When an asset returns a value, an IO manager stores it; when another asset takes it as a parameter, the IO manager loads it back. The default, registered under the key io_manager, writes files under the instance's storage directory, one per asset (and per partition).
To store values somewhere else, subclass dg.ConfigurableIOManager and implement two methods: handle_output(self, context, obj) and load_input(self, context). Both contexts carry context.asset_key and, for partitioned assets, context.partition_key, which you use to build the path. Register it as io_manager to change the default, or under another key and point specific assets at it with io_manager_key="lake_io".
Two cases skip the IO manager: an asset annotated -> None, and an asset that returns dg.MaterializeResult. Both mean "this asset writes its own data", so a downstream asset must use deps rather than a parameter to depend on it.
Values stored as JSON text in a lake, one file per asset.