Book excerpt: All mental function - perception, cognition or action - is, ultimately, based on memory and its appropriate recall. The question of how the brain learns memories and their associations has been one of the most debated and interesting topics of research by cognitive psychologists and neuroscientists. Experiments show that human memory is largely associative, working through association between entities seen in the real world. In such a memory, recall occurs when the system is presented with a partial cue for the content, which then triggers the rest of the memory by association. A very important aspect of human memory is that associations are highly context-dependent, allowing the brain to represent a very large body of knowledge concisely through different combinations of a relatively small set of memories. This can also be seen as the basis of the brain's ability to generate novel ideas. The most successful computational models for associative memory are based on attractor dynamics in recurrent neural networks. However, relatively little work has been done on context-dependent associative memory. This thesis presents modular recurrent neural network architecture for context-dependent associative memory (CDAM) that provides a natural and biologically plausible way to represent context-dependence in the network dynamics. The simulations shown indicate that this system is capable of both exploitation (recalling previously learned memories in a context-appropriate way) and exploration (generating novel but sensible "memories" that can be seen as representing new ideas.) The role of noise in both exploitation and exploration is an especially intriguing issue, and is investigated systematically in the thesis.