Hippocampus

The hippocampus is a seahorse-shaped subcortical structure in the medial temporal lobe that serves as the brain's fast-learning episodic memory system — a biological instance of sparse, high-fidelity recording that enables rapid acquisition of individual experiences while preserving their distinctness. It is not merely a storage device but a computational architecture: a network optimized for pattern separation (making similar experiences distinct), pattern completion (recalling complete memories from partial cues), and sequence generation (constructing trajectories through remembered and imagined space).

The hippocampus occupies a critical node in the brain's memory hierarchy. In the Complementary Learning Systems framework, it functions as the fast learner paired with the slow-learning neocortex. Where the neocortex extracts statistical regularities through overlapping, distributed representations, the hippocampus records specific episodes via sparse, minimally overlapping representations in the dentate gyrus and CA3 subfields. This architectural separation solves the stability-plasticity dilemma that plagues continual learning systems: a single network cannot simultaneously learn quickly from new examples and preserve old knowledge. The hippocampus sidesteps this by accepting rapid overwriting of its own contents, knowing that the neocortex will eventually consolidate what matters through memory replay during sleep.

Structure and Computation

The hippocampus is anatomically layered — dentate gyrus, CA3, CA1, and subiculum — and each layer implements a distinct computational transformation. The dentate gyrus performs pattern separation via sparse coding: similar inputs are mapped to nearly orthogonal representations by a vast population of granule cells. CA3 performs autoassociative pattern completion through recurrent collateral connections, enabling the retrieval of complete memory traces from degraded cues. CA1 compares the output of CA3 with direct cortical input, producing a match/mismatch signal that may drive learning or novelty detection.

This layered architecture has direct parallels in machine learning. The dentate gyrus resembles a random projection layer or an expansion layer in a neural network — increasing dimensionality to achieve linear separability. CA3's recurrent dynamics resemble Hopfield networks or modern recurrent memory mechanisms. The hippocampus is, in effect, a biological solution to problems that artificial systems are still struggling to solve: continual learning without catastrophic forgetting, one-shot encoding of novel experiences, and structured generalization from sparse data.

Spatial Cognition and Beyond

The discovery of place cells in the rodent hippocampus — neurons that fire at specific spatial locations — initially framed the structure as a cognitive map dedicated to navigation. But subsequent work has shown that the same cells code for temporal relationships, auditory sequences, social configurations, and abstract conceptual spaces. The hippocampus maintains a generalized representation of structured