musWM is not only a harmonic analyser. It was designed as a computational model of musical working memory: a system that decides which sounding events remain jointly available long enough to be heard as a chord, and when that material has become stable enough to be named. The design takes its functional constraints from the multicomponent model of working memory introduced by Baddeley and Hitch (1974) and extended with the episodic buffer (Baddeley, 2000), and from sensory accounts of rhythm and meter. The argument is developed in full in the author’s thesis (Tuğral, 2026); this section summarises it.

From a single store to a multicomponent system

The modal model of Atkinson and Shiffrin (1968) treated short-term memory as one capacity-limited store between brief sensory registers and long-term memory: holding information and operating on it drew on the same resource. Baddeley and Hitch (1974) put that assumption to the test with dual-task experiments. Participants could keep a few items in mind while reasoning, with only modest slowing — retention and higher-order processing were not competing for a single pool. Working memory was therefore recast as an active system of distinct parts: a central executive that controls attention, modality-specific slave systems (the phonological loop and the visuospatial sketchpad) and, from 2000, an episodic buffer that binds information from these systems and from long-term memory into integrated, multidimensional episodes available to awareness (Baddeley, 2000, 2012; Baddeley, Hitch & Allen, 2019).

Figure 1. Musical working memory following Baddeley’s core principles. Sensory input from the outer world is registered in pre-attentive memory, held as short-lived modality-specific traces, filtered into slave systems, and bound in the episodic buffer, which exchanges information with long-term memory under the attentional control of the central executive. Reproduced from Tuğral (2026).

Why music needs a different balance

Baddeley’s model was built around language and visuospatial reasoning, where units are discrete and quickly categorised: a word can be recognised and then rehearsed. Musical events rarely work that way. A single note is not yet a harmony; a chord acquires its function only after several sounds have been integrated over time. Two consequences follow. First, executive control cannot simply choose among ready-made symbols; it must also judge when sensory material has become coherent enough to interpret. Second, pitch material cannot be rehearsed internally without changing the very stream being heard, so rehearsal gives way to controlled decay. Sensory theories of rhythm point the same way: beat and meter arise from low-level temporal regularity and entrainment before conscious access (Fiveash et al., 2021; Emmery et al., 2023), and musical training sharpens these pre-attentive responses themselves (Olszewska et al., 2021).

The architecture, component by component

Figure 2. The processing path of musWM read as a working-memory architecture. Control and storage are separated, as in Baddeley’s model: the executive determines what is admitted and when, but stores nothing itself.

A worked example: from arrivals to a label

Worked example on a C major triad. Adapted from the figures in Tuğral (2026) on the set-like short-term buffer and the stabilised pitch-class set.

What this adds to statistical models of expectation

Statistical-learning models such as IDyOM describe expectation as the interaction of a corpus-trained long-term model and a piece-specific short-term model (Pearce, 2018). They are effective for that purpose, but their “short” and “long” components are learning timescales rather than memory systems; Rohrmeier and Koelsch (2012) note that the analogy with short-term memory “does not exactly hold”. The present architecture answers a complementary question.

How this relates to the figures on this site

The validation reported here tests one stage of the architecture: whether, once the short-term window and the checkpoint have produced a chord code, the long-term step names it as the textbook does (see Music Theory below and Ratings). It does not test the cognitive claims themselves. Those are framed as hypotheses to be checked against working-memory research, auditory neuroscience and music cognition — for instance, whether the windows and checkpoints of the model correspond to measurable signatures in the concurrent EEG and cardiovascular record (Tuğral, 2026).

Does musWM follow the textbook?

Measured against , over . The reference is the textbook itself: no human annotation and no other analysis corpus enters the calculation.

The 63 cores at every chromatic step

Loading…

Chord codes in the corpus outside the 63 cores

Loading…