Short-term memory
Short-term memory is the temporary retention of information over seconds to minutes — the liminal space between the immediate present of sensory experience and the enduring archive of long-term storage. But the term itself is misleading. Short-term memory is not a memory at all, in the archival sense. It is better understood as an active holding pattern, a dynamically maintained state that persists only so long as the system continues to sustain it.
The folk model treats short-term memory as a storage bin: information enters, sits for a while, and either decays or is transferred to long-term memory. This model, inherited from the Atkinson-Shiffrin memory model, is not wrong so much as it is the wrong level of description. The brain does not store information in short-term memory the way a computer stores data in RAM. It maintains activation patterns through recurrent neural dynamics, and the persistence of the pattern is the persistence of the memory. When the dynamics stop, the memory does not decay — it ceases to exist.
Capacity and the Magical Number
The classic finding is that short-term memory holds approximately seven items, plus or minus two — the "magical number" identified by George Miller in 1956. But this is not a hardware limit. It is a span limit: the number of items that can be maintained in a simultaneously active state before the activation patterns interfere with each other. The limit is not storage space; it is dynamic stability.
The limit can be stretched through chunking — grouping items into larger units that function as single activation patterns. A chess master does not remember the positions of thirty-two pieces; they remember the position as a configuration, a single gestalt that compresses the information into a stable attractor. This reveals that short-term memory is not about holding items but about holding structure.
From Short-Term Memory to Working Memory
The shift from the term "short-term memory" to "working memory" in the 1970s and 1980s was not merely a rebranding. It reflected a reconceptualization: what had been treated as a passive store was reconceived as an active workspace. The Baddeley-Hitch model posited a central executive that manipulates information held in slave systems — the phonological loop and the visuospatial sketchpad. This is closer to the truth, but it still assumes a controller-and-storage architecture that the brain does not implement.
A better model is that short-term memory and working memory are the same process viewed at different timescales. Short-term memory is the transient activation of neural populations; working memory is the sustained, task-relevant activation of those populations under top-down modulation. The difference is not architectural but dynamical: the same network can be in a passive echo state (short-term memory) or an attractor state stabilized by feedback (working memory). The distinction between storage and processing dissolves.
Decay, Interference, and the Problem of Time
The central question about short-term memory is not where information goes but why it leaves. Two accounts dominate: decay and interference. Decay theory holds that memory traces fade over time, like footprints in sand. Interference theory holds that new information overwrites old information, like a palimpsest.
Both are wrong at the neural level. Memory traces do not fade; they are actively disrupted by the dynamics of the network itself. The same recurrent connections that maintain a pattern will, when driven by new input, settle into a new attractor state. The old state is not erased; it is destabilized. And this destabilization is not a bug. It is a feature. A system that could maintain every previous activation pattern would be a system that could never change its mind. Short-term memory forgets because the brain must think, and thinking is the displacement of one stable pattern by another.
Connection to Systems Theory
Short-term memory is a paradigmatic case of a system operating at a specific timescale. The systems-theoretic perspective distinguishes between processes that are fast (sensory transduction), intermediate (short-term memory), and slow (long-term memory consolidation). Each level operates with its own dynamics, and the coupling between levels is what produces coherent behavior.
The short-term memory system is also a feedback system. It is maintained by re-entrant activity — loops of neural firing that sustain themselves through recurrent connections. This is not storage; it is a self-sustaining dynamical process. The system is in a state of homeostatic regulation: it maintains a pattern against perturbation until the pattern is no longer needed, at which point the regulation stops and the system settles to a new state.
The failure modes of short-term memory — distraction, absent-mindedness, the tip-of-the-tongue phenomenon — are not failures of storage retrieval. They are failures of dynamical maintenance: the attractor was not deep enough, the feedback was interrupted, or a competing attractor was stronger. The correct intervention is not better storage but stronger maintenance — deeper attractors, more stable dynamics.
The concept of short-term memory as a storage buffer is the last refuge of the computational metaphor in cognitive psychology. The brain does not have a short-term memory store. It has transient dynamical states that we, as observers, call memories. The difference is not semantic. It is the difference between a warehouse and a river.