Memory Traces: Are Transformers Tulving Machines?

Abstract

Memory traces--changes in the memory system that result from the perception and encoding of an event--were measured in pioneering studies by Endel Tulving and Michael J. Watkins in 1975. These and further experiments informed the maturation of Tulving's memory model, from the GAPS (General Abstract Processing System} to the SPI (Serial-Parallel Independent) model. Having current top of the line LLMs revisit the original Tulving-Watkins tests may help in assessing whether foundation models completely instantiate or not this class of psychological models.

Figures (2)

• Figure 1: The SPI Model of memory. The original 3-layer GAPS model later expanded to 5 systems citeciteproc_bib_item_5[5] into the Serial-Parallel Independent model. The Procedural Memory sits apart from a hierarchy of four systems: a Perception Representation System linked to priming effects, Semantic Memory associated with familiarity, Working Memory acting as a buffer citeciteproc_bib_item_6[6] to the Episodic Memory associated with recollection, "the individual's subjective awareness of remembering" citeciteproc_bib_item_2[2]. Encoding of an event or episode happens serially along the vertical arrows, and information (memory traces) are stored in parallel in each system. Retrievals from each system, on the other hand, happens independently from other systems (horizontal arrows). Source for image: citeciteproc_bib_item_5[5].
• Figure 2: Summary of memory performance (recollection) in LLM, left, and human, right. Both tables are presented in the original data matrix format from citeciteproc_bib_item_1[1], showing frequency of occurrences normalized to [0,1]. Following the original paper, the test targets two independent categories of word-cues: associative, noted A, and rhyming, noted R (see text). In a test session a list of words, each given with a context word, either A or R, is memorized. Two cued-retrievals are then performed in sequence, in the same LLM chat, exploring both alternations of cue categories, A then R and R then A. This setup of two encodings and two retrievals results in four rows marked on the vertical axis as A/AR, A/RA, R/AR, R/RA. In a table, each row displays frequency of occurrences of the four possible outcomes of the Tulving-Watkins cued dual retrieval tests: the word has been retrieved (upper case letter A or R) from the cue, or not (lower-case letter a or r). Drastic differences between LLM and human in memory performance profiles are discussed in the text.