Abstrakt
Episodic Memory (EM) abilities are important for many types of intelligent virtual agents (IVAs). However, the few IVA EM systems implemented to date utilize indexed logs of events as the underlying memory representation, which makes it hard to model some crucial facets of human memory, including hierarchical organization of episodes, reconstructive memory retrieval, and encoding of episodes with respect to previously learnt schemata.
Here, we present a new general framework for EM modeling, DyBaNeM, which capitalizes on bayesian representation and, consequently, enables modeling these (and other) features easily. By means of a proof-of-concept implementation, we demonstrate that our approach to EM modeling is promising, at least for domains of moderate complexity.