Memory is Reconstructed Not Retrieved

On June 15, 2606.06036 appeared on Hugging Face's daily papers. Memory is a hot topic at the moment so I decided to skim the paper and I found it a reasonable architecture. Having implemented various graph-based systems before (Hamilton, being one) I concluded that it was natural to implement this as well, however, I did something different this time, instead of spending my time writing code, I spent most of it thinking about system design.

Paper

I will talk about some decisions I took and how they affected the final product, which is now open sourced and available for testing. This is the first time that I implement a paper and to be fair nowadays it's really easy to do so, I think the bulk of the difficulty now is reduced to having a clear understanding of its contents, in order to avoid black-boxing the implementation an agent ends up making.

What do we really need?