Over the past two decades, numerous experiments have shown that spoken word perception creates detailed memory traces, containing not only word meanings, but also extraneous, perceptual or contextual details. This is shown, for example, by voice-specific priming effects. Based on such results, exemplar theories suggest the mental lexicon may consist of accumulated episodic traces. Although an episodic approach is well-suited to explain priming, ample evidence suggests that language also entails abstract representations. Certainly at the segmental level, there are logical constraints that require unitization. An optimal theory may include stable abstract representations, combined with context-sensitive episodic traces. This paper summarizes new tests examining word perception from a complementary-systems perspective, wherein reciprocal neural networks represent hippocampal and cortical memory systems. In this approach, detailed episodic traces and holographic, abstract traces combine to create behavior in real-time, allowing perceptual or memorial data to appear more or less episodic, depending on myriad factors. I summarize new results and simulations on perceptual priming, and discuss the model with respect to perceptual learning in speech.