[HS]

Like an ecosystem, a sustainable mind will need to make and break itself. Old stuff will need to make way for new stuff, because there is a limit on space and resources, but the system can still grow as a whole, in quality, like evolution. We tend to look unfavorably on subtractive/reductive processes, so most of our attention is focused on giving AI the best additive/constructive processes. These poor robots will be like high speed cameras, 8 GB/S, and rising, not to mention all the internally generated stuff. They will need the ability to stop thinking, they will need to sleep, and they will need to forget things without conscious effort. These things are already indispensable to us, and we don't have amazing lifespans, or great speed in the additive/constructive department. In this narrow sense, it doesn't matter how much memory an AI has, eventually the brain will need to settle on a balance point between making and breaking, like any lasting dynamic place. What are the best ways of doing this?

[LOCKSUIT]

Indeed evolution rotates us to make better structures. Sad.

With AIs, they can stay immortal and clone. They can update and transfer their minds to better bodies instead of restarting the process like kids must go through.

Yes, things like memories must go too, it's ok we still stay us. And importantly, our agenda is unchanged. I just know my work better day by day it is seeming. As korrelan said I may even lose how i got there though! Gotta go through your work and save it in notepad how you got there. All your facts you learnt.

The bad, and the good.

[WriterOfMinds]

My own forays into modeling things like this are primarily covered under these three posts:

https://writerofminds.blogspot.com/2018/04/acuitas-diary-10-march-2018.html
https://writerofminds.blogspot.com/2018/05/acuitas-diary-11-april-2018.html
https://writerofminds.blogspot.com/2019/04/acuitas-diary-17-marchapril-2019.html

Basically, I use the sleep half of the wake-sleep cycle for anything that qualifies as routine self-maintenance, including forgetting episodic memories and compressing or error-checking information in the semantic memory. The forgetting algorithm is fairly computation-intensive, so forcing it to run during sleep-time prevents it from inadvertently causing lag during conversations etc.

I can't speak to any best practices or competing approaches. I know there are other systems that have implemented episodic memory of some kind, but I can't think of any that do selective forgetting right now.

Neural networks may "forget" things if you re-train them, but I've only ever heard this spoken of as a bad thing ("catastrophic forgetting").

[goaty]

[edit] oops I think I just repeated and agreed with what you posted. I said it slightly dif, ill just leave it here anyway [edit]

The actual ending reality of making a actual functioning robot mind,  in my opinion, you have to fit it inside of limited resources,  so this means you will be "rotating" things in its head, especially when it comes to patterns arriving from the environment,   Maybe a bit of incremental learning could help distinguish the useful patterns from the less useful (But of course it is ideal if it didn't forget anything!) and you can keep a fixed count in its head, and once it hits its limit, it has to throw them as they come in.  Then maybe you don't get all the perks of ideal artificial intelligence, like "photographic memory", but maybe you could still keep the high reflexes, and others that are still there to grab, with just a small computer operating it, like an Arduino or old computer.

[HS]

Thanks goaty. I agree  with what you said too, just couldn't reply because the site was inaccessible here for a few days.