Checkpoint team while it is running

[asheeshgarg]

How to add a memory checkpoint as the system is running. I know we can do team save state. But how to persist for each agent communication so we can rerun at any point? So that system is recorded all the messge as it’s happening not to save as decided by user.

[EItanya]

I agree that a checkpointing system which allowed for easy re-running could be very very beneficial. I think potentially wrapping State types with a wrapper which saves them in an indexed way. Along with team options to save often could do the trick.

[asheeshgarg]

Indexing is really a important feature so we be able to move between different runs or between the runs of agent. After optimizing some setting like. a prompt and rerun to see the impact.

[rysweet]

In the current framework this is an exercise for the developer to determine when to call the state apis. If you write your agents as idempotent actors, you should be able to externalize the state and create flows that are easy to checkpoint/rehydrate. This is much more manageable in the core api than in agent chat.

[asheeshgarg]

Not sure if its a good idea to tie runtime to a optional Memory component like tracing we have. That can be passed to agent for checkpointing purposes. Every team run is indexed on its own run.
I tried using tracing for recreating agent state but it has lot of parsing required

[ekzhu]

One idea is to emit CheckpointEvent when running a team like SelectorGroupChat. The team decides on the frequency of checkpoints it is going to make. Each time a checkpoint event appears, the state upon the event exports a snapshot and can be accessed via a reference from the CheckpointEvent. This way, you can resume the group chat from any previous checkpoint by simply running run() again without an argument.

@asheeshgarg @EItanya @rysweet @gagb @husseinmozannar