With the increasing availability of Web services and adoption of services oriented paradigm, there is a growing need to dynamically compose services for realizing complex user tasks. While service composition is itself an important problem, a key issue is also how to support users in selecting the most appropriate compositions of services to fulfill a task. In existing dynamic services selection approaches, combinations of services are repeatedly discovered (e.g., using ontology-based matching techniques) and selected by users whenever needed. To improve their effectiveness, we propose a new technique that provides an efficient access to what is named a “task memory”. A task memory is used to provide users with a context-aware service selection by recommending combinations of services that are most appropriate in a given context. A task memory is formed using the service composition history and their metadata. We present an incremental approach for building the task memory in which we monitor how users use and rank the services. The continuous updates of the task memory over time will result in more fine-tuned recommendations for composite services.