MINERVA 2, developed by Douglas Hintzman (1984, 1986), is a multiple-trace memory model in which every encounter with an item lays down a distinct trace in long-term memory. Unlike models that store abstractions or prototypes, MINERVA 2 keeps raw episodic records and derives categorical and schematic knowledge at retrieval time through the emergent behavior of summing across many traces.
Architecture and Encoding
Each memory trace is a vector of features, where each feature takes a value of +1, −1, or 0 (unknown/irrelevant). When an event is encoded, each feature of the original stimulus is copied to the trace with probability L (the learning rate); features not copied are stored as 0. This probabilistic encoding means that individual traces are noisy, incomplete records of the original experience.
Retrieval: The Echo
Retrieval begins with a probe vector that is compared simultaneously to all traces in memory. The similarity between the probe P and trace i is computed as a normalized dot product, then raised to the third power to produce an activation value:
The cubic exponent is crucial: it sharpens the activation distribution so that highly similar traces dominate the response while weakly similar traces contribute negligibly. This nonlinearity allows the system to behave as though it retrieves specific episodes even though all traces participate in every retrieval.
Echo Intensity and Echo Content
Two quantities are computed from the activated traces. Echo intensity is the sum of all activations, I = Σᵢ A(i), and serves as a familiarity or recognition signal. Echo content is a weighted sum of all traces, C(j) = Σᵢ A(i)·T(i,j), and serves as a reconstructed representation of the retrieved information. Because traces from the same category share features, the echo content for a category probe will approximate the category prototype even though no prototype was ever stored.
Empirical Scope
MINERVA 2 accounts for a wide range of memory phenomena including frequency judgments, recognition memory, the prototype abstraction effect, schema formation, and the power law relating echo intensity to actual presentation frequency. Hintzman (1988) extended the model to show how abstract category knowledge could emerge from purely episodic storage, challenging the then-dominant assumption that semantic and episodic memory required separate systems.
The cubic activation function ensures that the echo is dominated by traces very similar to the probe. Without the nonlinearity, moderate-similarity traces would swamp the signal. The exponent of 3 was chosen because it provided the best fit to human frequency-judgment data, but the model works with any odd exponent ≥ 3.