Open minipasila opened 1 year ago
Thank for you suggestion. We're currently experimenting with embeddings. We could probably do a mix between the two at some point if "memory line prompt-pasting" is not efficient enough.
For reference: https://towardsdatascience.com/billion-scale-semantic-similarity-search-with-faiss-sbert-c845614962e2 https://arxiv.org/abs/2201.06009
Have a system that summarizes conversation history so you can have "long-term memory" kind of like how it's been implemented on this repository. This way even after you run out of tokens it will still remember most important parts of your conversation.