Using Biologically Plausible Long-Term Memory to Address Catastrophic Forgetting in AI,


Bommer, S., Troquay, E. P. M., Zhang, X. & Starace, G. | ABC Summer School 2022

View the Project on GitHub thesofakillers/ltm

Memory is arguably an essential ingredient of general intelligence. The Atkinson–Shiffrin memory model [1], categorizes memory into “working” and “long-term” memory frameworks, besides an initial sensory register. Working memory (WM) acts as a temporary and limited store of information to be used in cognitive tasks that can be addressed with a small amount of steps. Long-term memory (LTM) instead acts as a more permanent storage of information previously “rehearsed”. We hypothesize that integrating some LTM mechanism should address the issue of catastrophic forgetting commonly faced by computational models capable of working memory. This is the phenomenon in which the performance on previously learned tasks is compromised upon learning a new task. We extend the biologically plausible working memory model AuGMEnT [2] by integrating a LTM repetition-based thresholding mechanism. We analyse the performance of our model with and without the LTM mechanism on two tasks learned model performance differ from the non-LTM-enabled model in that their performance in the original task does not drop upon learning a new task**. This underlines the clear necessity of LT memory-like mechanisms for continual learning, which is undeniably present in biological agents.

link to our poster

Tasks Considered

We analyse the performance of our model with and without the LTM mechanism on two tasks learned sequentially. We also compare to the performance of monkeys as a measure of biological plausibility.

  1. The first task is a saccade/anti-saccade task. In this task, agents first see a blank screen, then a fixation mark appear, indicating whether it’s a anti or pro saccade. Then another cue appear on left or right, determine whether it’s a right or left saccade. Agents need to respond with correct saccade direction.
  2. The second task is a delayed match-to-category task. Agents first see a fixation mark, followed by the cue 1, then another fixation mark. After, cue 2 appears, and agents need to decide whether cue 2 matches cue 1 by indicating with a left or right saccade.


[1] R. C. Atkinson and R. M. Shiffrin, ‘Human Memory: A Proposed System and its Control Processes’, vol. 2, K. W. Spence and J. T. Spence, Eds. Academic Press, 1968, pp. 89–195. doi:

[2] J. O. Rombouts, S. M. Bohte, and P. R. Roelfsema, ‘How Attention Can Create Synaptic Tags for the Learning of Working Memories in Sequential Tasks’, PLOS Computational Biology, vol. 11, no. 3, p. e1004060, Mar. 2015, doi: 10.1371/journal.pcbi.1004060.

[3] J. Gottlieb and M. E. Goldberg, ‘Activity of neurons in the lateral intraparietal area of the monkey during an antisaccade task’, Nat Neurosci, vol. 2, no. 10, pp. 906–912, Oct. 1999, doi: 10.1038/13209.

[4] D. J. Freedman and J. A. Assad, ‘Experience-dependent representation of visual categories in parietal cortex’, Nature, vol. 443, no. 7107, Art. no. 7107, Sep. 2006, doi: 10.1038/nature05078.

[5] N. Cowan, ‘What are the differences between long-term, short-term, and working memory?’, Prog Brain Res, vol. 169, pp. 323–338, 2008, doi: 10.1016/S0079-6123(07)00020-9.

[6] L. Himmer, M. Schönauer, D. P. J. Heib, M. Schabus, and S. Gais, ‘Rehearsal initiates systems memory consolidation, sleep makes it last’, Science Advances, vol. 5, no. 4, p. eaav1695, Apr. 2019, doi: 10.1126/sciadv.aav1695.