Facing issues with passing memory checkpointer in lanngraph agents

[kishan_]

Hi,

I am trying to create a simple langgraph agent in Databricks, the agent also uses lanngraph memory checkpoint which enables to store the state of the graph. This is working fine when I am trying it in Databricks notebook, but when I tried to log this as Mlfow model, I couldn't find any documentation on where to pass the memory checkpoint

I want to know if there is way we can pass memory checkpoint when logging langgraph agent in Mlflow

[morenoj11]

Facing the same issue here.

[morenoj11 · Accepted solution]

I saw that you can compile the model without checkpointer, register it in MLflow, and then, after loading, assign it after compilation.

```

import mlflow

mlflow.models.set_model(build_graph())

with mlflow.start_run() as run_id:

model_info = mlflow.langchain.log_model(

lc_model="build_graph.py", # Path to our model Python file

artifact_path="langgraph",

)

model_uri = model_info.model_uri

[...]

loaded_model = mlflow.langchain.load_model(model_uri)

loaded_model.checkpointer = checkpointer

loaded_model.invoke(input_state, config)  ## config has the thread_id

```

It's not elegant or future-proof, but it might do the trick while we wait for a better solution.

[kishan_]

@morenoj11 The solution which you have mentioned, Have you tried to deploy the same in Databricks model serving ?

[sebascardonal]

Hi all. I have the same issue, could you deploy the graph with the MemorySaver?