Hacker News new | ask | show | jobs
by simon_rider 1192 days ago
Sounds interesting, how do you create the long term memories in the short 4096 token window? Can you share something?
1 comments

I let GPT summarize the conversation, and use embeddings for semantic search to recall relevant memories. The memories are inserted into part of the prompt.

It's imperfect because of the token window (much like human memory).

It can generally recall things I ask it to remember.

I do save a complete, exact chat record, but currently I don't use that except for the recent chat (in case we're working on some code or whatever that needs to be exact).

It still has plenty of problems, don't get me wrong.