Making memory dynamics (store / retrieve / decay / overwrite) the primary trained object, instead of model weights.

article, coming soon

most builders would want to put memory as part of the ai agent but it is not. the truth is Memory is not optimized as a goal in itself. memory is either:

• passive storage, or

• a training crauch to protect ai weights.

i starndard ML.

lose -> update weights memory is auxilary.

in this blog:

loss -> update memory policy.

weights stay mostly fixed.

to understand this you need to understand 2 things:

• learning a state transition system.

• not learning a function, but learing how a state evolves over time