This article is part of article series: ChatGPT Features
Previous article: ChatGPT Features — Context and Instructions
Memory
Lets begin by discussing what memory is. As per Wikipedia definition, memory is a faculty of the mind by which data or information is encoded, stored, and retrieved when needed.
You can switch to video version of this article
It is the retention of information over time for the purpose of influencing future action.
If past events could not be remembered, it would be impossible for language, relationships, or personal identity to develop
Memory and ChatGPT
I must say they are pretty neat definitions that let us build understanding of what memory is. Since we are talking in context of ChatGPT, one should know that there are two types of memory at play here.
One is the memory it takes for ChatGPT application to function i.e. the data captured, stored and used to train the models. Big chunk of above definitions apply to that. ChatGPT or should we say OpenAI uses your conversation history to train and improve the model. It has, at any given time, access to conversations of hundreds of millions of people (even if they delete as they claim).
This requires a lot of memory, but from perspective of our article, we are not concerned about this memory. This is probably being added in the cloud by adding more machines. When we refer to memory going forward, we would mean the memory ChatGPT specifically has ‘about you’, the user.
ChatGPT User and Conversation Memory
By default, ChatGPT has no memory, but as discussed in last article about context and instructions, it slowly begins to use conversation history as memory to add more beef to the responses it generates. Its your job to add context and instructions to get meaningful interactions from it. We won’t be repeating that in detail here except that the context window used by GPT 3 is limited to 4000 tokens but GPT 4 based version would get a major upgrade of 32768 tokens.
Custom Instructions
The real reason for coming up with this article is the new feature called ‘Custom Instructions’ in ChatGPT. This feature saves you from adding context about yourself and instructions to format response by saving some context and instruction permanently. Its like adding a hard drive to your computer without which it wipes out as soon as it reboots. So this new feature, accessible on the interface using the 3 dots (currently) allows you to tell ChatGPT
About yourself
How you would like it to respond.
The about yourself part is the context part, where you can tell ChatGPT about what kind of work do you do, what is your education and experience level, what interests you, where you live etc. Not random information about yourself, but information that would help ChatGPT to give you more personalized response.
The example they themselves provide is that of a teacher which provides information about what level teacher she is so she doesn’t have to repeat that the topics she queries are for developing lesson plans for 3rd grade science class. That way, when she queries anything like ‘electrical circuit’, ChatGPT would spare her the graduate level complex intricacies of electricity and would keep the response understandable to kids (without her explicitly requesting it each time).
The next part i.e. how would you like it to respond, is the instructions about output formatting. Its like setting defaults e.g. would you like answers to be detailed or summarized, do you want them to be formal or informal etc.
ChatGPT and privacy
You always need to be careful on what information you are divulging over the internet. For pessimists among us, we hear about stored data being sold or abused by small and big names alike all the time. With shady play with words in privacy statements and default opt ins setup when you sign up without reading privacy statement. And, from that perspective, who cares what privacy statements have to say anyway. But if we go extreme with this, we can’t connect the computer to internet ever.
So we’ll stick to the ‘pragmatic optimist’ version and lets have a look at how you can control your data with ChatGPT. We already discussed your conversations are used (via default opt in) to train ChatGPT model and are retained for 30 days (currently).
You have the option to opt out of model training using your conversation history though via explicit actions on OpenAI website. You can even disable history retention via data controls provided. I want to talk with specific reference to the two fields introduced with custom instructions where you are providing information about yourself.
Now you are doing that with intent of getting better response from ChatGPT, but you are also providing personal information online, which can be used in ways you don’t anticipate. So best to be careful and selective. Best to avoid any financial or sensitive information which can be harmful in case of a data leak or something. Like its best if you don’t try to ask ChatGPT to remember your ATM PIN code.
OpenAI may be a very honorable organization dedicated to data privacy, but it is a valuable target for hackers and people with less honorable intent. The possibility of data leak and data abuse is always there. Just like words you use in conversation, once you say it, you no longer control it. It applies both to your physical/personal life, and your life online. Be careful, be selective.
Hopefully you know enough about custom instructions and memory retention of ChatGPT. If you liked the article, please consider to clap and/or share.
Next article: ChatGPT Features — Output Formatting
This article is part of article series: ChatGPT Features
