. campaign, ads information, etc.
Only answer the question below, no need to answer for previous questions.
If the question has been answered, no need to run further.

instructions
....

Question: {question}
Answer:"""
message = [HumanMessage(message_str )

response = await agent.ainvoke(
        {
            "messages": [message ],
            "chat_history": chat_history,
        },
        config=config,
    )

Basically, I added a dedicated chat_history variable. What do you think about this approach? i tested it and it seems to work.

[cocoza4]

@hwchase17 @vbarda hi, may i follow up on my previous question?