This is the model most people like to use for that kind of thing but there are many others to try out. Also give Wizard/Vicuna a go.

I'm not familiar with LM Studio so I can't speak to their memory system but generally, you can think of tokens as words. The token limit for a model is how many it can handle at the same time. For local models, you're looking at 2048 for older ones, 4096 for more recent ones and some have been tweaked to work up to 8192.

The model itself has no memory. It's frozen in time and will not change as you use it. Instead, the "memory" is simply the entirety of the conversation, also known as context. GUIs make it look like you're only adding the latest query but in fact, ALL of the conversation is fed into the model as a prompt with every query. As such, if you reach the token limit, the prompt will be truncated and the the oldest parts will be "forgotten".

Some efforts have been made to extend "memory" by either condensing the conversation via a summary or by saving parts that have fallen out of the context to a file and then searching that file for relevant text with every new query. If it detects a match, it will re-inject the text it found in the file into the prompt.

These have varying degrees of success.