Best Cursor Memory Bank Feature & N8N Strategy (Not Documented)

hey there I'm Parker Rex I led technology and product for a startup that did well over $50 million a years in sales we were there the whole time helped scale it after that I did a bunch of different startup ideas in the SAS space tried to go the VC route it didn't work out the execution was good the tech was good but the market risk was too high for what I was trying to build but I've been using AI to build things pretty much every single day since it's been popularized since the SDK came out and so on this channel do daily uploads where I teach what I'm doing to build an AI Services business in public we go through some news we go through some build stuff some strategy let's jump into it first thing is Vibe coding can break a lot of different things and so I'm figuring out the best practices for how to avoid that we'll making a lot of content on that on a new channel called Parker Rex explains so let's get into the Whiteboard one of the questions that came in is on the different tool calls that cursor has so I want to just cover that real quick a tool call is something that an AI uses to basically have a function a very well- defined function that it can call because you don't want it to just have all sorts of Freedom if you wanted to do a specific thing so example of this in cursor is it has a read file tool which reads the contents of a specified file has a write tool which does what it sounds like it can write it can delete so the CR operators then there's code modification tools like editing doing parallel applies so similar edits across multiple files simultaneously a code based search so that's when it's doing a GP and it's looking around um I guess these are separate you have GP search and codebase search you have file search web search run command which does internal stuff and then reapply so some of those are the tools and you can Define these in your own code bases so an example that would be with perplexity they have a research tool they have bunch of different tools that make you get that experience when you type in a query and it comes back so that was one of the questions that I saw that I wanted to answer next up is LG company that typically makes televisions launched a model kind of wild to think about but it's going to be a trend like there's going to be people that just make more and more and more models it's uh kind of crazy but it's worth talking about because they achieve number one on a which I don't even know what that is so let's look up aim AI Benchmark highest performing benchmarks so this just looks like it the mathematics examination so it's really good at math sort of good at math O3 manying being the best rock not on here interesting Rock three at least but I don't know too much about it I just wanted to call that out because I think it's kind of crazy that people like LG cting stuff out I don't know if there's any examples of people using it oh that's interesting so outperformed a bunch of different models with just one 20th of the model size so these things are getting smarter they're getting smaller which plays into something that Apple might be doing longterm a lot of people are giving Apple crap but at the end of the day if you look at Apple they have this massive ecosystem that they've built so you have obviously the iPhone and you have have the MacBook have airpods you have the watch obviously other things in there but those are the main ones and then each one of those has access to a couple proprietary features that no one else has so airdrop one of them really easy file exchange you have the Bluetooth connectivity if you're in the e system you know that it's just better the handoff process between devices and the connection you have iMessage you have photos then you have iCloud so all these are kind of under this umbrella this whole thing basically oops is Under the Umbrella of the Apple ecosystem the Walled Garden so to speak and and it's really hard to get out of so a lot of people are saying like where's Apple when it comes to AI because they've chosen to they did a partnership with open AI which I don't know how well it's going but I've used it on my iPhone and it's kind of like you have to dig through a couple different uh folders to even find it or settings but what I think they're doing is really smart in a way because the amount of Capital that has gone into the foundational layer so in AI have the foundational layer which is the models and the actual like metal things that are powering everything like the big databases or uh data centers and the companies like Nvidia really benefit from this they sell all the stuff that everybody has to have in order to run their models and build companies you have the app layer which would be the actual things that people are using so like chat.com which is chat jpt or the grock app and then the third is De tools you these three different segments and the most expensive one and and by far most expensive one is this one right here so this one is really really Capital intensive like for example there was a project star something star base now it's there's a project that half a trillion dollars is being put to work to build out AI infrastructure structure and that would be just one of hundreds of Investments so trillions of dollars are going into this one right here and then what what's happening is in any new kind of paradigm shift so like when YouTube came online we spent a whole lot of money to be able to support videos streaming but then it got cheaper and cheaper and cheaper with getting faster and faster internet speeds so more people could upload but a lot of money went out up front so I think the way that you could kind of chart this is like if it's money investment then time there's a big balloon up front the cost and then what is kind of interesting is if we had like let's call it like uh Enterprise Value so the companies that are putting the investments in they're also ballooning but then as the price and the cost of the compute comes down so in this case like GPU and inference and all that stuff as that comes down so does the Enterprise Value typically unless they pivot unless they uh land and expand because things get commoditized so for instance like early on you know chat.com or chat jpt was the only player in the room so they've commanded you know a990 billion Enterprise Value or market cap but it's kind of in question like are they really a foundational company that's what they've been or are they becoming more of an app player company and I think they're more of an app layer company over time that's what their Enterprise Value High because companies like Microsoft who own 51% of chat GT or open AI are now kind of ridiculing the space saying these things need to make money they're not making money so what Apple's doing by not getting into this space is they're actually avoiding a lot of this they're avoiding this like big area of loss they're letting other people other entrance and other money go into the space to advance it and then they'll come out and you'll see this they did the same thing with the MP3 player so if you have like Zoom was Microsoft's I think versus the iPod you also saw this Blackberry versus iPhone you also saw this with Bose versus airpod max but then they come in and they just wreck it obviously um you know wireless headphones you can't name one set besides airpods and their business for airpods alone is massive like hundred billion dollars so I think that what they're doing is actually really smart and it plays into this thing like LG dumping a lot of into it they're just helping Apple in the end because they're showing what the best practices are and at the end of the day to bring it all home what they own is the iPhone which is what everybody's on in the US a lot of people are on globally and the chips that are in them they have like M4 that runs on a Mac but those will be able to eventually run these so you'll end up just connecting model directly to the phone and they're very PR privacy Centric privacy focused so the next up is Claud Max so Claud Max is something that cursor released and what it does we'll just read the docs so has Claude 3.7 thinking at its core uses the entire context window is a very high tool call limit so traditionally if you're using cursor you'll hit a tool call limit it's not good you have to hit resum but whatever but they're basically just maxing this thing out it's kind of controversial a lot of people are saying like well you dumbed it down and then the guy's like no it doesn't I just wanted to call it out because it is in cursor and it's a new a new thing so we wanted to take a look at it so we're actually going to use it now we're going to skip around a little bit because I have something that's broken on the website that I need to get fixed with some level of urgency because if Google hits that and there's broken links it will d rank me in search luckily I pushed this stuff at like 2 in the morning last night so it's fine so what I'm going to do is introduce another new system that's called the memory bank and this is put out by CLE kle if you haven't heard of it they are the innovators in the space of AI agents they invented the first pattern of the AI agent that recursively answers itself in an IDE and what it does is really special so this memory bank has a couple different files in it active context is the thing that when you're starting a new task because client is Task Centric it will go and read this product context is context about the product progress is what we did last project brief is what we're going to type in so we're going to put this we recently updated routes for the blog and now when I click on individual blog posts in the blog page it goes to Broken links we need to fix this so that all the blog individual posts work so I've put in something here more detail the better so what I can do is I want to just do a little bit of Investigation because a the difference between an 80% prompt like you know it's going to work most of the time right like four out of five times it's going to work the difference between that and then five out of five times is knowing how to do the task that you're getting it to do so what I would do is i' actually do a little bit of Investigation so I'd pull up the file and I know that this page is leading to something that's broken so files involved and already new that's great so let's take a look and see if we can't find a link tag in here link these are for the search queries let's make sure that works so let's open up this let's go to our blog we're going to type in prompt enter and I'm not sure even how this is working so that's an issue so in the search I'd want it to show highlights of the words that I've typed in there's another issue here so we have like a dual clear button so let's type in issues for project brief current bugs fix cool number one or I'll do uh context we recently did this okay we also added search functionality but it's not working as expected okay great so on this bug number one would be clicking on individual blog posts on the blog page go to Broken links so let's click on one goes to this something is happening in here after we added search where if I did a search it's adding this so let's see if we're going to put this in current behavior the uh what is it search bar reads that expected Behavior should just be this let's look at the slug so we'll actually go into the database check what the slug should be posts and slug copy that and let's test it out that's wrong too interesting it seems to be fixing things is that the slug is a slug okay so even that's busted I didn't expect that part so the slug stored on the database is that this may be an issue with our trigger the database so I had had a function that made or sorry I had a trigger that made slugs so what we can do is we can bring that in and the reason I had that was because I was seing the platform and frankly I just was trying to Vibe code part of it and I didn't read it so that's a mistake you don't want to do so let's look for slug log post slug well let's grab anything that has slug in it create new function up post status there quite a few things in here 18 results let's just tag you can see the trigger here tag the schema and then we also want to here's a snippet of the trigger we'll grab it so let's go to the first instance and I don't even think I'd need this I think this is causing more harm than health so put that in there put the tags around it okay great and then that's another file there any other files involved this blog blog page close these log slug that all seems good well so so we have everything in there now what we'll do is we're going to grab incline I'm not going to use it because I ran out of tokens but I'm going to use the same pattern with cursor so has this big custom instruction that explains what the memory bank is the structure of it all the like flowcharts explaining it you can find this in the links in the description actually it'll be on the blog post once this blog is working and I'll go over here and I'm going to say read your user instructions because what I did was I put them in here and user rules and I won't tag any files and then I'll hit let's try this Max thing let's try Max okay never mind we're not going to use that because I don't have it set up let's try this we're tag it because I couldn't figure it out okay so now it's going through and I'm wondering why I didn't pick that up but that's okay it's reading through everything it's getting everything loaded into memory and it should be able to knock this task out e okay so it found the issue that slug was doing the problem that was causing the problem because the ones that were seeded had that applied but not the ones that have been created after because I was just in lining them like in the gooey so this should handle the slug from the URL and then we need to update the generate metadata then we need to update this what generate static params does is when you go to build your project it will make sure all those routes are there ahead of time and so they're static and that matters because Google will have more pages to read from the get- code they'll be a part of the site tree versus if you didn't have it then it would not work you wouldn't have access to this okay so now we want it looks like we made a change to our SQL by making a new update so it'll fix all the records that makes sense oh and what we'll do is we'll just run this on gooey oh I need to get orb stock running second so let's start the server or the database so that's running okay it needs to apply it so that get that did apply it to our database created a temporary function it executed it and then it dropped it that's cool so now it should be running and it should have fixed those URLs locally so we'll run it locally and let's click on blog let's click on article Tada it works great so what we'll do now in order to fix this is we will do a commit which will push the Mig ation up now we're going to do super base DB push we like what happen we're going to push both actually we don't need the initial schema because it's already up so we don't have a migration history in here because it was messed up and I like having clean working stuff that's not confusing um so what I'll do is I'm going to actually rename let's cancel that name this schema.sql we're going to move it out of the migrations that'll be the base and then we're just going to push that one that we have yes cool so what that should do is hopefully fix them [Music] see might have to do this manually create and call it let's do it let's grab this and let's run it okay then let's check our database to tables or sorry table editor posts much better I think let's see now let's push these changes and that's going to take a second to build so we can come back to that in a sec we'll just pull up verel and pull up the deployment it should take about I'm assuming like yeah a minute and a half so we'll come back in one minute cool that's from yesterday next up uh okay okay these N8 and automations are things that I'm going to be selling into businesses because they're super helpful but I want to do it in a scalable way I want to do it in a way that I can learn the fastest rather but I can learn the fastest on the breadth of what's possible with them so what I found was there's a bunch of open- source ones so if I look in here if you're familiar with what one looks like um like I'll just pull up a YouTube video and I think it was let's go to our history let's find one okay this guy built Jarvis this guy built an assistant that does a whole bunch of stuff it basically like has a telam uh trigger and then it can talk to you it can update contacts and do uh like pull things from Hacker News it can do calculator tests it does a bunch of random things this is not a business use case but this stuff on screen here is basic basically just a Json file so luckily a bunch of people have open source these so I can find all of them and I want to be able to have all of these on a server so that I can have them for myself and do the self-hosted way but then also so I can know which ones are good which ones are bad so this is a great resource right like what there's how many folders in here let's call it like okay one two three four five maybe 25 in here 2530 a lot of different use cases there's another one in here it's like Advanced web researcher and all these again they end up looking like that one we just looked at so we had 30 in there we have another 40 in here there's 70 different automations that you can select from you probably end up selling these not even to businesses but I see people doing this all the time as like I guess you could use them as lead magnets but that's the plan is I'm going to start picking these off and hosting them see our build completed so you can see that we have this SSG so that's static site generated which meant that our generate static prams worked and that's what I was speaking about earlier where all these paths are completed like they're already built ahead of time of build so that Google can go and index them so let's see if our method worked let's grab our normal site the live one let's go to blog let's click on one all fixed so that worked great let's test another one just to be sure and we'd want to actually wrap tests around these but great we fix our issue so that was teaching you guys how to use the memory bank method within cursor and that is awesome so I'm really happy with that and I needed to get that done anyways you guys learned something there but back to this n8n thing so my plan with it is to basically have it won't be a $5 VPS because what I noticed is there's this meme where people top you know it's called the $5 VPS which is a virtual private server what it means is that there's some computer that's sitting in New York and you just get a little piece of one of those computers problem with that is if you run one nextjs app on it it will Spike the CPU to the point that it can't even build on it and you can get around that by building off of it and then uploading it but I just want to be Fe your machine so what I'll do is there's a company called hetner and they have a really cheap compute and then that'll allow me to run all my tools on it so it's a bfar machine I can run Parker rex.com on it I can run mapthe map.com on it which is just a marketing site but also then the app itself I can run all the n8n workflows on it and then I can run any sort of miscellaneous KRON tasks on it which is really cool so KRON is just uh basically something that's built into Unix and built into Linux the operating system level where it has a syntax where it's like timestamps and asteris so I can say like at 12 o'cl go do this task and then you can have it be associated with AI and having it run on a server is really nice because you can't really do it on yourself for cells got what's called On Demand compute and so each time that you want to do something it's basically got to like spin something up and that's what leads to these things called cold starts that everyone complains about and it's built on this Tech called lambas which is from Amazon so I think this is a good plan I think other people should be doing this too because I was just scooching around on the internet lesson I'm like holy crap all these are available on GitHub it's just the best um one other thing is I'm going to be having this Satellite channel for all the cursor news because it there's just so much of it so you should check that out and then lastly before we wrap up I just want to update on the content pipeline so if you've watched my videos before I've been talking a lot about how do you get the most leverage out of AI and how do you put the most amount of content out in a world where are going to be valued for their authenticity because there AI there's going to be so much crap so the system that I was thinking about is like okay well how do you get more surface area across all these different platforms with just one person and so I made this long list of all the steps that would be involved for taking one piece of content which is this and then turning it into every platform and so I made some progress on this yesterday until I had a ceiling with the account type that I have but um why do I have to log in again okay so you can see I already have exceeded the data transfer by 2.8x because I tried to do one video obviously need to change that but I started with this YouTube to blog post I can turn off what it did was it was watching for a file in a folder so when I make this video I'll save it to G drive and then the G Drive downloads it this is where it broke because and then it goes over to transcription so it's going to transcribe all the words after that will write a blog the blog prompt is very specific so it takes conversational transcription and makes a comprehensive tutorial blog and this is using best practices from HubSpot they're really good at inbound mark That's What content is the end of the day it's inbound marketing so it has all these things in it it pluses it up and then it's going to store that in a variable then we're going to have an excerpt writer because if you look at superbase or anywhere that you're storing your information most posts have an excerpt so if I look at my site I go to SL blog that's what this excerpt is and you can even see I'm missing an I here so it would write this and then it's actually going to write a title as well so it's going to write a title and it has a specific prompt and then it goes to markdown to HTML as the next modu and then finally it creates the row so this is something I did manually for the last two weeks it's just another 15 minutes but if I take one hour to automate it then by the fifth time that I've done it manually I've already I'm in the green so that's one part of the content pipeline that I finished and then the next one I was working on is let's see YouTube long to short so right now there's only two steps in it and it's they're going to be watching for a post that's already been processed from an audio perspective I use in a different automation when I post a video it's going to a third party called aonic aonic does two things one is it creates a transcription as well which which is then used to generate the chapter markers which are the things that let you navigate the video and then it creates a summary both short and long then it creates tags and then obviously creates a separate audio and video file and so it's doing that and then it's going to store it somewhere else which will be the process folder and then I want to take that and then do an HTTP call that will go to an app called descript Plus's we'll go to an app called descript descript does clipping so it will take the video that I send it and then make short clips for it and then I don't have to hire an editor for that so that will be something that we continue to do and this brings up to what I was talking about earlier about self-hosting is I don't mind doing the dev Ops I think that's what a lot of people are afraid of doing but I've done it before and it's better to know how do it then running into these walls so the last one that I'm working on for the pipeline is this Instagram theme page so what it does is I'll show you a quick example of what the end piece will look like but the question was is how do I make the image and text posts for Instagram without actually doing anything so if the end post this would be one that I've made right so it has a brand name has an image and it has text and then it has a caption and so all this will be on a channel that won't have video I'm going to have a different one for Clips so this would be another Satellite Channel but what it does is it looks at an RSS feed it pulls articles from that RS feed which is bundled by the keyword chat gbd prompt and AI news it classifies the news that's coming in so it acts as an Editor to pick which pieces make sense for what we want to do we want to have pieces that aren't about celebrities or non event like basically just like filler news so it's given 15 stories of picks the best three then it stores those article URLs because we need those later then it iterates through the list of the three it passes it off to fir crawl which is for optimizing llm readability it'll go and read the sites because we need to actually get the Articles the RSS doesn't do that it gives us headlines images descriptions URLs so then we go through and we scrape those and it takes any of the garbage out and it optimizes for readability for the M which we'll use to write the Articles then we'll iterate through those markdown files that we've gotten from fire call we'll do an artical summary because we need that for a caption we'll store the summary back to the associated URL then we'll generate an image then we'll update the whole thing so we'll end up with spreadsheet full of posts so if you have three of them being written every 15 minutes then what you three you get 12 an hour you get 24 hours in a day 288 times 365 you end up with 105,000 posts so I'm going to run this for a couple days just to see but that's kind of wild you had 105,000 across a year then it's really on you to understand how to make the best classification how to make the best summaries and how to make the best images and you would likely do this same process when you bring on someone to make custom art or uh maybe a Content assistant You' use the base of this this is fully automated I think it would perform better if it was like 75% and other 25% is bringing it home good content strategist but those are the things that are possible and I think it's just wild so I had a piece of fuzz on my chin the whole time hope you guys didn't notice that I'm also Rudolph the red noos reindeer but hope this is really helpful I'm going to be working more on this today and putting out more content I'm also building automation for upwork for clients doing client Outreach because I'm trying to work my way up it's somewh to YouTube where it's like at first you have no one and it's just like sets and Reps and you have no reputation then you continue to show up you build a rep on there and then you go to the top I'll be building one of those out and I think the title of this video will be about cursor and its memory bank so there's going to be a bank right here this is for the thumbnail give me some sort of Bank all right I'll see you in the next video oh and like And subscribe obviously do all the YouTube stuff leave a comment ask me a question so I can answer it see you