i wanna make some memories

So much of the discussion around LLMs and generative AI has been about composition of language chains, retrieval-generation, improvement in foundation models. Are we sleeping on memory?

I’m not saying that memory is not taken into consideration. Of course it is. I almost never create a LangChain model without it, but it seems like we only talk about memory for language models as if it’s RAM. In order for us to be able to push boundaries as engineers, we need to think about what it means to store our data on disk.

We spend so much time setting up vector databases in order to tune our optimized prompts with excellent context, and then we might add a memory buffer, but once the back and forth with the chat is over, the memory buffer dies, we move on.

I know there are tools in the space, I’m sure there are many more than I know of. But I don’t hear about them like I do about vector databases. But if we want to be able to build the types of language models that improve our lives to a whole new level, we need to be able to chat with a single interface that is able to store our information for the long term, condense it, and hold onto it. Long term memory was important for the last generation of language models. It needs to be important for this generation too. I’ll have more to say on this subject as I think through it, but I want to at least pose the question for now.