What are the use-cases for adding a common checkpointer to subgraphs in a multi-agent system?

[thisisarko]

I'm trying to understand the scenarios where adding the same memory checkpointer from a parent graph to a subgraph in a multi-agent system could be useful.

In my experiments, the StateSnapshot from the get_state() method appears identical for both the subgraph and the parent graph after a super-step.

It seems useful to add a different memory checkpointer, like a vector store, to a subgraph that requires semantic search over conversation history. However, is sharing the same memory checkpointer redundant?

Alternatively, could it be necessary for a subgraph to resume execution from a halt or error, even if the parent has the snapshot?

[hinthornw]

It's useful for the same reason checkpointers are useful for any graph! Multi-turn interactions, and interrupts.

For multi-turn interactions, if you have state in the subgraph you want to persist between turns but that isn't accessed by the parent graph, then you'd need it checkpointed.

If you have an interrupt in the subgraph, then you'd need the state persisted in order to resume from that location.