🚨 Installing & Building Out My Hermes AI Agent LIVE 🚨

Added: 2026-05-09

[00:00:04]All right. All right. All right.

[00:00:10]Okay.

[00:00:15]Should be good to go across the board here.

[00:00:30]All righty, guys. What's going on, everybody? Just going to let a little more time pass here really quick.

[00:00:40]While we're waiting, what do we think of the new setup here? We got the nice backdrop here. Um, I'm liking it. I'm liking it to say the least. I will not lie. Super super clean. Super clean.

[00:00:53]What's going on everybody? Thank you everybody for coming in.

[00:00:57]Let everybody else populate.

[00:01:02]And uh let me actually pull this up really quick.

[00:01:16]And if I do this. Oops.

[00:01:22]Clutch.

[00:01:26]He said, "I made it. Let's talk Hermes."

[00:01:28]Appreciate you. Appreciate you, Bean Boy.

[00:01:31]Appreciate you, man. All right. So, I got a couple accessories and things to utilize on this live stream today.

[00:01:40]Specifically, we got Elgato Stream Deck.

[00:01:42]So, that way I can just boom, boom, boom. Switch those really quick like that. You see what we did there?

[00:01:49]Uh, also I am going to make this a little smaller. So we can just see how this looks.

[00:01:57]Make this a little smaller here.

[00:02:00]Yeah, that's better. Much better. All right.

[00:02:06]So yeah. Yeah. Well, these are not just like acoust these are not just like uh regular like black wood panels. They're acoustic panels more importantly. Um so they should have helped with the audio which they are as it seems. So I like that. That's nice. Um so everybody grab your coffee, tea, Coke Zero, whatever you drink. I got my tea here because I'm not feeling good at all.

[00:02:40]My throat definitely hurts. It's sore. I I don't know what I I got was super sick yesterday. Super sick. Congested.

[00:02:48]And uh throat was kind of hurting a little bit. Worse than today, for sure.

[00:02:52]But the sore throat isn't as bad, but it's still pretty bad. So, we have uh we have our tea. Okay. We've got our nicotine. I ran out of pouches, so I had to grab this. Don't judge me.

[00:03:07]and uh we're going to get started. So, let's uh let's let's go ahead and get into this, man. So, what we're going to do today is we're going to set up our Hermes agent today. I've tired of like not having this set up. So, we're just going to go right in and set this up.

[00:03:23]Okay. Um so, as you guys can see here, I've got different screens that I can switch to. Obviously, this one specifically is the one we're going to really be on. Um and uh we're going to set up Hermes today. All right. We're going to set this up. We're gonna get it fully functioning, ready to go. And, uh, I'm really, really excited. So, uh, we'll give it maybe a couple more minutes, just let everybody kind of come in here, everybody who was maybe wanting to see this or whatever, uh, we'll kind of just shoot the [ __ ] here for the next two to three minutes, and then, uh, we're going to go ahead and get started.

[00:03:53]And, uh, just so you're aware, let me actually, um, I need to I need to SSH into my terminal anyways. So, we're going to have to do that. So, let me uh let me just SSH into my into my stuff really quick. Anyways, while we're waiting, I had to pull one of my screens over.

[00:04:14]There it is.

[00:04:17]There it is.

[00:04:19]Yeah. Let me just SSH in here real quick. All right. So, I've sshed into my Mac Mini, which my Mac Mini is where Obviously, I run my uh OpenClaw agent.

[00:04:35]So, we're going to go ahead and we're going to dive right in here. I'm going to go ahead right now and we're just going to get started here, guys. Okay, we're just going to get started. So, what I'm going to do is I'm going to go ahead and grab the Is it bad that I don't even know where to grab the bash? Let me Ah, here it is.

[00:04:55]Okay.

[00:04:57]So, we're going to go in here.

[00:05:02]Uh, all right. We're going to go in here.

[00:05:06]So, we've got this. We pulled that directly from the Hermes site.

[00:05:11]Says, "I know that Hermes is supposed to allow you to switch between LLMs. Is there any way for it to do that automatically to save tokens for specific tasks?" Um, I believe there probably is a way to do it digital. Um, but I I per I purpose, holy lord, I probably wouldn't have it do that automatically. It just creates more complexity. Um, and it could give you much more errors and issues than you ideally would want. Um, so I tend to stay away from that, but obviously to each their own. All right, guys. Let's go ahead and uh let's go ahead and uh get this bad boy started here. Okay, let's go ahead and get this bad boy going.

[00:05:48]So, we're going to set this up and make sure this is set up correctly.

[00:05:55]Let me see if I can make this a little bit bigger for my just for my own just for my own uh selfish purposes here.

[00:06:10]Let's see.

[00:06:14]If I'm a little quiet when I'm doing this, I apologize. I'm just kind of just making sure everything runs good and I'm paying attention to my screen over here to my left. So, apologies if I'm a little quiet, but we should have it installed and then we'll go through everything. I'm going to set it up with my uh GBT. So, we're going to obviously ooth in. I have obviously a pro plan on GBT, so we're going to do that. Also pretty interesting to know that you can uh I think it was on the 5.3 update they fully allowed the OOTH in for Open Claw to uh for you to ooth into your GBT account. So now if you have like a pro plan, you could spend $100 to $200 a month and you could have pretty high limit rates uh usage rates on your open claw. So that's pretty cool. Um, and you can utilize GBT 5.5 without having to utilize um without having to utilize the API key. So, that's cool.

[00:07:09]I'm doing that for one of my clients.

[00:07:11]He's pretty [ __ ] happy about that.

[00:07:12]I'll say that much.

[00:07:17]All right.

[00:07:19]You know, I will say Hermy's in the in the uh terminal just looks just looks nice. Would you like to uh Okay, open call installation detected. Would you like to see what can be imported? Absolutely not. I would not like that. No, thank you.

[00:07:35]We're going to go with the full setup here.

[00:07:44]All right. And then we're going to go with I'm actually going to switch out of here because I got to I got to obviously plug in my u my chat GBT stuff. So, we're gonna just not do that right now in front of you guys. We're going to go ahead and go to um where is it at? Where the [ __ ] is Open AI at? Am I tweaking right now? Am I absolutely tweaking right now?

[00:08:12]Okay. I was about to say I was like, "What in the [ __ ] is it at?"

[00:08:18]All right. Let me go here.

[00:08:22]Let me just sign in really quick.

[00:08:27]Um, we're going to sign in here. Sorry, guys. There's just codes and stuff like that. So, I don't want to I don't want to leak any of that. You know what I mean? You know what I mean?

[00:08:43]Okay. What?

[00:08:47]Oh, okay. I need to I guess I need to change something in my security settings. had no idea.

[00:08:52]Uh, what do I need to change in here?

[00:08:57]Run the codeex.

[00:09:00]Okay. Yes, we need to change that. There we go.

[00:09:04]Okay. Maybe do a little refresh here.

[00:09:07]Didn't know that I had to change these settings, but okay.

[00:09:11]Uh, yeah, there we go. Okay. Let me do refresh again.

[00:09:20]Let me just recopy and paste this. Hold the phone, folks.

[00:09:24]So, apparently you have to change the uh Apparently you have to change a setting inside the back end. So, you could do this.

[00:09:34]There we go. Now, it'll work. Let me put that code in, chief. Yes.

[00:09:41]Oh, duh. Put the wrong [ __ ] code in.

[00:09:44]Was this amateur hour over here?

[00:09:49]Okay, let me come here. Boom. Boom.

[00:09:51]Continue. Now I may close this page.

[00:09:56]All right, perfect.

[00:09:58][snorts] All right.

[00:10:01]Now, let me go back here.

[00:10:04]Let's go ahead and uh I'm going to get past this really quick.

[00:10:09]Well, actually, no, because you guys, even if you have my code, actually, no, I can't do that. No, I can't show you guys that. If I do, then you could take it. Can't do that. We're gonna go on the GBT 5.5 model, of course.

[00:10:24]Okay, now we can switch.

[00:10:28]Now I can show you guys again. There we go. All right. Um, okay. Select the TTS provider.

[00:10:40]We'll set this up later. Honestly, we're going to set this up later.

[00:10:45]We're not going to do this right now.

[00:10:47]Uh set select terminal backend.

[00:10:52]We're going to do a local as always.

[00:10:59]And of course that was there. Golly.

[00:11:06]Uh okay. Enable pseudo support.

[00:11:11]We're going to hit no there.

[00:11:17]for a install, etc. You know what? We'll hit yes. Let's enable some pseudo support. How about it? Huh?

[00:11:28]Wait.

[00:11:34]Okay, there we go.

[00:11:38]There we go.

[00:11:40]Okay, guys. I think now I can finally have you fully back in here and with working with me on this here. There we go. Okay. My apologies.

[00:11:52]All right. Um, [music] tool progress display controls. How much tool activity is shown? The CLI in messaging. Show every tool call. Show tool name. Oh, no. We want that off.

[00:12:07]We want that off.

[00:12:11]Can I just do I type in just off here?

[00:12:13]Yes, I probably do. There we go. Okay.

[00:12:17]Uh context compression. Okay. This is the context compaction.

[00:12:23]All right. It's set at 0.5. Okay.

[00:12:25]Automatically summarizes. Okay. This is just we'll keep it at 0.5 for now.

[00:12:31]Session reset mode. Okay. Okay. Uh this is the recommended reset whichever comes first.

[00:12:42]Uh daily only reset at a fixed hour.

[00:12:44]We're just going to keep it at the we're just going to keep it at the current settings. We're not going to touch anything yet.

[00:12:50]And we are going to go with the telegram.

[00:12:54]And when it comes to the tools for the CLI, well, what do we not need here?

[00:13:05]Okay. Okay.

[00:13:08]Yes. Yes. Yes. Yes. Yes.

[00:13:12]If you guys see me doing anything wrong, you better tell me this. You're getting like the raw first time I go through this. Like, you're getting it just as it is. Just so you guys are aware.

[00:13:26]I don't need image generation.

[00:13:29]I do need that. Yes. Yes. Yes. Yes. Yes.

[00:13:36]Okay, we're good there.

[00:13:39]Choose a provider local browser.

[00:13:46]Keep defaults. We'll configure this later.

[00:13:51]All right.

[00:13:56]We'll do this later. I don't want to do this right now.

[00:14:05]Okay. Select search provider.

[00:14:08]I would rather use Tavi.

[00:14:13]Oh, I don't have the [ __ ] key right now. Gosh darn it.

[00:14:19]I mean, I will use it, but like I'd rather not do this right now. You know what I mean? I mean, [ __ ] it. We're We're setting up the whole thing anyways. You know what I mean? [ __ ] it.

[00:14:27]Let's just Let's just go ahead and do it. How about it? Huh? Let's just go ahead and do it.

[00:14:34]Let's set this up real quick.

[00:14:40]Okay. I need to switch out so you guys don't obviously see that API key.

[00:14:44]So, we're going to switch out.

[00:14:48]Let's copy that.

[00:14:51]Paste that.

[00:14:59]Are we Are we Are we smoking gun right now?

[00:15:07]Copied.

[00:15:10]Paste.

[00:15:13]Okay. It really doesn't want to let me paste my stuff in right now. That's absurd.

[00:15:20]What do we Yeah, that's odd. Won't even let me paste.

[00:15:27]Okay, we'll do the rest of this later.

[00:15:30]So, I can bring you back here.

[00:15:34]This is going to be interesting, man.

[00:15:35]This is g This one's going to be real interesting. Let's Yes, let's launch this [ __ ] now.

[00:15:43]Launch me. Launch me, Hermes.

[00:15:49]All right. Invalid argument. What does that What the [ __ ] does that mean?

[00:15:55]Okay.

[00:16:02]Now, the main thing I need to figure out here is how do I get this started?

[00:16:14]Now you guys are getting to see how this goes live for me. Holy [ __ ] All right, let me go into here.

[00:16:22]We're going to go into Hermes setup.

[00:16:24]What does this do?

[00:16:28]Let me make sure that obviously I don't uh leak anything. Give me one second.

[00:16:35]Come inside the Mac Mini.

[00:16:42]But we need to do my question now becomes what do we need to do to get this going?

[00:16:58]Okay, let's do help.

[00:17:04]No such file or directory.

[00:17:11]You guys ever run into this when you guys are setting up your armies before?

[00:17:13]This is interesting.

[00:17:24]Me uh let me actually change this to this so you guys can actually see as well. I wonder if I can just do this so you guys can see the full layout.

[00:17:35]Take my off the screen.

[00:17:41]This doesn't make any sense. See, open claw was so much easier.

[00:17:52]I'm going to look something up.

[00:17:58]Okay, let's do this.

[00:18:03]Is it hasn't found it. That doesn't make any [ __ ] sense. We just watched the install happen. Are we tweaking or what?

[00:18:08]Am I Am I just dumb?

[00:18:11]No worries, Nick. Thank you, bro.

[00:18:12]Appreciate it. Yeah, I just That's what I'm saying. Uh Bean Boy, I just tried Hermes dashboard. It's It's literally just saying nothing's even found.

[00:18:21]It's literally saying command not found.

[00:18:23]Do I need to Dude, Bird Boy Kids, I agree. I'm about to like I I don't understand what I just did. So, I just went through the whole installation.

[00:18:39]I didn't obviously merge everything over.

[00:18:43]Even if I do help, no, nothing in the help. I'm obviously sed into my Mac Mini because of this. So, I'm I'm I know that I'm not dumb here.

[00:18:54]Like, let's try [clears throat] that.

[00:18:56]Yeah, let's try that. Bean boy.

[00:19:03]No, it's not even finding anything Hermes-wise. I wonder what happens if I run the if I run the uh the install again.

[00:19:12]I wonder what happens if I do that.

[00:19:16]It doesn't make s cuz like it shows it here, you know, like it shows it.

[00:19:23]Do I need to like kill it and restart it?

[00:19:28]Usually it would just when I've watched people do this, it just jumps right in.

[00:19:41]You know what I could do?

[00:19:43]I could use codeex.

[00:19:46]I could use codeex to figure this out.

[00:19:49]Uh let's try that.

[00:19:53]Let's try that.

[00:19:56]Give me one second here.

[00:19:59]I could do that.

[00:20:13]Okay.

[00:20:16]So then we could do codeex and then we would just do uh let's just do y let's do is Hermes.

[00:20:32]Yeah.

[00:20:40]Yes. I trust Oh, now you guys are going to get to see the real juice here.

[00:21:16]Okay.

[00:21:21]Write tests for file name. No, I don't want to do that. Is it Okay, I'm going to just do uh All right, let's see what this says.

[00:22:02]>> [snorts] >> first time as well. Just troubleshooting it this way. This is super odd.

[00:22:17]This is why I wanted to do this live because I feel like a lot of people can learn from this when they especially when they rewatch it on the back end.

[00:22:30]You were right. You were right, Bean Boy.

[00:22:43]I don't I want to exit. Let me exit out of here.

[00:22:49]>> Okay, hold on. I got an idea. I have an idea.

[00:22:57]It's not.

[00:23:03]Okay, we're going to try this.

[00:23:06]We're going to try this. We're going to try Can I move this camera?

[00:23:14]Can I move this to a different area?

[00:23:15]Like, you guys can't see that.

[00:23:19]If it's down there at the bottom like that, it's kind of driving me nuts. And if I was you guys, that would drive me more nuts, right? It's not allowing it.

[00:23:28]Okay, whatever. Let's do Hermes Doctor D-fix.

[00:23:43]Hermes itself is working. The remaining actionable issues is configuration installation. Hermes set up.

[00:23:51]Yeah, I'm not I don't care about any of that.

[00:23:58]Okay, we're out of there at least.

[00:24:00]Jesus.

[00:24:04]Oh, we can keep going.

[00:24:15]Okay. So, then let's do Bro, this [ __ ] is going to drive me absolutely nuts. Oh my gosh. You guys have no [ __ ] idea.

[00:24:30]Okay, quit. Okay, now can I exit this [ __ ] please?

[00:24:37]Done. Okay, thank you. But I'm still in here in GBT55.

[00:24:44]This is not the day. This is not the day to [ __ ] with me right now.

[00:24:49]This is not the day for it.

[00:24:56]Okay, you exited, but I'm still seeing you there.

[00:25:13]This is going to drive me nuts, guys.

[00:25:14]This is going to drive me nuts. It's going to drive me absolutely nuts.

[00:25:20]It really is.

[00:25:23][snorts] >> Okay. Yeah, we've set all this up. I'm literally looking at all the stuff that it's telling me to do or that it's like going through in the background.

[00:25:42]And I've done all this already.

[00:25:45]I did the whole [ __ ] setup.

[00:25:51]>> Hence why we're here.

[00:25:55]>> Oh yeah, this is going to drive me nuts, guys. You guys are in for a rude awakening. You're going to see me get real real real irritated real quick.

[00:26:03]Real irritated real quick. [snorts] >> You know, maybe this is why no one does this [ __ ] live. Maybe this is why Let's see.

[00:26:27]Okay. Would you like to run the following command? Do you want to stop the stuck interactive Hermy setup process?

[00:26:38]>> Yes.

[00:26:52]No, I Okay, it doesn't still wants me to do the same thing.

[00:26:59]I'm just going to push this [ __ ] through.

[00:27:14]Okay.

[00:27:19]Configure missing API keys for full tool access. Okay.

[00:27:24]Okay. I just want to exit this. I want to exit. I got an idea. This is going to drive me nuts, but I'm going to do it anyways.

[00:27:32]Yeah, it's my dog. My I have a French bulldog. She is absolutely out like a light right now.

[00:27:39]How to exit.

[00:27:44]The fact that you guys can hear that is kind of crazy.

[00:27:47]Uh, okay. Let's do this.

[00:27:52]Thank you. Holy [ __ ] Hermes doctors.

[00:28:00]There's Am I tweaking right now? Am I absolutely tweaking right now?

[00:28:07]What in the [ __ ] is going on?

[00:28:13]Holy [ __ ] It doesn't It's as Nothing with Hermes is there, but I just told Codeex to do it and it worked, bro.

[00:28:29]Okay. Okay. Let's think. Let's not be let's not be completely let's not let our our anger hurt us here. Let's see.

[00:28:42]Okay. So, nothing will go through whenever I type in anything with Hermes.

[00:28:47]Even if I did I'm spelling it right.

[00:28:52]Nothing with Hermes again.

[00:28:56]And I and I'm literally looking.

[00:29:00]Okay.

[00:29:01]Yep. Even that is still not working.

[00:29:07]Still not working. I don't understand what what the deal is here.

[00:29:15]This might be a failed stream, folks.

[00:29:17]This might be an absolute failed stream.

[00:29:18]I might just have to go into a Q&A because this is just pissing me off.

[00:29:26]All right, let's do nothing with Hermes found. I I I genuinely I genuinely I genuinely am going to lose it now.

[00:29:38]Mm- Oh my god, dude. Okay, so I don't know. Anything I do with Hermes, it doesn't show up. It doesn't show up.

[00:29:55]Now the question is is dude don't turn around. I just tried to have Codeex fix it and Codeex just told me that it's working and that it's there. Now my question is is even if I was sshed into my I wonder if it's because I'm even though I'm sshed into my [ __ ] if it installed it on my main like on my laptop like what I'm streaming off of right now. So let me let me do something really quick.

[00:30:25]Okay. No. So, it wasn't that. Okay. It wasn't that. I thought it was. Okay. I'm going to do this really quick, guys. Let me SSH in.

[00:30:35][snorts] Let me SSH in. Damn. I can't believe you guys can hear my [ __ ] dog snoring right now. That's absurd. She is a loud [ __ ] snorer. I will say that.

[00:30:43]It's kind of ridiculous.

[00:30:48]Let me uh Okay, that shit's [ __ ] annoying. There we go.

[00:30:52]Let me do this.

[00:30:55]Let's go here.

[00:30:58]Okay, we're sshed in. Okay.

[00:31:01]And now I'm going to go back so you guys can actually uh get angry with me potentially.

[00:31:08]Yeah, [snorts] I'm about to I'm literally about to just rerun it again.

[00:31:11]Um there. Okay, now it wants to [ __ ] work.

[00:31:18]God damn. All right.

[00:31:23]I just had to get out and get back in, I guess.

[00:31:28]I don't Okay, I'm logged in. Log on in.

[00:31:31]All right. So, let's do run Hermes setup to configure the Okay.

[00:31:39]Okay.

[00:31:41]Okay. Okay. Here we go. Jesus, sweet mother of Mary and Joseph. All right.

[00:31:46]All right, let me get out of this just to make sure obviously don't leak nothing, nothing crazy.

[00:31:51][snorts] Use existing credentials.

[00:31:57]Okay, we're going to current keep using that. That's good. That's what we want.

[00:32:03]All right.

[00:32:06]EdgeTS, we're going to go with we're we're going to keep it on the edge.

[00:32:16]We're going to keep it local.

[00:32:20]Max interactions. Yes. 60. Sorry, guys.

[00:32:24]I I know that I'm not showing you my screen right now, but it's just because I just don't want to I don't want anything to pop up that shouldn't off.

[00:32:32]Okay, that's what I wanted. Yes, that's where I was at. Yes, that's fine. That's just Yes.

[00:32:45]Telegram.

[00:32:47]Okay, I can't show you guys my bot token, so we're going to have to get out of that for a second. We'll create a new uh a new bot.

[00:32:55][snorts] Okay, we're going to name that guy Hermes because he's a G. Hermes.

[00:33:05]Okay, someone already took that username. That's absurd.

[00:33:12]Oh yeah, these are really moving. Okay, Hermes.

[00:33:19]We're actually getting somewhere now. Oh my gosh. I can't.

[00:33:23]Jesus, man.

[00:33:28]All right, there we go. All right, now let me copy the bot token.

[00:33:36]It's so wild. It's not letting me input.

[00:33:39]It's not letting me copy or paste anything in at all. Like, I can't paste nothing in there. I can't even type.

[00:33:48]Like, it won't let me type [ __ ] in.

[00:33:50]That's what I'm not I I I don't I don't understand.

[00:33:53]Like, I I genu that's what's irritating me because I can't [ __ ] put anything in there.

[00:34:01]It's not letting me type anything.

[00:34:05][snorts] And even if I like right click and paste, doesn't do that either.

[00:34:12]Yeah, still not working. Oh, it did. It did. There we go. Beam boy. Much obliged.

[00:34:24]All right. Um I'm gonna I'm going to share a screen in a sec. Don't worry. Don't worry. Uh we can do it now actually because you still can't even see my bot token.

[00:34:35]Okay. Allowed user IDs.

[00:34:41]Fire Telegram.

[00:34:43]God. Find your Telegram user ID message.

[00:34:49]Okay, I know my Telegram user ID. Let's get out of that.

[00:34:54]Hold on.

[00:34:57]Okay, now it lets me see this one.

[00:35:06]Okay.

[00:35:08]Sorry, I know you guys can't see this yet.

[00:35:14]Okay.

[00:35:20]Uh, use your username as the home channel. Yes.

[00:35:31]Yes. Install the gateway as a launched service. Runs in background. Starts on boot. Yes.

[00:35:40]Yes.

[00:35:44]Configure CLI.

[00:35:50]Do I do I have to continue this?

[00:35:54]Launch Hermes chat now. Yes, we are in. We are in, gentlemen.

[00:36:05]We are in, gentlemen. Holy [ __ ] Holy [ __ ] All right. after I almost just broke my [ __ ] Max Studio display and 15 Nick hits later, we're here.

[00:36:26]Holy [ __ ] All right, that only took 36 minutes.

[00:36:40]That only took 36 minutes.

[00:36:45]Oh my gosh. All right. Now that that's done, now let's go in. And God, my I wish you guys could know how much I feel like [ __ ] right now. Holy. All right.

[00:37:02]Hey, what can I help you with? Well, you can help me with a [ __ ] lot right now is what you can help me with. Um, can I move this so that way you guys can actually see what's being typed out? I feel so bad that you guys can't see that, right? Like you guys, let's just make it big screen for you.

[00:37:22]How about it? Huh? We'll just remove my ugly face here for a minute.

[00:37:27]Okay.

[00:37:29]What can I help with? Well, what do we do now, guys? It's fully set up. Um, I want to wait, let me just do am I able to chat with you inside of Telegram? Now, please do don't disclose any sensitive information as I am live streaming. Say hey to the live stream.

[00:38:05]Better be respectful.

[00:38:12]At least he's respectful and says hi.

[00:38:17]It doesn't appear connected, so I can't chat with you yet from here.

[00:38:22]Okay, what do I need to do in order to make that happen?

[00:38:32]I love how it tells you though, like you are how close you are to context compaction. I absolutely love that [ __ ] Okay, let's [ __ ] relax. How about that, Hermes? Geez.

[00:38:44]All right. Yep. I already did that.

[00:38:50][clears throat] Literally pro. I literally already did this.

[00:38:55]Literally already did this.

[00:39:00]Let me do something really quick.

[00:39:08]Okay. It's not. Yeah. All right. So, we'll set that up. What's annoying me is I and I wonder if it's because I'm sshagged in.

[00:39:19]Dude, I'm not going to lie. Just how it looks here is beautiful. Like, it looks absolutely beautiful. But what I don't what is annoying me is that I'm and I'm wondering if it's because I'm sshed in.

[00:39:33]If I am if it's because I'm sshed in, it's not letting me do some of these things. That's the main thing I'm wondering right now, right? Um because if that's the case, then I have to literally jump onto the Mac Mini and do this. Kaizen, appreciate you, brother.

[00:39:52]You are you're a goat. Um that's what I'm curious about. And maybe because Telegram bot token may not have uh have the right security settings, privacy mode. That's super odd though because like I've when I do this like when I done this any other time let's see.

[00:40:15]No, I don't think so. I I just find it super odd.

[00:40:24]Um I find it super odd. Okay, but it's essentially telling me that I didn't set it up correctly, which again, it doesn't make much sense to me because when I did set it up in the first place, like I did paste in the bot token and then like Bean Boy said, like it wasn't visible, so I just hit paste and did it anyways.

[00:40:42]Um, and it worked, you know, uh, worked normal. So I don't, uh, that one I don't understand, but okay. So, we can't get it inside a Telegram yet. What else can we have this thing do? Um, let me ask it this.

[00:41:02]Let me ask it this.

[00:41:06]Actually, let's make this big screen for you folks.

[00:41:11]Can I just paste the bot token in here with you and you handle the rest?

[00:41:25]>> [snorts] >> technically. Yes.

[00:41:32]What a nice guy. He He's He's really looking out for me. He's really looking out for me here. Uh Kaizen, give me two seconds. I'll read your stuff in a sec, man. Let me paste this bot token in really quick, folks. Let's see if this works.

[00:41:44]Yeah. See, there it goes.

[00:41:49]I just pasted it in there.

[00:42:02]No, no, don't worry. They can't see this, dude. Hermes is really looking out for me.

[00:42:16]Okay, while we're letting that do its thing, um, Kaizen asks, "Two partners, not so difficult. Uh, want to set up some automations for them. Should I make their own setup or want or one main and from there run agents that manage the clients?" Appreciate the content, bro.

[00:42:31]Okay, appreciate you, man. Um, yeah, I mean, you could do it all on one if they want if they're like wanting to have their own setup and [ __ ] you know? um if they were like wanting to have their own setup and stuff, then yeah, you would have to you would essentially need to go in and like have them buy a Mac Mini or something like that and you could set it up that way and then you can just use tail scale to log into their stuff or like screen share and get everything set up until you unless you're close to them. Um you could do that or like you could do like an easy like screen share on Zoom and set everything up and then you could have access to the agent on Telegram >> and then uh you could do that. That's for sure. Or if they don't really want their own setup, then yeah, you could just essentially have uh one main like I do. You just have one main Mac Mac Mini and then go from there. Um I wouldn't do a VPS if I'm being honest with you. Um I wouldn't I wouldn't do a VPS if you want me to be completely honest. [snorts] Um do you plan to use Discord uh with Hermes Agent? Probably not unless like I see an opportunity to do so. Like other than that, probably not. I don't know yet though. On that one, I don't know yet.

[00:43:46]Oh, Hermes, what do you Well, right now, gez, first impressions are [ __ ] annoying me right now. Um, but other than that, like I want to actually like utilize it. Utilize it. You know what I mean? Okay. Uh, Hermes Gateway restart and sleep status. Okay.

[00:44:02]Um yeah, let's go ahead and yeah, allow for this session.

[00:44:18]We might have it. We might have it connected here.

[00:44:24]We might have it connected here.

[00:44:30]He said, "Uh, why not BPS? Want to run it want to run in locally?" Uh, first want a good Mac for it. Yeah, but you don't need a good Mac, bro. You can literally just run that bad boy on anything. Um, I I wouldn't uh I I wouldn't use a VP.

[00:44:46]That's just me. I just wouldn't use a VPS. It just adds much more complexity to the mix that like I just don't think you really need to be quite honest with you. um in my opinion.

[00:44:59]Next, what's your top local LLM? I used to use Quen. I used to use uh what was it?

[00:45:04]Quen, was it 2.5?

[00:45:08]I think I was 2.5 27B, I believe it was as well. I tried it out. It wasn't horrible. Um, but again, my Mac Mini doesn't have the capability to run a lot of good local LLMs, so I just don't, you know, I just don't do that. Um, I don't really try to mess with it too too much. Dude, what the [ __ ] is going on here? What am I even looking at? I wish you guys could see what I'm looking at right now.

[00:45:37]It's trying to do API calls, and I don't understand what the [ __ ] deal is. All I want you to do is very simply just take the goddamn key, the her the the bot father key, the bot token. Take my [ __ ] bot token and [ __ ] plug that [ __ ] in. That's all I want you to do.

[00:46:00]And it failed again.

[00:46:03]This is really irritating me right now, guys. This is really irritating me. All right, I have my own Mac, so let me just download it locally then. I have a normal Mac Air M4. Fine to run it. Yeah.

[00:46:15]Yeah. I mean, I always prefer local over a VPS, that's for sure. Uh, dude, don't turn around. Laughing at my own demise right now. It's absurd. Um, yeah, you can run it on there, bro.

[00:46:28]Okay.

[00:46:30]Okay.

[00:46:32]Okay. Okay. We might have gotten somewhere after 15 hours. Good [ __ ] Hermes. I like it. I like it. All right, let me do this.

[00:46:43]Oh, it worked. Oh my god, it worked. All right, why not? Deep sea kidney, bro. Guys, listen. Listen. Let me tell you something real quick. Okay, let me tell you something real quick.

[00:46:58]I don't know my ass from my elbow with all of these local LLMs. There's too many of them. It's like Pokemon and I can't catch them all. Okay. There's just so [ __ ] many of them. Um I can't one I don't even get into them that much because I can't even run one of those [ __ ] worth that is even worth it. You know what I mean? Um that was that's the problem there. Uh and then with Deepseek and Kimmy, I know that they're powerful as [ __ ] I see them on Twater, so I know they're powerful. Deepseek V4 is like the equivalent to like Opus 46 or Opus 47 or something like that. And I know Kimmy, there was a Kimmy model that came out that was also absurdly well. Um, the Asians just aren't playing, okay? Like their their local LM are on point.

[00:47:39]That's no disc to Asians. Their local LM are that's actually a gas to them. Their local LLMs are on point. Um, but yeah, that's why I don't I just don't know too much. Um, I haven't gotten into them that much because I can't run them. U Okay, real quick. Let me let me do something. Let me go ahead here and let me give it the pairing code. Uh, don't worry, the stream can't see this, lol.

[00:48:04]But here is the pairing code.

[00:48:10]Imagine I just started speaking the code out loud. [laughter] Like I just started like talking out loud in the code and someone just hacks into my [ __ ] Okay. All right. Yeah. Yeah. You just got sometimes you just got to you just got to let it know. You just got to slap it around a little bit. Be like, "Listen, [ __ ] listen to me." Okay.

[00:48:29]Please, for the love of God. Um, what was the issue with Telegram? Honestly, I don't know. Um, it it said that I didn't set it up right in the and then the the botfather token.

[00:48:41]It said that it wasn't working or whatever. Uh, so I just literally gave it the botfather token inside of the terminal chat and then it worked and then now I had to pair it and then that worked and now we're here. So, let me do this.

[00:48:58]Now, let me ask it that. No, home channel is separate. Telegram home channel is where Hermes delivers crime job results. Okay.

[00:49:07]Uh yeah, let's make this the home channel.

[00:49:11]Let's make that the home channel. Okay.

[00:49:13]Beautiful. There we go. And he is in here. All right. Let me delete this so you guys can see the pairing code.

[00:49:22]Let's delete that.

[00:49:25]All right. I love that. I love that.

[00:49:27]Okay. Now let me uh share screen here and not do let me stop this. Share green screen share share access window telegram.

[00:49:43]All right.

[00:49:45]All right. There we go. Add to stage. Oh my gosh.

[00:49:50]All right.

[00:49:53]Okay. Now we're talking.

[00:49:57]Now we're talking.

[00:49:59]All right, we got it set up.

[00:50:02]Okay, so I am live streaming and until I tell you otherwise, please do not release any sensitive information.

[00:50:23]Now, are we good to chat in here solely and not use the terminal UI anymore?

[00:50:39]Uh well, what you can do, Kaizen, is you can actually set up uh uh you can set up uh you you can essentially like port everything over if you wanted to, but if not, yeah, you can literally just copy pasta all those uh all those uh all those MD files, but there's an easier way to do it.

[00:51:00]Thank you. Much appreciated.

[00:51:03]Now you are running on my Mac Mini, right?

[00:51:10]I'm not going to lie to you. I genuinely kind of like [ __ ] with the terminal.

[00:51:13]I like the terminal. It's just satisfying. I just like seeing colors and [ __ ] you know? Oh yes. Oh yeah.

[00:51:20]That's what I'm talking about. Right on that Mac Minsky. That's what I love to see. Right on that Mac Monica Lewinsky.

[00:51:26]We love that. We absolutely do. All right. Uh what should we do, guys? I mean, we've got some time to kill. How about it? Huh? We've got about This actually was a lot faster than I did.

[00:51:36]You know, I could do a series where it's like every Thursday I just keep on building upon it. You know what I mean?

[00:51:40]That'd be kind of cool. Uh [snorts] but we got some time to kill here. What we What should we build? Like what do we do now? Like we got it done. I'm kind of like, dude, I really like using it in the terminal. I'm not even going to front. I'm not even going to front. Um tell it to make me a million dollars.

[00:51:57]No, I can't do that. Can't do that. Um, it probably could to be honest if I gave it enough gave it enough capital. Uh, okay.

[00:52:08]Setting up obsidian with it is going to be ah, I know what to do.

[00:52:14]I know what to do.

[00:52:17]How can we make I'm going to say this. So, I know out the gate that your memory is not the greatest, but there are probably some settings that we can tweak in order to make this better.

[00:52:39]What are those settings? And I always want to start with the memory, guys. What do you recommend we change those two in order for you to be bulletproof when it comes to your memory?

[00:52:58][snorts] You guys want to know how I figured [ __ ] out on OpenClaw? It was literally just prompting it like this. [sighs] This is literally all I did.

[00:53:09]But the real thing, the real kicker, what the [ __ ] does that mean? Retrying in 2.4 seconds. I just asked you a question. Does anybody know what that means?

[00:53:24]Does anybody know what that means? I I I could do that. A, what the [ __ ] does that even mean? Retry again. I just typed in a [ __ ] question.

[00:53:33]I connected it with GBT 5.5.

[00:53:40]>> [laughter] >> GBT 5.5 [laughter] Vader. You absolute GU.

[00:53:51][laughter] You've heard correct, my friend.

[00:53:57]You've heard [ __ ] correct.

[00:54:01]Now, holy [ __ ] that's soothing to the throat.

[00:54:07]Now, [snorts] who who was in my chat on Tuesday talking about Honcho? Who the [ __ ] was in my chat Tuesday talking about Honcho?

[00:54:19]Hermes just brought up Honcho?

[00:54:22]Why is the cup so big? Is it really that big? You think so?

[00:54:31]It is kind of big, isn't it? I don't know. Well, I just hate small cups, you know?

[00:54:37]Then I got to go run, get another fill up, refill, rinse, repeat, recycle. You know what I mean? I don't want to do all that. [snorts] Okay.

[00:54:44]Okay. It took it a little bit to do that. Uh, it's tea. It's tea from Starbucks, but uh it doesn't look like tea, so I that could be misconceiving.

[00:54:56]That could be misconceiving. It's tea. I promise I'm not weird. Okay. Anyways.

[00:55:01]Okay. Okay. It says, "Yeah, good calling the bullpers sessions." Okay.

[00:55:08]I wonder if it does memory com I wonder if it has like a compaction like post compaction protocol like OpenClaw does.

[00:55:18]Okay.

[00:55:20]So, both of these are enabled. Memory enabled. You user profile enabled.

[00:55:25]Lets me retain durable environment projects. Lets me retain your preferences. Communication style constraints. Add a stronger memory provider. Built-in memory works, but more reliable long-term retrieval. I'd recommend adding a dedicated provider.

[00:55:39]Okay. Honcho, MIMO, Memo, Super Memory, and local. Then pick a provider. My recommendations, if you want best general reliability, Honcho, privacy, local first, local. If you want hosted searchable memory, I wonder if I can use QMD. Make sure session search is enabled.

[00:56:00]Okay. Uh, this is huge. Memory only saves distilled facts. Session search lets me retrieve old conversations when you say things like, "Remember when we fixed that? Enable the tool set if needed." Okay.

[00:56:13]Enable it for me. What the [ __ ] Like, keep skills enabled. Skills are better than memory for procedures.

[00:56:23]Dude, you're talking to me like I'm like I'm stupid right now. Like, I I'm not dumb. I get what skills are made for.

[00:56:30]Okay, I know that I should save stable workflows as skills instead of stuffing them into memory. Uh, tune compression so longs don't lose context badly.

[00:56:39]Compression prevents long conversations from overflowing.

[00:56:43]Okay. Okay. So, here's what I'm going to do because it just spat out a whole bunch of vomit at me. Okay.

[00:56:52]Uh, we're going to do this. We're going to do this.

[00:56:58]Okay, so let's attack this one step at a time, you psychopath.

[00:57:08]First, let's optimize any we can within the config in order for us to maximize memory effectiveness.

[00:57:27]are all of the settings inside of the config done uh or correct and optimized or do we need to change them?

[00:57:47]Let's start there. Let's just start with like do it do the basic [ __ ] for me first.

[00:57:55]Okay.

[00:57:56]Jesus. Hermes.

[00:58:05]Wait. Okay. Here we go. Here we go with the question. Kaisen, as you brought up the context compaction in one of your videos, I saw that you mentioned that with 1 million tokens and then all the context compacts explaining, etc. But would you would you say but what would you say to do when you have an LLM that only has 200,000 token context? Okay.

[00:58:28]200,000 token limit. I do understand what your question is. Yes, I do understand what your question is. So then what I would do is I would simply you keep the settings the same because you're setting a soft token thresh a soft threshold token. Okay. And what that is essentially doing is it's telling the it's it's telling the the config essentially like it's it's default settings. What it's telling it is is okay once we reach this level of of like uh usage in a session before we compact before like our context compacts it then triggers it to to automatically do a context compassion or a memory flush.

[00:59:09]Okay. Not a context compassion. I'm sorry a memory flush. All right.

[00:59:12][snorts] So, you keep those settings the same, right? It doesn't change. It just it depends on it doesn't depend on like let's say the the the uh context window like how big the window is. It your settings could stay the same regardless. Does that make sense? Um let me know if that if you have any questions on that, Kaizen, because it's a good question and it shouldn't change, right?

[00:59:36]God, it answers so long. Holy [ __ ] Current status already good. Okay.

[00:59:44]Okay. Already good. Uh so the core pieces are present. What is not optimized yet? Built-in memory only. Uh memory provider. That means Hermes is using built-in memory only. That is fine but not bulletproof for stronger memory.

[00:59:56]We should eventually configure a dedicated provider. Okay.

[01:00:04]Thank you.

[01:00:06]Memory character limits are probably too low for strong long-term memory. That is pretty small. I would change it to this.

[01:00:13]Why? The memory character limit controls how much durable memory can be injected into future sessions. User character limit controls how much user profile preference memory can be injected.

[01:00:25]Increasing these gives me more persistent context to work with. The trade-off slightly more prompt context usage built or but on a large context model. This is worth it. Okay. Okay, that makes sense. Now, here's the thing.

[01:00:43]When it when it tells me [ __ ] like this, when [snorts] it tells me [ __ ] like this, I'm not just going to go with like what does it say? You know what I mean?

[01:00:50]I'm not just going to go with like what it tells me. I'm going to be like, is that the most optimal? Right? Is that that memory character limit at 8,000 and the memory for the usage character limit at 4K? Like, is that optimal or do I need to go higher? Like, I don't give a [ __ ] about like contact. Like I don't give a [ __ ] about you know if like you know on a large model like essentially like gez it's been a long day guys I'm sorry like I essentially want it to be as optimal as possible so that way I don't have to keep feeding it new context like it should pull from the previous. Does that make sense? Sorry it's been a long [ __ ] day.

[01:01:25]All right here is another one. Memory save cadence could be more aggressive. The current is 10 and six on the nudge interval and the flush min turns for normal use. This is okay for me. For more remember more reliably, I'd take it more aggressive.

[01:01:41]Five and three. Why? I'll be prompted to persist useful facts more often. Short chats are less likely to end before memory extraction happens. The trade-off, more frequent memory consideration, slight chance of saving more noise, but I can manage that by only saving durable facts. Compression is too aggressive.

[01:02:00]Okay. Damn. Okay, this is giving me some useful information here, folks. Okay.

[01:02:05]Um, these are the settings I have. This means compression can kick in when the context is only about half full, then crush the retained content down to 20%.

[01:02:14]That's efficient, but not ideal for memory and context fidelity. I'd recommend this. Why? Wait longer before compressing. Preserve a richer summary.

[01:02:23]Keep more recent messages untouched.

[01:02:25]This should make long conversations feel much less amnesiac. Okay, like your terminology there. Appreciate you.

[01:02:33]Next, uh, no explicit model context length set. Current relevant model config does not show. Uh, I'd set it explicitly if your current provider model supports it. I think our context is 272,000.

[01:02:51]That's what that is. So, we could set that at 272K, but only if the selected model actually supports that. If unsupported, it could cause issues or simply be ignored. This is this is this one is beneficial, but I this one is beneficial, but I treat it as slightly more cautious than the memory compression settings. All right.

[01:03:09]I mean I mean this it's giving us a fuckload of information. Uh holy [ __ ] Yeah, you see this is how I learn [ __ ] Uh real quick, Kaisen, I I seen your chat here. Um also really quick uh AV I will do that in a moment. Uh you said got you bro. So same settings don't change the context window is not the main issue. The settings. Yes, it's it's the settings. It's all about the settings. Um, and yeah, I appreciate it.

[01:03:31]Yeah. So, guys, I want you to make something I want to make something very clear to you here, okay? Like this this is how I learn. This is like my way of like learning, right? If I if not I have to like dig through I mean, no one's putting out YouTube content on like walking you through this [ __ ] No one's explaining all these little things, right? No one [ __ ] does that around here. Uh, as as AV said, I believe it was. No, no, I'm not. it was Vader, Jacob, uh, all these other YouTubers, they're not putting this stuff out.

[01:04:01]Okay, this is why I am. Um, but also to another degree, like this this is the type of stuff that like this is the type of stuff that I do and prompt it with so that way I can learn.

[01:04:12]So, I hope this is really helpful for you guys. Um, okay, let's go back here.

[01:04:17]Yeah, let's do a little bit of this, a little bit of this action. Okay.

[01:04:23]Okay. So now what I'm going to do is I'm going to go back here. My dog's not storing anymore. We love that. Okay.

[01:04:32]So number one, uh uh we are not going to install a memory add-on yet for two.

[01:04:49]Okay. It wants me to change this to this. you are wanting me. Okay. No, I'm going to say I like your suggestion, but is there any reason you didn't go more aggressive to learn to retain and store memory better [snorts] also it it gave us some context here. Um just for context we are working on model GBT 5.5 and the context window is 272K.

[01:05:38]So I would adjust the settings accordingly.

[01:05:45]Let me know your thoughts.

[01:05:48]For number three, a memory save cadence.

[01:05:52]Okay.

[01:05:54]Um, we'll leave that as is. Sounds good. I'm going to say sounds good.

[01:06:03]You can change this accordingly as well. Let me know guys if the screen is like too small and you guys can't read anything cuz I'm looking at it on my screen and I can't read [ __ ] But like I don't know if that's for you guys as well. So, please put in the comments or in the chat. Let me know if it's too small. I'll try to find a way to make it bigger. Um, I wonder if I do this. If I do this. Oh, yeah. Okay, now we're Now we're talking here.

[01:06:30]>> Now we're talking. Okay, hold on. I'll get it back for you.

[01:06:35]There we go. That's probably way [ __ ] better for you guys. I am so sorry. Um, I am so sorry. Okay. Uh, we're at number three. Yeah, that sounds good.

[01:06:47]You can change this accordingly as well.

[01:06:52]Okay. Number four.

[01:06:55]Compression is too aggressive.

[01:06:58]This means compression can kick in when the context is only about half full.

[01:07:07]Uh, okay. I'd recommend this. Wait longer before compressing. Preserve a richer summary. Keep more recent messages untouched. Okay, hold on.

[01:07:17]But if we wait longer to compress, doesn't that mean that you can lose nuance within what is getting compacted?

[01:07:39]Also, the more we the more often we compress, the less we have to worry about losing context slashnuance of the information that was compressed.

[01:08:06]Correct.

[01:08:07][snorts] Or am I wrong?

[01:08:11]Number five.

[01:08:14]Um, for the model, are you talking about the LLM that I am using?

[01:08:25]If that is the case, we are using GBT 5.5 and it has a window of 272K.

[01:08:42]Okay, auxiliary models are on auto. The [ __ ] does that mean? This is not wrong, but for maximum reliability, we may eventually want to pin these to to a known good model and provider so compression and session search doesn't silently use a weak fallback. I would not change this first unless we know exactly which auxiliary model and provider you want to use. Okay.

[01:09:10]Okay, for number six, what do you mean by auxiliary models?

[01:09:22]I need more explanation on this. Okay, this is what I did when I first got open claw. I did this [ __ ] for [ __ ] 16 hours a day every day to figure this [ __ ] out.

[01:09:40]And I would just talk to it all day.

[01:09:49]Oh, I'm glad you're finding some value in this, guys. And I hope everybody else is as well. If you guys could do me a solid favor, just smash a like button on the on the video, man, I would greatly appreciate it. It's really all I ask of you. If you're not in the Telegram chat, go ahead and uh join the Telegram. It's completely free. Nothing to sell you in there at all. Links down in the description below. Follow me on Twitter if you haven't already. You guys been Some people have also been DMing me on Instagram. So I obviously the Instagram is on my channel as well. Um but yeah.

[01:10:16]Yeah, if you guys could do that, that'd be great. Appreciate you, Bean Boy.

[01:10:19]Appreciate you. All right, we're going to see what this says back here. Okay, here we go.

[01:10:26]Here we go.

[01:10:30]Done. I applied the more aggressive config memory optimization with no memory add-on. Okay. Changes applied.

[01:10:35]Boom. Boom.

[01:10:38]Okay.

[01:10:40]Number two. Actually, for number one, are you able to use QMD?

[01:10:49]Okay. Why I originally don't go more aggressive. I was being conservative because I didn't know yet the context window. With a normal 32K to 128K setup, injecting too much memory can become counterproductive. But now we know this.

[01:11:00]It makes sense to go much bigger. I figured it did. That's like, dude, we have a [ __ ] fat window. You know, I'm very surprised that GBT 5.5 doesn't have a memory like a context window of a million tokens. I'm actually thoroughly surprised. You know, they should definitely do that. But whatever.

[01:11:17]Anyways, who am I? I'm not Sam Alman.

[01:11:20]Uh, okay. So, instead of this, we're going to go to this. This gives me much more persistent memory room and still avoiding the mistake of dumping everything into every prompt. I would not recommend going unlimited or absurdly high. No, more memory is not always better if it becomes noisy. The sweet spot enough durable context to remember your world. Not so much old relevant facts crowd out the current task or context. Okay, perfect. So, it changed this uh character limit to 24K, user character limit to 12. Boom.

[01:11:49]Boom. Okay, it did. It changed that. All right. Perfect. We love that. Appreciate you.

[01:11:54]Makes sense. And thank you.

[01:11:58]All right. Number three. More aggressive memory. Save cadence. Changed from 106 to 32. Nudge interval 3. Flush min interval 2. Nudge interval. Boom. Boom.

[01:12:07]Okay. Did change that. This means I should consider memory persistence more often and flush and save useful facts.

[01:12:13]Durable facts sooner. Um, this should help with the con with the classic issue of we talked about this. Why didn't you remember it later? Yeah, I don't want to have to [ __ ] remind you of that.

[01:12:24]Number four, compression. Your intuition is partly right, but there's a trap. Ah, there always is. There always is. Um, you asked if Dude, I love how it takes bits and pieces. All right, Kaizen.

[01:12:36]Appreciate you, my man. Um, thank you for joining the live stream, bro. And, uh, and I hope you have a great time.

[01:12:43]Um, I love how it pulls I love how it pulls like bits and pieces of my [ __ ] uh from like my previous message. Okay. If we wait longer to compress, does that mean you'll lose nuance within what's getting compacted? Potentially, yes. If you wait until a conversation is massive, the compression job has much more material to summarize, so it can miss details. But the opposite has a bigger hidden problem. Compressing more often can cause repeatedly loss uh lossy summarization. I don't know what the [ __ ] that means. Think of it like saving a JPEG over and over. Each compression pass can lose a little nuance. If Hermes compresses early and often, then later it may compress a large summary of a summary of a summary. Okay, that makes sense. That can that can be worse than waiting longer. So, with a 272,000 context window, the better strategy is keep raw conversation in tag longer, compress later when compression happens, preserve a richer memory or summary, and protect more recent messages from being compressed.

[01:13:41]Okay.

[01:13:43]Okay, that makes sense.

[01:13:50]Okay, we're gonna set this up, guys, and then we're going to call it and then I will pick up back where I left off on the next live stream. Okay, this makes sense. Meaning, don't compress until around 85% of the context is used. When compressing, keep a much richer summary.

[01:14:07]Preserve the last 60 messages untouched.

[01:14:09]Avoid message count hygiene kicking in too early. Okay, this is more appropriate for huge context model than the default.

[01:14:17]The old settings were efficient but too lossy lossy for your goal. Okay, perfect. So, I'm just going to say so are we now set up to have the settings be most optimal?

[01:14:35]All right, number five. Yes. Model context length means your LLM you're using. Okay. Since you're using this. Okay. Perfect. This tells Hermes the budget on the Okay.

[01:14:46]Beautiful. Okay. So, what did it set this to? Yes. Okay. Good. Good. Good.

[01:14:51]Good. Good. Good. Good. Good. Good.

[01:14:53]Good. We loved it. We loved it. Yes, we do. Perfect. Thank you.

[01:15:00]Auxiliary models are secondary models call calls Hermes uses for helper tasks separate from the main chat model.

[01:15:09]compression session vision and title generation. Okay, they can be used for things like summarizing and compressing long context, searching, summarizing past sessions, generating session titles, analyzing images, approving low-risk commands if smart approvals are enabled, skill hub and curator tasks.

[01:15:25]Right now, yours are set to whatever that means. Hermes picks an available model automatically.

[01:15:31]That is convenient, but it can be less predictable. For memory quality, the important auxiliary ones are compression and session search because if those use a weak, cheap, incorrect fallback model, then compression summaries may be lower quality and session search summaries may miss important details. I did not change the auxiliary models yet because that requires deciding what model or provider should handle those helper tasks. Given that you're on GBT55, we may eventually want to pin and compress pin compression and session search to a strong model too, but that's a separate step.

[01:16:02]Important note, these config changes may require a fresh Hermes session or gateway restart to take full effect.

[01:16:09]Since we're staying in Telegram, the relevant command would be that. Okay.

[01:16:13]Okay. So, for the auxiliary models, can you please have them both set to GBT 5.5?

[01:16:26]Is there any benefit to having it having these auxiliary models for search and compression to be a high-end frontier model like this?

[01:16:45]I would assume the better the model, the higher quality the output.

[01:16:55]Am I correct?

[01:17:01]Let's see.

[01:17:07]We shall see here. We shall see here.

[01:17:12]So, what I'm going to do, just so you guys are aware, is we're going to we're going to finalize these settings.

[01:17:18]And then, uh, I'm most likely going to save this as like a something we do on live streams. Like, this would be kind of cool. So, you guys can see this like on a live stream all the time. And you can just see me continuously build this out. Um, okay, done. I set their both relevant auxiliary models to GBT55 QMD if you mean you mean markdown files QMD. No.

[01:17:44]Yes, I can work with them. No.

[01:17:48]Okay, I'll do that. Are we now set up optimally for config memory optimization without installing a memory add-on? Yes, we are now in a very strong setup. All right, so if you guys want to copy this setup and you could take it, then uh by all means go ahead. I'm operating on GBT55, so uh and I owed in. So you guys could literally take all those settings and go from there. Um, so yes, for your current constraints, it's a pretty close to optimal. The only remaining upgrades would be later adding a dedicated memory provider, possibly creating more durable skills and workflows as we discover repeatable patterns and restarting the gateway session so all fig can changes uh config changes fully apply. Is there a benefit to using GBFD 5.5 for auxiliary search and compression? Yes, especially for your goal. You are correct. Better model generally means better auxiliary output, especially for nuance tasks like compression and session search. Compression benefits. A frontier model should be better at pre preserving nuance, distinguishing important versus unimportant details, retaining intent decisions, constraints and unresolved threads, not flattening the conversation into vague summaries, maintaining technical accuracy across all chats. Yeah, it's exactly what the [ __ ] we want. This matters a lot because compression is lossy by nature. A stronger model gives us a better compression codec.

[01:18:58]Okay, session search benefits for session search. GBT55 should be better at these things trade-off. The downside, higher cost, possibly higher latency. Um, if the main provider has rate limits, auxiliary tasks can also assume that. Okay, that makes sense, but I'm only on the $100 plan. So, honestly, we're going to leave it there for now. Um, but if your priority is memory quality over cheap fast operation, then yes. Yes. See, that's always my case, guys. That's always what I want to do. I would always much rather have like a higher memory.

[01:19:29]Like I care more about the memory than I do the cost. I care more about memory than in this case than I do the usage.

[01:19:35]Um I I I would rather that because yes, it will use a little bit more of my usage rates, but at the end of the day, I'm literally going to not have to repeat myself. It's going to have better memory, etc., etc., etc., right? Uh so my verdict for your use case, big context window, live streaming, want to maximize memory reliability. Yes, this is correct. Okay.

[01:19:53]Okay. Perfect.

[01:19:55]No. Uh I'm going to say so when I am talking about QMD, I am talking about the plug-in that was created by the CEO of Shopify.

[01:20:09]And do you want me to give you a link in order for us to uh in order for you to know what it is?

[01:20:23]If I give you the GitHub, I do not want you to install it right away.

[01:20:40]Makes sense.

[01:20:43]Also, for the auxiliary models, if you haven't already, please go ahead and install the or please go ahead and set those to GBT 5.5 and especially for search and compression.

[01:21:10]So we get the highest quality output and summaries.

[01:21:22]Everything else that we discussed is already done I believe. So the auxiliary models was the last bit of the puzzle to complete.

[01:21:41]Am I correct?

[01:21:46]So far, now that I'm like, you know, kind of messing with this a little bit more, I haven't actually got to use it fully, but like I'm not I'm not I'm not upset with it.

[01:21:58]You know what I mean? It is what it is.

[01:21:59]Not upset with how this is.

[01:22:05]Let's see.

[01:22:08]I don't know. Let me know. Let me know if you guys like watching stuff like this. I I I mean I could set this all up on my own, but I figured I would uh I would do this in front of you guys.

[01:22:19]Okay, what's it say?

[01:22:22]Yes, please send the GitHub link if you want me to understand exactly what you mean and understand clearly. I will not install just because you send the link.

[01:22:27]I'll only inspect, read, and blah blah blah blah blah. Okay, cool. Auxiliary models already done and verified.

[01:22:33]Good. Current config. Uh so the search and session recall and compression are now both for GBT55. Are we done with them? Yes, for the plan we discussed the config is side is complete. The only caveat config changes six week prior refresh session or gateway restart.

[01:22:48]Okay.

[01:22:52]Restart this.

[01:22:57]Let's see now. I would love for it to tell me when it when it's done restarting because if it tells me when it's done restarting then that would be nice.

[01:23:14]That would be nice.

[01:23:19]So, let's go ahead and just let this run and then we'll be good to go. Gateway restarted successfully. Your session continues. Okay, perfect. I restarted the gateway and everything is now done.

[01:23:33]Correct.

[01:23:38]I really like this. I really like that it tells me when the gateway is done.

[01:23:43]I really like that a lot. Correct. If you restarted the gateway after these config changes, uh then the memory optation pass is now complete. You should now have built-in memory enabled, user profile memory enabled, larger injected memory limits, more frequent memory flushing and nudge behavior, 272,000 uh model contexts configured, less lossy, later compression, more recent messages protected from compaction or compression. GBT 5.5 pin for the main model compression and session search.

[01:24:07]Privacy and secret redaction enabled for live stream safety. No external memory add-on installed yet. So yes, everything we discussed this phase is done.

[01:24:17]Alrighty, guys. All right. Now, I want you guys to do me a favor. Smash the like button if you haven't already. Go ahead and join the Telegram. Follow me on Twitter. DM me if you have any questions. Uh I'm getting a lot of like requests in there, so I will respond to all of those um probably by tomorrow.

[01:24:34]Uh, so that's all done. We're going to end this live stream here and then what we'll do is we will go ahead and continue and pick up where we left off.

[01:24:43]Um, as long as everybody wants me to. If not, I could just zoom through the rest of this and get this taken care of on my own. Um, so other than that, I will uh we'll end it here. I I hope you guys enjoyed. I hope you guys got a lot of value from this. This was super super fun for me and uh you guys got to see me almost pull my hair out a little bit at the beginning because it was kind of annoying. But we did all of this through sshing into my Mac Mini to do all this.

[01:25:08]So this was actually like very interesting for me. Um, I'm going to continue to build out the foundation and the config and all of the things to ensure that we are essentially like setting up this Hermes agent for complete maximum effectiveness and efficiency and memory optimization and all that stuff because that's going to be the most important and that's kind of like you guys are starting to see now like what I'm doing here is essentially like what I would do what I did with OpenClaw like this is exactly what I did and my plan is to obviously run Hermes along alongside Open Claws. So, that's really the game plan. Um, but I hope you guys enjoyed this. I really do. And, uh, yeah, I'll uh I'll run a poll inside of the Telegram so that way we can see if people like that and if they actually want me to go back and continue to build build on this.

[01:25:58]So, I'll I'll do that. This is going to be this Hermes agent is going to start off as like my personal assistant uh and get a lot of things done for me. So, that's what this is going to start off as. And we'll just kind of see where it goes. you know, after that I'll I you know, if I want to include it in some of the business side of things, then I will. But, you know, for the most part, I'm going to keep it like this. Um, I appreciate that. Uh, Brayton tons Brayton, appreciate that, man. Thank you very much. Um, video is going to be releasing tomorrow, 8 a.m. Central Standard Time. You guys should definitely tune in. Uh, it's going to be, I believe, video four, five. My god, I can't even remember right now. It's been a long day. video four, video five of the uh openclaw blueprint. So, um we've I've been getting a lot of good feedback on that. And so, I'm going to continue to obviously post uh the rest of that u series. So, that's going to get done as well. But, smash the like button if you guys enjoyed. Subscribe, hit the bell notification, that way you guys are always notified. We go live every week on this channel on Tuesdays and Thursdays at 700 p.m. Central Standard Time. And uh this was a nice little test, something new to do on live stream other than just answer questions, which I could still do in the midst of all this. So anyways, hope you guys enjoyed. Appreciate you guys a ton. And uh yeah, see you guys in the next live stream.