Linked Data applications will typically need to discover and load data and schemas from the web. A user may request that extra data is displayed from specific locations, and the loading of a new data source may trigger loading of additional schemas, e.g. to discover labels for properties and types, or to pass to a reasoner for infering additional data and relationships. Some resources and vocabularies may be very commonly used, e.g. the RDF, RDF Schema and OWL vocabularies, while others may only be encountered during run-time.

Build a local cache of retrieved resources, refreshing the cache only when source data has changed.

Retrieving resources from the web, like any other network access, is prone to failure. Repeated fetching of the same resources is waste-ful of bandwidth on the client and the server: a large number of clients can easily overload resources on a system serving up popular vocabularies.

Applications should cache remote resources wherever possible. The cache may be handled entirely in-memory but with sufficient permissions and access to the local file-system an application could also build a persistent cache. Desktop applications may ship with a pre-seeded cache of commonly retrieved resources such as ontologies. Efficient use of HTTP request can ensure that cached versions need only be updated when the remote copy of the resource has been updated. Certain vocabularies, e.g. RDF Schema, will only change rarely, if at all. These could be cached for longer periods, if not permanently.

Ideally applications should provide configuration to support the user in managing the amount of local resources (memory or disk space) that can be used by the cache. Control over the location in which cached data will be stored is also useful.