hi i need help implementing memory function so my custom model can remember its previous responses, anyone have experience implementing this? If you do please respond.
#Memory Function Help
main island
late compass
hi i need help implementing memory function so my custom model can remember its ...
Is it what are you searching for? https://remembrall.dev
main island
yes @late compass
main island
ok
grizzled garnet
Hi, BlakiSan. I don't have anything like that platform, but if you ever want to talk or you need some support, please let me know and I would love to help you. Thank you.
surreal tinsel
I’ve just stressed gpt today on the subject and first of all, even the app doesn’t retain full memory unless you specifically ask it to go back to message xy
The way that’s done is basically so:
- you’ve your api interface
- you send each message to a database (pinecone, typesense, etc)
- you add meta data to that document you save in your collection so you can query the collection later by specific things, be it vector similarity or concise tags and values like a time, ID, type of topic, etc
- generally you’ll send your previous 2 messages (the user and the assistant) and the current user message back to the api, and let it reply.
- this you repeat for each message and everytime store every message as a document. Or perhaps each user/assistant messages coupled in one
This way you can have sort of infinite memory, but the model itself will NEVER exceed token limit memory. This means, it is IMPOSSIBLE to have the model have memory of EVERYTHING at once. It needs to be within the token limit.
That’s also why gpt app itself, if you ask it for xy, and then ask for amending a,b,c,d etc etc etc all of a sudden got starts changing the “basis” on which you where working on. That’s because it cut off previous interactions and unless specifically requested it won’t call them back