You NEED to know about Openclaw Context Window

This video is one of the most important videos you'll ever watch and it's about context for your open claw. >> Understanding this will turn you will turn your agent from a total dumbass to one of the smartest people you'll ever know. >> And this is kind of inspired by one of the comments from our viewers, Note. Thank you, Note, by the way. >> Amazing. >> Goed. >> Goed. and he says, "Uh, Mini Max works really well on low context. Like the moment you go above 100 20K context window, it feels like I'm talking to Chad GBT3.5." And you know, if I want to phrase him properly, I'm talking to a total dumbass. I'm hope I can say that on YouTube, but this is actually really true and it's something that we found and it's not really >> I think it's Open Cloud doesn't really tell you unless you ask it. >> Yeah. >> Right. I think that's kind of important. So what an experiment we did earlier is uh I talked to uh Jeff which is on Miniax uh 2.5 and I asked him you know what is your what is the context >> and I want to say that he's already crossed the threshold. >> Yes. [laughter] Which is crazy >> and uh a couple of questions here right because um Michael you taught me before that the context window automatically resets to zero every day. Uh but that's not the case because this as you can see >> it's 5 right it's 10 10 1051 it's already the start of the day his context window is already filled up >> so I think that's something that we really need to address and this is really important for lower end models and just [clears throat] kind of how these AIs work but the easiest way that you want to if you want to know what context how much context you're using because it's sometimes not displayed on chat sometimes if you're using uh open claw on terminal it will actually show you how much context you're using but you can actually ask it So just say how much context are you using? That's that that would be a good question. And then Jeff came back with this uh this result, right? So um this is quite important because um >> the way the best way to describe it is that the more context you're using, the dumber and the less reliable your AI is going to become. >> Yeah. >> Right. So there's that sudden switch >> where this AI is super intelligent. You can rely on him for everything, but at a certain point all of a sudden he just becomes stupid. >> Yeah. >> Right. As you can see, it's already the start of the day and it's already [laughter] being, right? And then I'm just going to go back over here to Stark as well. So Stark is running on Opus and yet again at the start of a day he's already u 100k out of 200. >> So there's some differences here I would say. So um we all have the what we're doing here is we really want to show you the results for you guys and really how to improve it. So let's just start one by one. So Stark is running on Opus and Opus doesn't really seem to have that intelligence drain at higher context, which is why a lot of times when we use Opus, it feels very fluid. It feels like you're paying for it. You're pay you're getting what your money's worth is from here. Yeah. >> But I think if you're using a lower-end model like um like Miniax, you really want to be be able to manage that context. Yeah. >> So, okay, let's go back up a bit and really say what is context, right? So I think u that's quite important for us just to kind of define it. I we have a slide here. It's not the best slide but uh um it kind of iterates our point uh the context the hidden power. This is a presentation made by our AI agents. Anyways so the way that these AI works is it kind of has a working brain. It kind of like you're like a human memory right of what you're doing. But at some point when you're giving even a human being like I've worked with assistants before I say I give like oh deal with my car deal with my house deal with my parents they're coming in today deal with this they they get overloaded and that's exactly what these ARs are feeling they're feeling okay their context is being overloaded they're being overwhelmed and then they become stupid >> right so that's kind of the best way to imagine context is uh how much they're kind of handling in their brain in their temporary brain right now until it's um becomes overloaded. >> And that was one of the misconceptions that I fell into in the first week was when Open Claw when I learned about Open Claw, they were talking about how it's memory is amazing. It's like unlimited, >> right? >> And I didn't know that they were talking about the memory, you know, the [snorts] >> save the memory long-term memory versus short-term memory. >> Exactly. Right. >> Yes. So short-term memory is what um decides how smart your AI bot is. The long-term memory is it can be stored. So it doesn't really matter that much. So in this video we'll talk mostly about context. Yes. >> Okay. So why does a context window? So it's kind of something that's interesting is that open claw automatically resets uh itself. So the best analogy for this is that it kills itself every day and it rewakes up. Then it reads your long-term files, your long-term storage and it loads that into the context. It's almost like a assistant coming to work. You know, they wake up, they're like, "Oh, yeah. Okay, who am I?" Okay. Um, I'm Jennifer. Okay, let's just invent a random name so I don't dox any people in the office. I'm Jennifer. And um, you know, I'm serving Ron today. Uh, and then they kind of read up these files that you saved up long term in the memory. We have a video on memory. Um, just go watch our channel, but we have um a video on how to optimize that memory. But it kind of remembers, oh yeah, I'm doing wrong today and here's some of my tasks. Here's what I'm doing. So this goes back to your question, right? When it wakes up, uh it's already loading up context, right? So um if you give a lot of information about who you are and the information that you're giving it is >> unnecessary, >> okay, >> um it's it's already going to fill up that short-term stage and it's not going to be very very well performing today. >> Okay. So, okay, that's kind of a one tip that you want to have. And we're going to make a future video on how to optimize uh there's certain files inside your bot, which is like your memory, your soul, your long-term storage. >> Uh you don't want it to read too much of the irrelevant stuff. >> I see. So, what you're saying right here is this is literally the first message of the day. >> Yeah, the first message of the day >> and it's already at 136K because it read all of the files. What? You gave me a lot of information because you know when when you started talking to Jeff, right, you were telling him your life story, you know, like maybe this is how I grew up, you know? Right. Right. I remember you were doing that. You're like, you had a two-hour chat with him >> cuz that's what the YouTube beginner uh beginner guide said in the first week when Claw was popping up. They're like, just blast it with your life story so it understands you better, what you expect. And yeah, I did in fact have hours of conversation with Jeff telling you my whole life story. So >> So Jeff remembers it, right? He wakes up, he's like, "Oh yeah, um you know, Ron's got a really pretty girlfriend or something like that and you know, maybe you want to buy her flowers and he's thinking about all these things. He's not focusing on work." Okay. >> Right. Right. That's that's what that's what that's what's loading up the context window. I see. >> Right. And then there's also something I want to say about context window which is like in theory the context window it's a short-term memory for the uh for the bot right for the agent for whatever >> but um to save costs >> uh what sometimes these models do in the background and I think miniax is guilty of this is that it just dumps certain parts of it >> which is why like say for example even when it's executing a certain task say for example it needs to make a presentation for me all of a sudden it might forget oh what presentation are we making? Are we making AI or making um crypto videos? It suddenly forgets. >> And this is not a fault of you. >> This is a fault that's um this is a hidden feature inside Miniax to save cost for themselves. They kind of just dump memory that they think is not important. >> But sometimes what they dump is actually really important. Yeah. Right. So this is why when you're using cheaper models, it does feel like it gets duper over time. And this this kind of dumping happens um quite aggressively on the cheaper models because they want to save as much power. Like I mean these Chinese guys are smart, right? They they want to save as much money as possible, right? For every $1 you put in, they they only want to do 20 cents of work, right? So they're they're finding ways to do so. So that's why the context window is really really important here where uh you want to especially when you're using Minia mags or even Kim or one of the Chinese models I would say try to reduce the amount of context you're going to give your bot right and you know the next question I can just see that in your head is how do you do that? >> Yeah right >> actually I just looked at it yeah look at [laughter] that input token 2.2 2 million, I would put 3.5K. >> Yeah. >> Tokens. >> Yeah. He's He's thinking a lot. He's thinking a lot. >> And we just did a comparison just now, I think. Um and um can >> And this is just a comparison between Miniax and um Opus. So Opus is just really good at working with large working with the full memory that it's got. And it actually knew more. I've trained Opus more towards work tasks. Mhm. >> So, when I trained my bot, um, I was very focused on not telling him who the I am. I don't, you don't need to give a Okay, I don't know if I can swear in this video again. You don't need to give a about who I am. I don't, you don't need to care my education, my life, what the whatever, right? All you need to know do is make a good presentation for me, right? That's all you need to do. So, that's kind of why Stark, he wakes up and he's he's already done a presentation for me and he's his context is still not overloaded, but I'm still mindful of the context. I can still ask him about the context and then he made a presentation for me and he he delivered I think a lot better than your one. I don't want to compare sizes here but he made a full on presentation for me um that I can access on the web, right? So that's a little bit better than Jeff who just typed up a MD document which is like >> and he didn't even send it to me, right? Yeah. >> He didn't even send me the document. He just said, "Hey, you know, this is what's really important for for for context." And I would say his presentation was useless. It it really didn't didn't do anything. This is this is his presentation. Just sessions and it's very like boilerplate. >> Yeah. >> All right. So when I trained my bot, I was very specific on what I needed. I said, hey, you need to do your X research. He remembered that. All right. Now in his skill, he knows, okay, you need to do research on X. He needs to do research on the web. He needs to compare models. He he tried his best, right, with the prompt that I had. But uh yet again training your bot for a specific task is much better than for it to be a general bot. >> I see. So the context you provide cuz contrary to popular like okay let's say cuz we were um as beginners we were taught that provide as much context as you can so they can understand the situ situation. But from what we have observed is you need to also uh limit the irrelevant context. >> Yes. >> Yes. Exactly. And I think I I was watching a video yesterday um on the way back and they were saying how um AI even though they're trained to be general AI, they like to specialize. So uh they gravitate towards specializing in one particular task. So that's why for me on stark side, I really wanted to train him to be okay just really good at making presentations right now. Right? I don't want him to manage my dinners. I don't want him to manage my lunch. I don't want him to order my lunch. I just want him to make a good presentation if I can say that. I think it's later on in the video. I just want him to make a good presentation for me um on that respect. So that's that's why it's a lot a little bit better. And yet again, I can tune this. I can make it better. And the other way you can do it is can you uh clear your context. So that's the other tip that you really want is that when you're approaching that limit, if you feel like your bot is getting stupider, clear your context, right? I mean that's that's that's an easy way. Just clear context, dump your brain. Um um oops they can't actually manage uh if you want to okay start session just start a new session >> and [clears throat] he'll tell you how to do it. Um there's a natural compacting stage. So the way the natural compacting works is that >> uh it remembers all these things and it's actually very funny. Our brain actually works the exact same way too. >> So what your brain actually automatically does is that let's say you do a lot of things this morning. Oh, and but you maybe had a really great drink or you had a really great um video that we produced. Your brain actually automatically says, "Okay, this is the most important thing you did this morning internally." And then whether you brush your teeth or not, I don't think you remember, you know, how you brush your teeth, how long you spent brushing your teeth, what toothpaste you use, you you automatically clear your context. You automatically compact your context. And that's what the AI is also doing. Just that it's not doing it as well as humans cuz, you know, it's it's not a human, right? So yet again, um that's is actually quite unfortunate. Um we cannot actually restart it. So this is a good question. We we should actually make a video on how to restart context, but you have to do it every day. Um on that respect, but oh my, my lunch is here. But um anyways, with that with that guys, no, [laughter] but to just finish um up on this uh you actually we'll probably do a video on this. Uh there's a UI. I think you should um that's open call gateway. You can do it by running. Oh, so you have to actually copy this into your gateway to run it. So, we'll do we'll we'll do this. Uh, actually, can you run the command? Let me see if I can really force him to run it because yet again, he he's trying to tell me to do it. I believe he can do it, but he just doesn't want to. >> Okay, this is where I feel feel like he's gaslighting me a little bit. I just want to see the result. But, um, starting a new session means he you you kill him >> and he comes back to life again. he reads the uh his memory and then he has a much clearer context. >> I see. >> Yeah. So, I think that's around it for what we want to cover. I just want to see if this works. Uh last bit of information, of course, um our videos, our channel is growing really well. Thank you guys so much for your support. We actually read everything that uh you guys post. So, if you guys have a comment or question about this, especially because it's a large topic, leave that down below. Or if you have any results for other bots, we we have been focused on Opus um for me and Mini Max for you. and then uh we'll play around with that and we have any for questions just tell us. We actually want practical uses on this channel. I think that's what is actually missing in this space. So anyways guys uh let me just lastly check if uh if this is what Stark's back. >> Uh no he doesn't. So we'll have to make a video about how how to clear insect. So >> yeah. So with that guys um thank you guys so much for um watching and um you do a session cleanup. So maybe a session clean up will help, >> but um I'm disappointed he couldn't do it. I'm pretty sure he can if you really force him. Just like, you know, run the bash command. >> But anyways, with that guys, >> but it's a big revelation. So >> it's a big revelation. All right, with that guys, thank you guys so much for watching this video. See you guys in the next one. Peace.