The SOB-CS model, developed by Stephan Farrell and Stephan Lewandowsky (2002) and extended by Klaus Oberauer and Lewandowsky (2008, 2011), is a computational model of serial recall and working memory that uses energy-gated encoding and interference-based forgetting. Critically, SOB-CS requires no decay mechanism: forgetting occurs entirely through the superposition of memory representations.
Architecture
SOB-CS stores items by associating item representations with positional markers in a shared weight matrix, similar to matrix models of memory. Items are represented as distributed vectors, and positions are represented by a separate set of vectors. The association between an item and its position is stored via a modified Hebbian learning rule:
where eᵢ is the item vector, posᵢ is the position vector, and W is the weight matrix. The term (eᵢ − W · posᵢ) is the novelty of the item given what is already stored: if the current item-position binding is already well represented in W, the weight change is small. This energy-gated encoding is crucial because it means that similar items interfere less during encoding.
Interference-Based Forgetting
In complex span tasks, distractor items presented between to-be-remembered items are also encoded into the weight matrix (though with the same energy-gating mechanism). These distractor representations create interference that degrades retrieval of target items. The more similar the distractors are to the targets or to each other, the more they interfere. Forgetting in SOB-CS is therefore driven entirely by the content and similarity of intervening material, not by the passage of time.
Retrieval via Competitive Queuing
Recall uses a competitive queuing mechanism: the position cue is applied to the weight matrix to generate an output vector, which is compared to all item representations. The item with the highest match is selected, then suppressed to prevent perseveration. The retrieved item is also "unlearned" from the weight matrix via anti-Hebbian adjustment, reducing its interference with subsequent retrievals.
Benchmark Tests Against TBRS*
Oberauer and Lewandowsky (2011) conducted a systematic comparison of SOB-CS and TBRS* across a large set of complex span benchmark findings. SOB-CS provided a comprehensive account of cognitive load effects, distractor similarity effects, and serial position curves without any decay mechanism. The model demonstrates that the apparent time-based effects in working memory can be explained by the amount of encoding that occurs during processing intervals, rather than by temporal decay during those intervals.
The energy-gating mechanism in SOB-CS acts as an automatic novelty detector. Repeated or similar distractors are encoded with progressively smaller weight changes, reducing their interference. This elegantly explains why increasing the number of identical distractors does not proportionally increase forgetting, a finding that is problematic for pure interference models without energy gating.