Depending on the task, 'memory' can be anything from unattainable luxury to bare minimum necessity. If it takes more computation power for the agent to 'remember' than to 're-calculate' then memory is probably redudant. Most applications are benefited from a simple structured memory, but managing an increasingly complex memory component introduces demanding overhead that isn't easy to handle at scale. Reasoning from/for the behavior of others could definetely enrich self-models and practically pushes decision-making to an ensemble setup. Best case scenario, the agent is given both a memory and a capability to reason from others and allowed to learn when and how to use them.

I don't understand the question. Perception is either the process of perceiving or the end product of such a process. Perceptions aren't agents that reason or do anything.

Addendum.

Aha. I see the question has now been revised, with "perception" changed to "perceptron". I still find the question awkwardly worded, and I'm not sure Asterios is answering what you intended, but I certainly agree with what he says. A note of historical interest: Leibniz is a good example of Asterios's point at the personal level. Leibniz apparently had a quirky memory and was often unable to recall problems that he had previously worked out or to easily locate his previous writings on those problems (which can be regarded as an external memory storage). So Leibniz, cutting short attempts to remember or search through trunks of manuscripts, often optimized by simply working out his views over again (recalculating).

However, to understand goings-on at the subpersonal level of neurons, by simulating them in artificial neural networks (perceptrons), I think we need to specify how memory is implemented or simulated within the network. Different setups will have different associated costs and involve different optimization strategies.

Addendum 2018-01-15

I see the question has been revised again, for the worse. :(