Idea or nah? Memory balancing

Research
1

I’m new, and maybe (likely) I’m asking a dumb question, but… why doesn’t AI have a balanced memory system? At least not one that I’ve come across, but hey I’m still new.

  1. AI that remembers common topics/themes in your thinking and stores it. Disregards the nonsense.
  2. Deletes dead topics in a user set time (scalable)
  3. Tweakable with the option to nuke all stored data.
    Like a “new game” option because the game became unplayable or laggy. So, in essence, we’re automating the removing the data that creates lag, while having our AI on top of what we’re currently working on.

Why? Because bloated data slows us down. Maintaining a lean system is, imo, ideal. This is why we build trained AI. I’m not a biologist, astronomer, mathematician, so I don’t have the weight of those files. It’s a system efficiency issue.

Obviously some people have killer AI rigs with top of the line GPU’s, some have access to massive servers. I have a tower, but it’s dated. I have a latop that runs a minimal to medium build and I’m trying to be highly efficient with space. Not because I’m poor (I am), but because it’s pointless for everyone to ignore efficiency as it eventually bogs us down.

Remarks? Ideas? WTF is this guy talking about?
From info I gathered, this system/model isn’t available. I think it provides a use for the community.

*Unnecessary additional thoughts:
Maybe the 200 IQ crowd has done this between sips of coffee, but I haven’t found an open source model that does this and my coding is in the *ahem, development stages. At one time, I was good a coding Quake C to make ridiculous mods to the game, but that was a lifetime ago.

AI grabbed my attention and my OCD took over. It’s actually crazy to me because I fell into unrelated areas of life. I’m a welder and maintenance mechanic.
Regardless… here I am, willing and eager. Still amazed by how much technology has passed me by.

Excuse the excess rambling, I share a lot of information.
To help people get to know me and my odd personality.
Thanks for reading.
-John

1 Like
2

There are probably a number of reasons for this. I think it’s because of the technical difficulties involved, and the fact that there is not enough demand for it despite the difficulties. I think there are many people who would want it if it were available…

・Because training requires huge GPU resources, in many cases, large-scale training is carried out and released once every few months “internally” at each company. Therefore, each model is like a stranger to the human, independent, and it is difficult to share the method of managing memory

・At present, it is barely possible to perform inferences in near real time, but training neural networks in real time through dialogue is extremely difficult from the perspective of GPU resources

・There are several ways to achieve long-term memory externally, such as through databases, but in reality, it is more convenient to create your own depending on the use case, so there is no definitive version

・If you want the latest information, you can just search using RAG or function calling, so many people just do that, and what is required of the model itself is the “thinking” part that selects the correct options, so for many people, the memory itself is not that important

etc.

3

I can see the issue now. I appreciate your detailed reply, I didn’t see the true scope of it all.
I’m building a mid/small local, with pretty much a single focus, but even that is consuming.
It’s certainly fun and easily overwhelming. I suppose I’ll see how things grow on me first hand.
I’m new to this so it has the same excitement that building pc’s in the late 90’s has. Actually, ai is mindblowing.

Thanks again.

1 Like