ChatGPT rolled out a cool new memory feature.
Discover More From Me:
🛠️ Explore thousands of AI Tools: futuretools.io/
📰 Weekly Newsletter: www.futuretools.io/newsletter
🎙️ The Next Wave Podcast: @TheNextWavePod
😊 Discord Community: futuretools.io/discord
❌ Follow me on X: x.com/mreflow
🧵 Follow me on Instagram: / mr.eflow
Sponsorship/Media Inquiries: tally.so/r/nrBVlp
Mailing Address: 3755 Avocado Blvd Unit 287, La Mesa, CA 91941
#AINews #AITools #ArtificialIntelligence
I wish OpenAI would just added grouping/tags/folders for our chats, so we can organise our chat history into projects like a sane person.
Yeah it’s pretty hilarious how few options they have in this regard.
There are some plugin that could do that (google plugins)
Huh. That sounds like a strangely specific type of need. Isn't one session per topic organised enough?
I would find a search feature more valuable.
I hope they read that comment.
I've had memory on for a few weeks now. The problem is that ChatGPT begins to remember things from side-projects or throw-away conversations, which can impact how it might steer the majority of unrelated future conversations. Memory would be better implemented if users were given the ability to group like conversations into "Projects" or "Topics", where unique memories could reside within those defined groups and be applied to those group chats. Ie; If I have a project about puppies, memories of these chats won't flow onto my other project about learning Linux Ubuntu.
Yes I totally agree with this. Also, some chats have output that is not useful/off-the-mark, but I keep it just in case it may be useful for reference in the future (or knowing how not to prompt). But I don't want that bad output to influence any other chats.
It's an extra step, but you can save and load your memories grouped in topics by using JSON objects/files. Save memories for a particular project or topic and then when you need just those, delete current memory and load just those.
Your channel is great because I get all the tech news without having to expose myself to the toxicity of twitter!
Oh ya? Well let me tell you something racist.
@@makavelismithwell? we're waiting
@@ryzikx unfortunately I think that all races are equal. You're free to pretend that I've just said something horrible though.
@@makavelismith It's always an option to hate everyone equally.
Haha same
A few points I discovered talking to ChatGPT about the memory function: * memory is not unlimited and may drop memory that no longer appears relevant but there is no specific limit/space for memory entries * using the memories doesn't contribute directly to token count * memory is referenced before outputting a response (so the output may change based on what it has remembered) * while ChatGPT doesn't know the architecture it agreed that a good guess as to how it works is through converting the memories to embeddings (numeric versions of the text) * memory is all stored in the backend but it does contain privacy controls (which is probably why the "memory" didn't record your location - but I am not sure why it remembered your dogs name)
Chatting with ChatGPT about its functionalities has not worked for me -it was constantly hallucinating
You learned this from ChatGPT itself? Use a huge grain of salt with this information. It is extremely prone to guessing about its abilities. It's better to do real testing.
2:24 I actually made my Custom Instructions into memory instead, freeing up my actual custom instructions for more succint, general preferences and instructions that can fit any chat, which is pretty nice as 1500 characters just wasn't enough for me 😅
Was thinking along those lines. Most interested in using memory to greatly expand custom instructions. Not in giving billionaires and government detailed information on me personally
Gosh... I cant even get friends that don't say, "Oh... I remember that"
Thank u this was a great idea! 🎉
Anything that can fit any chat should still be able to be converted into memories. Custom prompts should be local to the next new chat. Currently, as a workaround I have to use a separate chat to save different custom prompts and tell it to shut up when I paste them in. Naturally, it forgets to shut up after just a few entries...well, I guess I can make that into a memory, now. If I ever got the pro version...
We are using it to learn our writing styles. We have it give us information, Ethen we edit it to more our styles, copy and paste it back and ask ChatGPT to remember our edits for the future. By the end of a project, there are very few edits we need to make
great idea!
Wow this is an amazing IDEA thanks 🙏🏿
Amazing IDEA THANKS 🙏🏿
Is this for articles or books?
@ViRiXDreamcore Sounds most likely for some company's marketing materials. But it also sounds usable for book; why not? I understand the need for volume and nuance is greater for books - but that should be totally doable.
Those features that explicitly ask to remember and forget are nice, but it also remembers certain things from your queries on its own, and shows the 'Memory Updated' link. I liked that.
You can also go to past chats to add that stuff to memory. Additionally If it's a chat that is archived due to plugins, you can still edit an existing response to add that chat to memory.
Huh! That's pretty good to know, actually. Does anybody use that plural "GPT's" option that replaced plugins? I wonder how common it is. To me it just looks like a hassle. But I was a paying and active user during the plugins phase, and completely ignored it for a while, then came back for this "GPT's" phase - and have no idea what to do with that new feature. I mean... it doesn't look necessary enough to need "custom GPT's" for anything. It browses, does code and and generates images, all in regular mode, so why would anybody need it?
@@OdinsHarem GPT's are good for niche purposes, and even then, only the ones that have a proprietary backend.
I got memory a few days ago too. Here are a couple useful things you can do: Ask it to save your memories as a JSON object and print it to a code block so you can copy and save it. Do the reverse, load your memories in from a JSON file. You can use this to group memories into discrete chunks, like for different projects or topics and then load in just the ones you need for a certain chat. When you save memories, you can also tell it to separate the memories into smaller single items, so that you have more fine-grained control over what you keep and what you delete.
Thanks!
This explanation what I was waiting for!! Thank you Matt!
This is so epic! I really needed this! Very happy to see it rolling out! Thanks Matt! 🖤
I love your videos, I always wanna keep up with AI News, update myself with New techniques coming up and You promptly deliver them. Many Thanks !
2:56 I am super excited to hear that strategic move to give Ai memories (or a memory bank) to give more context to the whole conversation. Super awesome 😎💯💪🏿👍🏿.
I wish they'd give it to Phi
LMAO "just like in real life sometime" legit spit my coffee out. Thanks Matt
@@yeenaryea seriously, i found it funny too. Now go back in your man cave
@yeenar ...because sometimes my memory banks are empty. How many kids you have? 🤣
🎉 this weekend I just had my kind of blank bday 🎈
It sent me 😂😂😂😂😂😂
Aww love the personalization and sharing some insights about yourself matt. And so useful love the remember this feature of ai to not have to repeat self well in future for now this is worse it will be and in my opinion pretty dang great to say it once put it in its mind and store for reference. What we put in is what it remembers.
This is essential. Without long-term memory AI can't gain human-like intelligence, relationships, or be a useful assistant. This is the first step.
If a person uses this instrument for work and for good purposes, this is a very good update... I hope it will be improved and expanded in the future. Thx Budy for vid!
I love the idea, but only locally. Data brokering is already so profitable. Let's not incentivize corruption by putting it all out there, then expecting them not to change term & conditions later.
Welcome to the creation of your covert psychological profile for the CIA in the NSA.
Hi Matt, great video. It’ll be interesting how all that information will be used to buster our own personal assistants in the future. Still waiting on the memory upgrade 😢
Be careful. When you have multiple roles or are *role-playing,* ChatGPT will learn the wrong stuff about you. (On Project A you may be a programmer, on Project B an artist, in another chat you could pretend to be Alexander the Great or a coin collector. ChatGPT will learn all this and may mix stuff up)
Great video, thanks!
yes, and in EU I still don't have it. I hope I get it soon, cause I am tired of explaining everything to chatgpt in every new chat from the beginning
I have a text file which i dump in at the start of every chat session. Not ideal, but works.
@@AndrejMejac I also use text, but sometimes when my job is 50% done and chat crashes because of errors or because it's too long, my base text won't help, I still need to explain almost everything to get to this 50% and then start progressing forward. I am not sure whether this "memory" feature helps with that a lot or just a little bit, but at least something )
Buddy, you’ve got to think differently: if you use ChatGPT on a project by project basis, what are you gonna do between projects? You’re gonna have to delete old project information so that it won’t reference that, and add new project information. Which is going to be a headache. THE SOLUTION (that I’ve had for over a year): create an Evernote or Microsoft Word help document. In that Evernote document paste all of that initial information at the start of every chat. That is still going to be the best way to communicate with ChatGPT, again, if you’re a project based user.
This tends to happen with every feature update. I wonder why the EU tends to be so far down the line for OpenAI; is it data / privacy laws? Language considerations? Lack of red teamers relevant to the area?
@@BrassDragonI'm not surprised. The EU is by far the most specific about privacy and personal data. I'm so glad I'm from a European country outside of the EU, lol... phew 😌
Got it two days ago, I think it will help me a lot, I'm an English teacher, I have like 100 students so to create stuff, come up new activities this is going to be very helpful
the students got it too, they will use it to answer you chatGPT created questions. Cheating each other is the game you both are playing and nobody, apart from chatGPT and OpenAI is learning from it. Bot you and your students are just tools used to train the next version.
I have had this for 2 months now.. not sure why the delay Matt... but thanks for metion it again.
This will be very helpful. You can add items you never want chat gpt to say, use, or do. Like avoid compound modifiers or other terms
Yeah, you can do that. And it will continue to ignore it. Like it does in custom prompts.
Thanks Matt
great video man
Thanks!
I wonder how knowing personal data will have influence on real life general use of ChatGPT. Love you vids.
It will probably make it feel like a more personalised experience. And ChatGPT will probably feel even more like a capable personal assistent.
Good find! Didn’t realize that I now also have this.
have you realized that octopuses have distributed nervous system ;)?
Glad to see people are getting updates for memory but i still didnt get it so this is very teasing on me
Finally!!! I had headaches with gpts when I wanna to remember certain instructions for little time he do it right then forgot everything…
One thing i hope they have soon is maybe the ability to store certain groups of memory. I use gpt with mutliple hats (programmer, localizations, general office work etc). I dont need my office work stuff remember my programming memory and vice versa. The other thing is how much memory can it have and does it effect context window size in anyway? These are my 2 crucial cents relating to this new ‘memory’
Curious to know if gpt's new memory program will keep track of each of my projects; how much memory can be allocated for each project; how far back in each project they / it is willing to go to gather info about me; and / or do we have to identify - starting from now - which data we want it to add to its growing memory of us and what to discard.
You can also prompt GPT to ask you questions for the purpose of loading memory. Ignore something if you think it is prying, but I did it twice, and the questions were not compromising. And the difference in more frequent, favorable prompts is most welcome.
I find it super useful. But I'd also like an "attention"-based memory management, so ChatGPT will store all information it finds important, just like how the human brain does - but based on the tasks that are given to it, instead. Maybe that could be achieved already with the initial instructions? Or does it not evaluate every piece of information automatically to see if it's important enough to remember?
I just got it. Hopefully, it is the self-sustaining continuously improving model.
Its not out for us living in Europe jet but seems to be a dope update!! Looking forward to hiving it!
Morning. Anyone know how to get information from Excel spreadsheets. Will GPT4 do it. Thanks
I love AI and use ChatGPT a lot as well as other AI tools. I personally don't think its scary. We use it or we lose it. I love learning new things all the time. Also I love your channels Matt 🙏💕
There must be a limit it can remember, otherwise it will have to have a large context window. If not, then there must be a backend function that somehow searches the data on behalf of the AI. (Which is what my AI has been doing for 12 months now)
I haven't got it yet but I am based in Portugal! But as an author of novels, I reckon it's going to be super useful! I hope I get it soon!
Boa tarde :)
Digital Jiminey-Cricket
I had access to this about 3 weeks ago already. I'm surprised Matt only just got access, being that he has a paid subscription, and I'm just freeloading on ChatGPT 3.5. It kinda scared me at first because there was no notification about it. I had pasted a bunch of my code that the formatting was messed up on and asked it to reformat the code to PSR-12 standards, and it said "Memory Saved" and output the code. I was confused for a bit, but after poking around I found the memories tab, and it had saved a memory that I like my PHP code formatted to PSR-12 standards.
Can you please make a video about Hume ai 🙏🏾
The real value lies in ChatGPT in understanding how you prefer responses. So THAT should rather be what is retained so that you go through less iterations in refining answers with further prompting. Preferring results in table format for instance. In terms of knowing personal information, there isn't a lot of value FOR ME since I don't ask it about products or things to MY taste, but I'm instead after factual answers. There is a risk in an AI becoming biassed in a way if it only presents things that fit your preferences since you then run the risk of missing insights from things appreciated by others with different likings and views. It's wise to know what's happening on the other side of the rode also.
You are worrying about THIS particular chat bot possibly becoming biased? We are way past that point.
when you upload a screenshot of your youtube every day/week, will it update old information or will it create a memory for every screenshot? If there is no timestamp or similar for the memory that can get problematic for various information over time.
There is a timestamp, internally, even a date. It displayed as date, then memory
I have had memory in GPT4 since it was announced, but had no idea you had to manually save memories. Did I miss a notification message or introduction guide?
I got a little pop-up window like Matte showed in the video. I got access to memories the same day he did (4/26). Maybe you missed it, or maybe since you got it early the pop-up wasn't ready or something?
i like it , do i have have to tell. it to remember something or is it automatic
It seems like a lot of people got access to Memory last week. I'm guessing last week was more or less the common release date for most users. It seems kinda weird that they announced it so long before the wide release.
I for one welcome our new artificial "assistants".
they are not your assistants, they are your future masters ...
@@swojnowski453 🙄
I don't! 😢 Is nice to know your background in music, i played guitar, too! ❤
This feature will be indeed very strong if will work on each GPT created by user in its unique way
I imagine memories could unlock some interesting new ways of jailbreaking it. Depends on the technicalities of how it retrieves the memories.
I'm curious what the limits of Memory are. How much can we store before we start seeing serious problems develop? I haven't seen anyone push Memory to its limits yet.
I got a dynamic option now too. Where it choses between 3.5 and 4 based on the prompt
I see it this way: I am paying 20 USD every month, I want access to their best model every time. Because that is what I paid for.
@@CockmageLVL99 chatGPT is still limited. You only get like 30 messages every 3h. So the dynamic mode helps you save messages where possible
What are the data privacy/data security parameters when using this memory capability? Does it differ in Plus, Team, and Enterprise? How?
I also just got it yesterday. It's good but what we really need is a larger token window.
Yea the memory feature is bad and doesn't do what alot of people thought it does. There doing so much stuff except doing the real issue that 32k and 128k is just not enough.
@MustardGamings The memory feature is very good. But we do need a larger context window.
But a token window and a memory are two different things. A memory reduce the need for a large token window. As an example: I use AI as an assistant in updating a book with say 20 chapters. Then I want to keep the book in the memory and make it recall each chapter as I update that chapter. This way I don´t need a large context window once the book is in the memory. AI keeping a memory is SO essential for efficiency. Both for me as well as for the hardware that must handle my work.
@@staffanlundberg I didn't say it was the same thing but the reason I prefer a large token window over memory is because memory summarizes what was said, it's not word for word memory. Gemini has a 1 Million token window and I can put all of my business information in one chat and it recalls it word for word. Chatgpt isn't doing that. I'm not saying it's not good or helpful and the use cases are different but if I were to choose a larger context window is preferred over memory IMO.
I think this is a great feature for anyone wanting a super intelligent friend.
Would have been a good place to try editing the memory, just before 6:00, you could edit in "as of April 2024".
My gpt refuses to search by social network profile, and generally analyze them in any way, how did you do it?
Is that only available in PC? Can't find it on the android app.
Is this a GPT4 only feature or is it coming to the free 3.5 ?
I haven't played with it yet but I do tend to try to do experiments with interactive fiction games. I will be trying to see if I can get it to remember the behavior of multiple fictional characters, store events that happened previously etc. It might solve some issues of the story getting stalled and say suddenly turning bad guys into benevolent good Samaritans etc.
From what I've seen so far, I don't think there's any chance it'll come close to being able to do any of that.
@@vainezaiven6677 I have done tricks on older versions to do it somewhat ... but I had to prompt it to put in summaries with each output so things stayed fresh in memory. For this it might be a matter of pasting things to keep things in memory. I just don't know if the things in "memory" will drift away further into the session with these new features.
@@ronpetersen2317 The Memory doesn't last very long. It really isn't a true "Memory." It doesn't create a permanent database or anything anywhere close. It only holds about ten line items of information before it starts overwriting them. It's not a very useful feature, frankly. I'm tremendously disappointed.
@@vainezaiven6677 If it's still there then you can keep it in memory by prompting it to remember again by asking to tell you what it remembers. I will have to test to see if that will work or not. In my old tests with interactive fiction games I would prompt in to keep events in memory, charcters items they have etc. in a brief way of course. but get it to display them so they are still current. It may be a matter of keep teling it to remember. I will have to do a long test to see where it is at. But yah it needs to remember such things each time so it is always fresh with that sort of thing. That shouldn't be too hard now it at least records what it remembers.
This I super useful and very cool
you're not the last Matt, here in the UK I've still not got the option for memory on my Chatgpt plus account 🙁
helps build a bigger profile "for you" or "on you". I'll be using this because it is useful, but I wonder what we trade for these conveniences.
your boss, the gov, will know you better, like the pocket of his trousers, that's the price.
Using 3.5 is useful because it is easy over time to get accustomed to its biases. You can adjust prompts to account for them. Pressing beyond them. With v4 and memory, this becomes harder to manage. You never know when its answers have lost their objectivity and when your own biases, real or assumed, are affecting its answers. This matters most when using it to help with communications with humans. Humans want your communications to be devoid of bias and objective in some cases where AI is useful, such as in translations and in summarising information. Making adjustment for bias harder and less objective weakens this use case.
you dont have to use it
@@ryzikx For now. But it might be that zealous developers with nothing better to do will think they need not support optionality if they do not understand the use cases. Some devs love to deprecate things they do not understand in an API. It is less to maintain. Regression is the bane of software.
Just don’t turn off your brain and your good to go. I have the impression that there a quite a few people that eagerly want to do that.
@@bogartICQa It might be that since the late nineties memory was detrimentally affected by habits of searching internet making memorising less necessary. Similar side effects might happen. Already there are reports of code quality slipping. How can the promise be fulfilled without benefits being outweighed by costs? There needs to be a regression testable standard. Some testable assertions. Some criteria for ongoing success. I suggest we look at why AI was invented in the first place and do some metaphysical reasoning at high level. Ask AI what its purpose is. I invoke Plato, Aristotle, and Ramon Llull.
Live in UK. Got Plus account. Love to use it but obviously at back of cue for memory. I still don't have it! 😢
Unlikely as the UK keeps the EU directives
Is it only available for Plus subscribers?
@mattwolfe Do you think mobile phones apps etc. Will be around or available accessible in the future or for lifetimes in general??? Or will they go extinct completely in my lifetime I’m in my 30s??? Very curious to know what you think!!! 🤨
Q: when you tell it to commit to memory, is it only committing the last response or everything from the thread?
I work in Germany and so far I don't have this feature. I'm looking forward to testing it 🚀
I don't have the personalization tab... :/
Why is no one asking: what bands were you in Matt?!?!
Probably great for study.
So with this update it can learn & remember various things about (whatever) and if it is in regular mode does that mean it keep a record of (whatever) it tends to be that you commit to memory when using the Chat Gpt outside of incognito mode? Bc if it learns from what ppl add or remember about a specific topic or (whatever), then thats basically a personalized Wikipedia page on yourself. Which then translates over to the next version of "Social Media". Uh no, this can go bad quickly.... Social Media reinvented.
Is its memory limited to a certain number of tokens?
Pro tip: Chat GPT doesn't remember conversations before the upgrade, so download your dataset, split the conversation JSON file according to file size, upload to chat gpt and it will be up to speed. It really is next level.
How do you upload it to ChatGPT? Are you saying don't through a custom GPT?
OK, wow, thanks for the tip, I might just do this. Completely forgot the option was there
Yep this is logical. Good looking out.
Does it has some limit ? Cuz everytime I save more than 10 memories, it delete all of them
You’re definitely not the last. As of today, I still do not have access to that feature and can’t find any information from OpenAI as per when they intend to finish the rollout.
I want to know how to set this up on a local model
Back to watching Matt again
Cool feature. Except I ditched GPT for Claude. I can always create a bio doc on me and paste it into any chat I crave with Claude. And I can chat with Claude about updating it.
To say it is sort of like custom prompts is incorrect. Memories are global. Custom prompts are local. You cannot go back to old chats and have your custom prompt apply. Memories are kinda like RAG.
Can it use memory for PDF uploads? or memory for project based info and other text file related? I do not care much about if the AI knows about me personally but rather it can remember my project. Also, can it remember mistakes it made so it does not do it again? I am not subscribed to ChatGPT so that is why I am asking.
Update: I decided to subscribe again to "PLUS" to get the memory options. I must say it is working great. I am in development and the biggest thing I wanted is for AI's to have memory. On with memory can I use AI for my project. Now I can feed it docs and have it remember things to recall later in my project. I have been waiting so long for this day to come. Have a nice day.
I think this is the most important feature ever.
I think you are drunk on news
I have no idea, what to think about. Vendors like Meta already know everything, without providing that info back to me, so maybe it seems to be a good way to be implemented like this.
So if you put live data into memory, like subscriber count, does it default to that memory, so it will be wrong as soon as the count grows?
Unless you update the memory, it will remember the info given.
Seems it does not work on team accounts even though they say it should. I can use it on personal not team.
I still have no Personalization option in Settings today, and I'm in the US.
Doesn't seem to be available in the UK yet. 😞
in the UK nothing works, haven't you noticed yet?
UK keeps the same EU directives.
Oh Nice!
Here I am on April 28th and no memory. Still taking their time with the rollout, but they make sure they charge me on time every month.
Just got it now - or at least - just checked but for some reason mine wont fetch anything from the web like a twitter profile.
I love how a guitarist is now an authority on AI. What a time to be alive.
he is as much authority on AI as politicians are on solving problems ...
Jamaica says Thanks for all your efforts to keep us in the know and on the cutting edge of AI opportunities! Big Up!
ok
Now I have to remember to delete both my browser history AND my AI history before I let anyone on my computer
Isn't this like a modified system prompt including "memories"?