What an agent remembers is decided by two independent choices: where the memory lives, and how it gets written.
The only store that grows with the interaction. Every token appends, so capacity scales with context length and cost scales with it too — which is why the questions here are eviction, compression and selection rather than capacity. Splits into dense (keep everything) and sparse (keep a selected subset).
Fixed in size and written by gradient descent. Durable but expensive to change, and changing it risks everything else the model knows — the failure mode continual learning and unlearning both work against.
A fixed-size learned representation. Like the weights it is written by training, so what it encodes is decided before the interaction starts, not during it.
The recurrent or latent state carried between steps — fixed in size, overwritten in place as the interaction proceeds. Constant cost, but everything not kept is gone.
What the update optimizes for. The same substrate behaves very differently depending on whether the target is next-token likelihood, a preference signal, a distillation target, or an explicit forgetting term.
Samples drawn independently from a fixed distribution — the standard training assumption. Shuffle freely, revisit anything, and the ordering carries no information.
Samples arrive in order, once, from a distribution that moves. Order matters, revisiting is limited, and the forgetting that follows is what continual learning has to hold off.