Recently in deep learning, there's been a surge in learning how to use memories as part of the optimisation process (i.e. LSTM's and Stacks). However, these aren't really analogous to how a cognitive systems learns how to use it's working memory.
Are there models of how working memory modules (where a saved value decays over time) can be learned to be optimally leveraged? Either via reinforcement learning or supervised learning?
ACT-R is a complete cognitive model that incorporates Working Memory, Declarative Memory and Procedural Memory, but also incorporates input (visual and auditory) and output (manual) buffers. It is a really interesting model about the human in its entirety but, it is at a high level of abstraction.
The paper about memory you want is "REFLECTIONS OF THE ENVIRONMENT IN MEMORY" by Schooler and Anderson. It describes how and why memory is like Reïnforcement learning. In this paper you can read how memory is used for studying. About learning to leverage the working memory, I don't exactly know what you mean, but I think these papers give you a nice headstart.
If you want to learn more about ACT-R, I would recommend this book, or just read the many papers they have published over the years. They have great examples that explain multitasking and time perception among others.
I could only find a single instance of learning to leverage a memory. "The Origin of Epistemic Structures and Proto-Representations" by Chandrasekharan and Stewart shows how to include the option to save the current state (by training a neural network to output '1' given the current state input) and input this memory into the state representation of a reinforcement learning task. In the paper, this method allows an agent to switch between foraging and returning home.
Frustratingly, checking the citations of this paper leads to a dead end, so further references would be greatly appreciated.
