Great thanks to all my co-authors @kristorpjensen.bsky.social @qlu.bsky.social @qiongzhang.bsky.social @marcelomattar.bsky.social!
Special thanks to Prof. Chris Baldassano @chrisbaldassano.bsky.social and everyone in Mattar lab for providing ideas and feedbacks!

In general, our work shows that different free recall strategies can emerge through task optimization, and the memory palace is the more optimal among them. If you want to know more details such as what factors influence the learned strategy of our model, check our preprint!

The memory palace strategy performs better than other strategies and is more robust to noise.

This index code is invariant of specific item identity only for the memory palace models.

The model learned multiple memory strategies, including a strategy like the memory palace, which learned representation of item index to guide memory search. Some other models showed retrieval mechanism closer to the temporal context model.

In this work, we trained an RNN + slot-based memory module on the free recall task with reinforcement learning.

Humans show individual difference in memory ability. One memory technique widely used by memory champions is the “memory palace”, in which people bind items with locations in a mental environment and go through specific trajectories to improve memory performance.