Second @keith comments as calling OpenAI’s endpoints is very straightforward.
The OpenAI API endpoints have a capacity number of tokens of 4096. Think of this as a means to see a big universe through a telescope., The 4096 tokens are what the telescope allows us to be able to access the more than hundreds of thousands of tokens that their LLMs (large languagee models) currently handle.
The arquitecture to build a memory to an AI app that is built conventionally or in any other way … i.e. … one of them being a Bubble app … requires the use of embeddings. These are actually vector embeddings (numbers that represent a semantic meaning) that can be handled by conventional databases but for best performance via vector databases such as Pinecone (that house the vectors which represent the combination of tokened-embeddings along with their text metadata).
Here a few resources that may prove useful:
https://twitter.com/johnjnay/status/1603779835238748162?s=61&t=4L9ClA18krHTHFPEl0YPmg
In all, what I meant to suggest is that to build expanded functionality into an AI app one requires an arquitecture that includes most of the above components in order to function adequately.