So, so, so curious how they are implementing this.
MemGPT I would assume + background worker that scans through your conversation to add new items.
Memory and new controls for ChatGPT
31–40 of 270 posts
Re: Memory and new controls for ChatGPT
#32I love this idea and it leads me to a question for everyone here. I've done a bunch of user interviews of ChatGPT, Pi, Gemini, etc. users and find there are two common usage patterns: 1. "Transactional" where every chat is a separate question, sort of like a Google search... People don't expect memory or any continuity between chats. 2. "Relationship-driven" where people chat with the LLM as if it's a friend or colle…
Personally i would like a kind of 2D Map of 'contexts' in which i can choose in space where to ask new questions. Each context would contain sub contexts. For example maybe I'm looking for career advice and I start out a chat with details of my job history, then im looking for a job and i paste in my cv, then im applying for a specific job and i paste in the job description. It would be nice to easily navigate to the career+cv+specific job description and start a new chat with 'whats missing from my cv that i should highlight for this job'.
I find that I ask a mix of one of questions and questions that require a lot of refinement, and the latter get buried among the former when i try and find them again, so i end up re explaining myself in new chats.
Re: Memory and new controls for ChatGPT
#33It is already ignoring your prompt and custom instructions. For example, If I explicity ask it to provide a code instead of an overview it will respond by apologizing and then provide the same overview answer with minimal if no code. Will memory provide a solution to that or will be a different thing to ignore?
I have some success by telling it to not speak to me unless it's in code comments. If it must explain anything, do it it in a code comment.
Re: Memory and new controls for ChatGPT
#34I love this idea and it leads me to a question for everyone here. I've done a bunch of user interviews of ChatGPT, Pi, Gemini, etc. users and find there are two common usage patterns: 1. "Transactional" where every chat is a separate question, sort of like a Google search... People don't expect memory or any continuity between chats. 2. "Relationship-driven" where people chat with the LLM as if it's a friend or colle…
I use for transactional tasks. Mostly of the "I need a program/script/command line that does X".
Some memory might actually be helpful. For example having it know that I have a Mac will give me Mac specific answers to command line questions without me having to add "for the Mac" to my prompt. Or having it know that I prefer python it will give coding answers in Python.
But in all those cases it takes me just a few characters to express that context with each request, and to be honest, I'll probably do it anyway even with memory, because it's habit at this point.
Re: Memory and new controls for ChatGPT
#35I love this idea and it leads me to a question for everyone here. I've done a bunch of user interviews of ChatGPT, Pi, Gemini, etc. users and find there are two common usage patterns: 1. "Transactional" where every chat is a separate question, sort of like a Google search... People don't expect memory or any continuity between chats. 2. "Relationship-driven" where people chat with the LLM as if it's a friend or colle…
Personally i would like a kind of 2D Map of 'contexts' in which i can choose in space where to ask new questions. Each context would contain sub contexts. For example maybe I'm looking for career advice and I start out a chat with details of my job history, then im looking for a job and i paste in my cv, then im applying for a specific job and i paste in the job description. It would be nice to easily navigate to the…
I think it’s less of a 2D structure and more of a tree structure that you are describing. I’ve also felt the need of having “threads” with ChatGPT that I wish I could follow.
Re: Memory and new controls for ChatGPT
#36I love this idea and it leads me to a question for everyone here. I've done a bunch of user interviews of ChatGPT, Pi, Gemini, etc. users and find there are two common usage patterns: 1. "Transactional" where every chat is a separate question, sort of like a Google search... People don't expect memory or any continuity between chats. 2. "Relationship-driven" where people chat with the LLM as if it's a friend or colle…
Personally, I always expect every "conversation" to be starting from a blank slate, and I'm not sure I'd want it any other way unless I can self-host the whole thing. Starting clean also has the benefit of knowing the prompt/history is in a clean/"known-good" state, and that there's nothing in the memory that's going to cause the LLM to get weird on me.
I have thought of implementing something like you are describing using local LLMs. Chunk the text of all conversations, use an embeddings data store for search, and for each new conversation calculate an embedding for the new prompt, add context text from previous conversations. This would be maybe 100 lines of Python, if that. Really, a RAG application, storing as chunks previous conversations.
Re: Memory and new controls for ChatGPT
#37I use an API that I threw together which provides a backend for custom ChatGPT bots. There are only a few routes and parameters to keep it simple, anything complicated like arrays in json can cause issues. ChatGPT can perform searches, retrieve documents by an ID, or POST output for long-term storage, and I've integrated SearxNG and a headless browser API endpoint as well and try to keep it a closed loop so that all information passing to chatGPT from the web flows through my API first. I made it turn on my lights once too, but that was kind of dumb.
When you start to pull in multiple large documents, especially all at once, things start to act weird, but pulling in documents one at a time seems to preserve context over multiple documents. There's a character limit of 100k per API request, so I'm assuming a 32k context window, but it's not totally clear what is going on in the background.
It's kind of clunky but works well enough for me. It's not something that I would be putting sensitive info into - but it's also much cheaper than using GPT-4 via the API and I maintain control of the data flow and storage.
Re: Memory and new controls for ChatGPT
#38My use of ChatGPT has just organically gone down 90%. It's unable to do any sort of task of non-trivial complexity e.g. complex coding tasks, writing complex prose that conforms precisely to what's been asked etc. Also I hate the fact that it has to answer everything in bullet points, even when it's not needed, clearly rlhf-ed. At this point, my question types have become what you would ask a tool like perplexity.
Re: Memory and new controls for ChatGPT
#39It got so difficult to force ChatGPT to give me the full code in the answer, when I have some code related problems.
Always this patchwork of „insert your previous code here“
This is not a problem of the model, but I suspect it is in the system prompt that got some major issues.
Re: Memory and new controls for ChatGPT
#40It is already ignoring your prompt and custom instructions. For example, If I explicity ask it to provide a code instead of an overview it will respond by apologizing and then provide the same overview answer with minimal if no code. Will memory provide a solution to that or will be a different thing to ignore?
I have some success by telling it to not speak to me unless it's in code comments. If it must explain anything, do it it in a code comment.
I love when people express frustration with this shitty stochastic system and others respond with things like "no no, you need to whisper the prompt into its ear and do so lovingly or it won't give you the output it wants"