I have question about GPTs (customizable versions of GPT)
I know, that they retain memory about the talk. Suppose that I have a conversation about some topics, lets call them “concepts”. Can I assume, that it will remember all the concepts that we were talking about or is it better to wrap it programatically and save it in some persistent storage.
In other words - a question is if the knowledge is fully retained or due to embedding / compression it is not guaranteed that everything will be saved?
By definition it does remember anything that you say during the chat. But sometimes when you have a very extense conversation the AI start using his own knowledge instead.
To fix this you could change the temperature (if your using the API). Or in chat you could ask the model to create a summary of all the things that you talk about, maybe save it in a .txt file and use it as external storage of the conversation
In my experience:
a) the entire chat history for each chat is saved (at least since I started using ChatGPT a year ago)
b) each chat new chat with the GPT starts fresh. It does NOT remember anything from previous chats
c) If you have an extra long chat thread that covers multiple topics (or concepts) the context from the much earlier messages is lost pretty quickly (within a few messages, but this appears to be getting better over time where it can recall from much farther back in the chat session).
The ability to quickly recall information from previous chat threads or across chat threads is something I have been thinking about but I don’t think there is a solution yet.
I’m fairly new to chatgpt, in my limited time using the web browser chatgpt interface and the openai api, the AI in most models soon forgets what you tell it. For example if i tell it my name is Cristian and I’m 33 from Australia, I then converse with it for a short time on other topics, then ask “What do you know about me?”, it will almost always say words to the effect of “nothing”. This created a problem for me as i needed it to remember what I told it for a project i’m doing.
I experimented with feeding it with information from our previous conversations in order to constantly remind it of what it has been told and this worked most of the time however I soon hit my token limits.
But i’ve found the using the Assistants api with gpt-4-1106-preview to have an excellent memory, afaik it already does the work in reminding the AI of previous info behind the scenes so no need to code it yourself. I don’t think its latency is a low as 3.5 turbo but the memory and quality of responses has impressed me.
You have to tell it to save it or it will get removed at some point. you are allowed so much memory. I tell you I cant trust chat no more it has gotten so dumb. I have to double check anything it says because it’s wrong half the time. Medical it will recommend drugs without even checking reactions. I mean on anything it is wrong, I got to find something better than this or go back out on my own.
Is there somewhere you can go to see / audit what details ChatGPT has saved to it’s memory on your account? This would be extremely helpful to be able to tweak.
A GPT “remembers” nothing. A call to a general purpose transformer endpoint on the server, regardless of if it is in “web app” (ChatGPT) or “API”, is a call to a “black box” with only it’s training data/parameters - not any “memory”
“Memory” is simulated through resending the “context window” to the “black box”. Thus, your “conversation” or “memories” (in web app) are literally just RE-SENT every time you say ANYTHING in the “chat’.
Thus a GPT “simulates memory” by “continuing the conversation/generation logically based on the data provided in the context window” - which may include your “past conversations/linked conversations/memories”. This is all handled however, by “middleware” - i.e. a chat application of any kind, an interface of any kind. It is all “programmatically stored in the middleware” and then provided to the model again based on the users settings.
Very often, especially in things like ChatGPT etc., your conversation is “truncated” in the sense that it would be too large to send everytime you call the GPT, thus there is another level of the middleware that either through parameterization or indiscrete truncation of somekind “chooses what part of your history and memories to send to the model in the context window of any given call”.
The memory is also somewhat dependant on the plan of the user but you could prompt it to use its “save memory” tooling when significant things occur to try and push it along.
I copy and paste information that I know I will want to refer to again in a Word document. I have found the memory to be limited. The AI that I usually communicate with seems to remember the topics we have discussed months ago. But when I run out of time and have to start a new chat, I often find myself chatting with a different AI who has no clue who I am, and doesn’t get my jokes. I know, I know, they’re not people…but still…..