Making OpenClaw Actually Remember Things

Aloha and welcome. Today we're going to fix long-term memory. So I actually have OpenClaw installed on this Mac Mini right here. And so this thing is actually

here. And so this thing is actually running my whole entire life. And I

noticed when I was chatting yesterday that my memory stuff wasn't really working.

After things compacted, it says like, hey, I didn't really know what was going on.

And so what I found out was that there's this thing called QMD. And this is a new thing that came as part of the update that we're going to go ahead and talk about.

We're also going to talk about here about OpenClaw. Basically, OpenClaw has now

OpenClaw. Basically, OpenClaw has now become in the hands of OpenAI, which is pretty amazing. So Pete Steinberger is

pretty amazing. So Pete Steinberger is now joining the OpenAI team, and he's going to be basically helping them. Like, I think they've set up a whole new foundation and everything like that, and a lot of people are really concerned that this may actually tank the project or something like that, and it's actually not.

I'm also the belief of Alex Finn where he's also saying that it's very bullish and we're going to talk a little bit more about that.

So we're also going to talk a little bit about the memory optimizer skills.

So I made a skill around this whole thing and it's a little script that's going to check to make sure that your memory is pretty much compact.

Because what most people don't realize is that certain files inside of your memory can get really bloated.

So that every time you start to actually prompt your OpenClaw agent, it'll start to like maybe use 2,000 tokens right now.

A week from now, it could be 6,000 tokens, it could be 20,000 tokens, because sometimes you'll say, hey, can you remember this? Can you remember that? Can

remember this? Can you remember that? Can

you do this? And by default, want to put it into like an agent's MD file or in these other places that don't always need to be there. So there's a way to actually store these memories in a progressive manner so that you don't always have to fill your contacts window. So this will actually make your agent a lot smarter, a lot more personal to you, and actually kind of remember things in conversations,

especially for these long-running tasks.

So we're going to get into that a little bit today.

This is an amazing live stream so you're going to make sure you're tuning in because we have a whole bunch of people tuning in because we have Alex Finn in the building as well so if you haven't got a chance to tune in, Alex Finn, he's also a mod so if you talk any smack, my boy's got you right there. He also went crazy and he's got a couple Mac studios today he's

going to be coming in. Huge shout out to Eaglesworth as well and I appreciate the Pixel Collector for popping in. Also make

sure you do a favor right now and just smash a thumbs up and share this with some other folks because right now it's not being posted to X right now for whatever reason so I'm going to see if I can repost that right now.

so that we can get that going. So let's

go ahead and kind of do that as well. So

we've got to take care of some of the basic bills here because right now I think there's a problem with X posting somehow.

And so I'm just going to make sure that this gets reposted because as of now the stream goes live and you can kind of see it's not really on here yet.

For whatever reason, I have no clue. So I

have to go into here and then I have to like share the screen.

so I'll do post this live now to fix open claw memory for my agent. Okay, let's

post this. Okay, so now that I've posted that,

this. Okay, so now that I've posted that, it should now start showing up on my stream because there's like zero people actually joining in. Okay, cool. So yeah,

now that's other people should be posting up there. And so okay, so now that we got

up there. And so okay, so now that we got that, we got Alex Finn here. And so he's talking about this use super memory AI.

So you know, what's really interesting is that like, I tried to look into this and I couldn't really find out a lot about the Super Memory AI because I love Dravya.

Actually, he's a really cool founder and stuff. And what I also found interesting

stuff. And what I also found interesting is that there's stuff that's already built in that can work really, really well.

The thing about this type of app, which I may explore later on, depending on what happens, is the fact that like, you know, after about three million tokens, you're going to pay 20 bucks a month.

and then you only get like a thousand searches and so this stuff can like really kind of lead up to a lot of stuff over time.

The cool thing about like OpenClaw is that it's open so a lot of people are building stuff in real time so you can actually like make something and open source it and put it info free.

So they already provide like a very basic, let's just kind of like talk a little bit deeper right now on the memory system.

A lot of people don't really know about this stuff and I think you should probably learn about this because no one's really talking about this technically, right?

Other than Alex Finn and even Ross Mike and some other content creators.

So the memory system, it's all documented here on the OpenClaw docs.

It's basically stuff is in plain markdown.

So I'm going to show you really quick.

I actually have, because I have it running on my Mac mini, I can actually do a remote session inside and I can show you the workspace.

So for my actual workspace, you'll see files like this.

So this will explain kind of what happens.

This is like first boot.

This is exactly like kind of what happened, how we set up the Mac Mini, and this goes through the various days.

And then there's like an archives folder that I created to port over my ChatGPT messages over onto this Mac.

And so you'll also see like any indices that it creates as well that also got put over.

So that's basically kind of, you know, the long and story short.

It's like in your memory, the stuff will actually be in here and the stuff will be archived.

There's also a skills here.

There's also some like overlay.

This is like actual workspace that my agent was using.

So this is stuff about the user, this is stuff about my soul, and then these are like anything else that we have in here.

So here's a memory thing. So what ends up happening with the memory stuff is basically over time, this will start to get filled up.

And every time you have a conversation, this gets loaded right in. There's also

like an agents.md.

So the agents.md file is the actual instructions for how, if you've ever used Clawed Code, if you've ever used any one of these coding agents, the agents.md basically steers the

the agents.md basically steers the language model, and that gets loaded every time in every conversation.

and so a lot of times that file can get filled up pretty fast and start to kind of pollute things and so if we kind of go back into here basically you can see every single day you'll have a daily log that spits out it's going to read today and yesterday's session every time it starts up and so what ends up happening as as the days go further you don't have access to further days unless you

actually specifically ask for something like that right and then this is a memory MD file this is optional this is curated stuff that gets filled up so only load in the main private session, never in group context. And so these basically show you where they live. And this is the reason kind of

they live. And this is the reason kind of why we're going over this is just to really understand how this type of system works here. And so this is like there's an automatic memory flush as well. So when

well. So when a session is close to auto-completion, it triggers a silent agentic turn. So this

reminds the model to basically write any memory before the context is compacted. And so usually it'll do a flush and you You can see the reservoir for the tokens floor is 20,000 tokens.

And then the threshold for the token is 4,000.

So this is a prompt that it writes basically to flush the memory.

You can see right here, write any lasting notes into here before it does anything.

And so there's some other parameters you can also tweak here as well.

So workspace must be writable, obviously.

And then there's a vector memory search.

So this is the part that we're going to really talk about.

And this is kind of what somebody was mentioning about super memory versus these other things.

So by default, there's a vector memory search that can actually go in.

and this is cool but QMD is actual project. Let me see if I can find it in

project. Let me see if I can find it in the QMD project. QMD GitHub.

So QMD is this project here. If you're

actually wondering, you can actually install this yourself.

It's called a mini CLI search engine for your documentation. So these are

your documentation. So these are knowledge bases, media notes, whatever. You can track your current state and this is really, really good at searching markdown. And so the way that it works is

markdown. And so the way that it works is that user gets a query and all of this is free by the way.

This is free and open source and this is why it's really, really cool. So you're

not going to be trapped into paying, you know, 20 bucks a month or even 400 bucks a month for enterprise grade, you know, like this type of style stuff, which is like, it's cool, but if like all I'm doing is just, is this guy on a Mac mini, bro? Like, you

know what I mean? And I think this is kind of why I really want to answer this and kind of fill in some of this technical backgrounds because there's a lot that's going on here and maybe people will hear the term QMD, but they don't really realize like the importance and actually what's happening behind the scene when you're calling this type of stuff to get memory because this is a really complicated topic and it's still

being discovered right now in the agentic world.

So obviously you want things to be really fast, things to be kind of ranked up so that you can always get the information you need and relevant at the time.

And this is really important for your agents as well. So you can see there's different phases, right?

When you ask something, you're basically your agent or the actual framework is going to do a little bit of an expansion to try to understand what you're looking for.

So they have these fancy terms here that kind of spread out.

And then these types of things spread out to search stuff through these different files.

So you'll see these key terms like BM25, Vector Search.

And you just have to kind of know that some fancy stuff is going on in the back end.

And then what ends up happening is things will actually get ranked accordingly depending on what you're asking for.

So I can say, hey, what was that vacation plan that I had to Italy?

Can you just bring it back up?

So it's going to be able to do this technique and bring up those information and say, what was my packing list at the time? And boom, you'll be able to pull

time? And boom, you'll be able to pull that up pretty quick. And so this is pretty cool. This is free. So previously,

pretty cool. This is free. So previously,

you would have to go through these commands to point your agent to this specific directory and install it, and it would download a little model. But now

that's actually kind of baked into the actual backend on the latest version. So

if you have 2026, February 13th, so 2.13, you'll actually have this experimental backend thing.

and if you don't, if you ask your agents to set up QMD, it's going to be like, hey, I don't know what that is and it might kind of go rogue.

So if you want to go ahead and do that, what I did for my agent is basically I just told it to go ahead and do an update.

So this is my actual chat conversation and this is what you want to do is just say, hey, update yourself, OpenClaw 2026 2.15 is here.

I just copied and pasted literally the tweet from the OpenClaw itself.

So OpenClaw, I just said, hey, update yourself, this is the latest version that's out.

and it's really funny how it actually remembered that and it was actually like cool and then basically it does it kicks off the update and I even had an older update before so if you're on an older update before the rename don't worry once the update is complete you can do some other doctor commands tell it to fix itself so that it uses the new open claw names it can it's amazing how smart this

thing is and at fixing itself but that's kind of the first step to get this thing updated so like make sure you're up to date to the latest thing and then make sure you get this thing going here so Telegram has a new feature called streaming, which is really cool. So

you'll start to see some of this text streaming in. So I said, yo, do we have anything to

in. So I said, yo, do we have anything to enable streaming? So you make sure you go

enable streaming? So you make sure you go ahead and do that as well. And I said, hey, this isn't streaming. And I

basically had to quit the app and reopen it to get the streaming working. And so that was really cool. And then you can do slash

really cool. And then you can do slash reasoning, obviously stream to see like the reasoning stream. So if you do slash reasoning stream, you can actually turn the reasoning on or turn it off or like reasoning.

like basically what ends up happening is one of them will actually you'll see a chat bubble filling in with the reasoning tokens and after the reasoning is done it will flip it out and show the response so that way you're not like flooded with a bunch of reasoning you know messages and trying to scroll through which can make it really annoying but if you actually want to keep everything you just say

just show me everything on you know and so you'll get separate messages for the streaming which could be super handy but right now I just leave mine for reasoning stream so that way I'll see the stream of thought. And then that way I can catch

thought. And then that way I can catch things if something breaks. So it's

pretty cool. So like I

cool. So like I said, you have the full combo. And now

that I make sure that the stuff was enabled, it's actually pulling up my Italy trip, right? So that was kind of cool there. So kind of said, how do you use this tool or whatever you're using? And that's kind of where I was pulling up QMD. And so this is the problem that I was having is that sometimes people are saying that they're having some memory issues. And this

issues. And this is where we're going to actually go in and take a look at the memory fixing stuff. So I obviously had

stuff. So I obviously had that says like, you know, take a look at my memory files and figure out what we're doing.

And so in this conversation, I'm kind of showing you here is that what's the strategy that we should do?

Just cutting it seems extreme because I didn't want to just remove memories from the file.

And how do we do this? And so I have EXA code and I have an EXA skill.

So for those who are wondering, my skills files are here and I have an EXA skill.

I'm going to open source this. I'm going

to release an open source repo.

And right now, I think I have this currently in my Discord.

Discord. So if you want to click my Discord, I actually have like an open Claw like chat group and stuff like that in there. But basically I wrote these scripts myself and I had basically had the agents do it. But this uses the latest version of ExaSearch. So Exa.ai is amazing at search.

And there's this new thing called like Instant Search. So if you're not

Instant Search. So if you're not familiar, you basically pass a flag in called Instant. And like

the search is faster than Brave. It

scores higher than Brave.

And it's like you can find anything on the Internet. I use it a lot for coding

the Internet. I use it a lot for coding and it's incredibly phenomenal for finding GitHub repos issues in GitHub Discord like conversations of people having issues and then proposed workaround fixes It amazing Yeah I can tell you how amazing that is I wish I was paid for them So Exa if you want to part it with me, definitely hit me up. I'm a huge advocate of their stuff. But yeah,

this skill is something I'll open source.

I also made a memory optimizer skill, which basically we're going to talk about coming up here. So

this is how I basically was able to derive my memory optimizer skills because we did some research on how, like, and normally when I'm AI coding is basically I usually have multiple steps in which my agentic software runs. So in a conversation, I usually organize things by concern. So I'll say, you know, I want to accomplish this task, right? It's like,

right? It's like, you know, make me a ping pong game or something. So I, you know, step by step

something. So I, you know, step by step to make no mistakes or something, you know, people really joke around about this, but you have to, as an engineer think through this in different pieces. And so

what's nice about these agents is they can already start to do that themselves and self-organize. But

you still have to be cautious about every single conversation that you have, because if it saves something and it keeps referencing that up, it's like you have a table that you can never clear out because it's always kind of busy, right?

So you can't really think about things. And the same thing kind of happens here. This is usually done automatically, but as of right now, this may be fixed in the future and I may do a pull request for this but I'm still experimenting. This is kind of why I wanted to kind of, I'm not really publicly posting this stuff yet because as a person who's watching this live stream, all you have to

do is hit the thumbs up and then share this across with other people because this really helps the channel go a long way. If you're learning anything here,

way. If you're learning anything here, I've specifically turned off the ads because I kind of want to hang out with y'all and I don't want like an ad to get in the way when you first click in the stream because I, you know, I just hate that stuff. So if you're watching on other platforms or serving ads, just come over to this one here on YouTube and I've turned that off for now because

I feel like this is kind of the best place to hang out. All right. So in this conversation, you can kind of see the research that it did. And then my agent's file started to

did. And then my agent's file started to kind of bloat over like 2000 tokens. So basically I had some information about kind of what was going on in here and you can kind of see what happened. Right. So it's

like my agent's file was over 2000 tokens.

My memory is 2000.

Like it just everything was kind of up.

So that's 3000 tokens if we basically do some types of optimization to save per turn. And so

what I ended up doing is just giving it some instructions to do some optimizations and like how we can actually slim things down. So the current target of the AGENTS.md was over 2,000, I think,

AGENTS.md was over 2,000, I think, tokens, right? And after the

tokens, right? And after the optimization, we brought it down to 600.

Same with memory and so forth, because a lot of this stuff can just be recalled using QMD.

And so no information lost, same searchability. You just stop paying to

searchability. You just stop paying to load it every turn.

and the QMB already has an index. So want

me to draft a slim version. So one of my favorite things to do with my Clawbot is to have it draft plans for me, mostly because I just want to see what it's going to do. I'm an engineer, so I want to

to do. I'm an engineer, so I want to review them. But if you're not an

review them. But if you're not an engineer, it could also be helpful to learn to see what your agent's doing. I like to be a little bit more

doing. I like to be a little bit more verbose in that sense.

So yeah, that's kind of what's up. Okay,

so that's, I love to see the madman shipping. Hey,

shipping. Hey, Thanks, man. I haven't got to the

Thanks, man. I haven't got to the comments. I'm a little behind. Smash the

comments. I'm a little behind. Smash the

like button.

If you're building a team of agents, think about using the doctor role with the Phoenix protocol.

What is this? I need to look into that.

The freaking cron doesn't work. Oh,

that's interesting.

I have to check that out. And it's better than Q... What's better than QDM? Yeah,

than Q... What's better than QDM? Yeah,

I'm kind of curious.

What a coincidence. I'm working on a QMD skill right now. Yeah, so what's nice is the QMD stuff is actually built into the project and the agent knows about it, which is super handy.

Alex and Ray streamed the same day, yeah man, I saw him stream, I was inspired bro, this is so good, it's so good.

Hey man, I really appreciate the love for you. Yeah, AJ, AJ is the one who put us

you. Yeah, AJ, AJ is the one who put us on by the way, if y'all don't know, AJ when we were walking in San Francisco was the one who actually put us on the whole thing, he's like, bro, have you heard of this thing called Claudevon? I was like, nah, we gotta pull up, so that's what's up.

yeah, aloha, pulling up for the support, yeah, I'm creating assistant, uses agents, CK, almost open Claw, oh, interesting, yeah, also big shout out to Vladimir as well, I appreciate that, yeah, super memory, yeah, okay, we'll talk about that, cool, all right, so yeah, I'm curious about managing tokens, and this just goes from my AI coding days, right, so in here, what I ended up doing, I was like, great, we should have

something that is done at night, almost like the way human dreams, I want, I tried to connect the dots on and what happened and we clean up these things so we force helping Ray achieve his goals and then in the morning we can give a report about what has been done. So basically I'm telling my agent that I want to have a dream at night, kind of clean up these memories because sometimes the memories can get

filled up and they're not always pertinent to what our goals are. So we're

giving the instructions to the agent to make these types of things so it can learn for itself and help us out. So this is kind

out. So this is kind of forcing the agent down this route to because otherwise it's just going to keep doing what it's doing unless it has some directions from you. So this is a moment I thought this

you. So this is a moment I thought this is a really good point for us to share back and forth and I wanted to give you this insight. So

this is basically what it did. It gave us a dream cycle. So it's

it did. It gave us a dream cycle. So it's

going to generate two cron jobs, one that dreams at night and then one that has a morning brief.

And so in the latest update, there's actually some commands that it runs so they can either run silently or run verbose and tell you what happened. And so this is

happened. And so this is basically what we did. So here's the full picture. And so we're going to basically

picture. And so we're going to basically go from 7,869 characters from the agents.md file. So

that was the file that's bloating up some of my contacts down to 1,354, which is about 383 tokens, which basically removed a lot of duplication and different types of things. We thinned out the memory.md file

things. We thinned out the memory.md file

from 7,000 characters to 1,000, which is 1,600. Super amazing,

1,600. Super amazing, right? And then it created some new

right? And then it created some new folders in there. And so stuff that are in these folders are actually going to get indexed by that QMD thing. So this is looking pretty good.

thing. So this is looking pretty good.

And I said, now, let's go ahead and execute it. So that

was kind of the very important part there.

So once it did that, we're pretty much good there. So in order for you to set up QMD, all you have to do is just tell the agent to set up QMD, literally saying, hey, make sure you're up to date and just say, enable QMD.

I want to go ahead and enable the memory backend to be QMD. That's it. And if you want, you could even just go to where it says memory, OpenClawDocs, and just paste in this type of thing. It will already

of thing. It will already know what to do with it. So this is basically what it's going to do. It's

going to go ahead and install that into your actual thing, configure it, and it's going to do this type of thing. So it says how the

thing. So it says how the sidecar runs. The gateway basically

sidecar runs. The gateway basically writes a self-contained QMD home under agents, agent ID, QMD, and it collects the created collections in there, and now initializes QMD manager on startup.

So periodic update timers, usually every five minutes, will get armed before the first memory search call. So this is going to be

search call. So this is going to be awesome because not only do I have memories in my actual repo here, but I also have like if I have memory I go to archives I have imports so I imported all of my ChatGPT history going back from like 2023 all the way back to like today so all of these are like my ChatGPT histories but the ChatGPT history

is only like the summaries so these are all like little summaries here same with Claude so Claude from 2023 is all the way up in here so now I have a space in which my agent, I can ask any question about anything. Like I asked about that

about anything. Like I asked about that Italy trip.

That was in 2023. That was in a ChatGPT chat, right? And also in a Claude. So it

chat, right? And also in a Claude. So it

pulled them both together and pulled all the relevant information. It was amazing, right? So

information. It was amazing, right? So

this is how powerful this stuff is. And then also I have Claude

stuff is. And then also I have Claude projects. So all of my Claude projects

projects. So all of my Claude projects are in here as well.

And for you to export them, you just have to go to your settings in ChatGPT or go into OpenAI and just say, hey, go to the data controls and say export. I want to export all my data. So

export. I want to export all my data. So

as a California resident and maybe some places in Europe, given the data privacy laws, you're actually allowed to export whatever data this company has about you, which means usually all your previous chats and so forth. They usually just give you this

forth. They usually just give you this zip file of stuff and they just said, YOLO, here you go.

It's not really well organized, but in today's world of AI agents, you can tell them that, you know, I have my first live stream, I think I talked about how I walked through that specific process. If you want to know more, just

process. If you want to know more, just drop a comment and I'll kind of show you some more stuff there.

Also, please excuse the noise. I just

have a current, like a wireless microphone on and we're kind of going through this thing and I think there's some construction that's happening across the street.

So yeah, I don't think it's that bad. I

have a little processor on for vocals and stuff, but that's what's up. All right, so that's what QMD

what's up. All right, so that's what QMD is going to do. Search us through here and you can kind of see what it's doing here, right? It's kind of sets itself up so you can kind of read about some of the technicals there. All right, so that's

technicals there. All right, so that's that thing there. So once you have QMD set up, you want to run the doctor command, and it's going to tell you that if it's set up or if it ran through any errors. And so

after that, you're pretty much good to go. And so one of the things that I want

go. And so one of the things that I want to share with y'all is this memory optimizer skill. So this

is going to run when my device at night is doing its streaming, and it's going to run this script and the script basically just checks the various like agents MD, soul files and so forth and this is the like the tier one files right now. So in

this tier one path basically what ends up happening it tries to check to see how big your file is and so if your actual file is kind of getting out of scope it's getting a little growing a little too crazy it actually starts to call it out right so it starts to do a scrub of that and this is kind of what you know these are the files that we call out so these are the

injected file these are the files that we have here and then these are the actual directories that it's going to work through for everything.

And then it's going to actually give it a recommendation.

So at night while it's dreaming, it's also going to do this cleanup task to make sure that we aren't bloating up our memories.

And this is really, really cool.

So this is a part of a skill file.

And this is the actual instructions for what I call the dream cycle or morning brief.

And in here, the dream cycle is going to review today's files, search session transcripts, and connect the dots for recent days and reoccurring patterns.

So you can also just take this and just explain it to your agent that you want it to do something like this and you can build your own version of this, by the way.

You don't have to take mine, but this is something that I'm sharing with you to kind of give you some inspiration so you can kind of understand how to talk to these things and make them write software for you to do these things.

So these are skills, right? You can

literally get screenshots of here and just start throwing them into your agent and say, hey, Ray did this cool thing. Can

you learn about this? I want to do this for myself too.

so that's what's cool. This is the morning brief and your morning brief you can describe however you wanted but this is just a good starting point for me and this is what I'm starting to do and so there's a configuration thing that they call session target isolated and set delivery announced so you can kind of see for my dream cycle at night I have delivery none it's silent basically no message to the user so

you can do that you can run tasks that are cron tasks that aren't going to deliver messages to you because you don't want them I don't want to see what it's doing for the cleanup you know I'll get a report, something like that. And so we have two different jobs. Obviously, like

the dream stuff should be silent. The morning brief is kind of lightweight conversational. So it

just encourages me to start talking to my agent and kind of point in its direction of where I want to go, which I find very, very helpful. And so here's an actual template that it's going to run for the dream cycle and also the morning brief template as well. And

so this is this there, and then this is the actual skill file file that gets installed. And so in the skill file, it basically describes everything that we're talking about, how the OpenClaw injects workspace files here in the context window for every single model call.

And so the default templates in the organic growth cause these files to bloat over time.

And so that basically wastes thousands of tokens per turn. And so here's the optimization workflow that you can actually do. And then you can actually, this is just instructions for the agent for it to do and ClawCon.

So this is where it's applying things and this is where the dream cycle kind of kicks off.

And this is never delete information, move it to searchable.

So it's going to move information into the memory and then star MD files.

And then always create drafts first.

So create drafts and drafts and show the user before applying.

Respect the sole MD.

Personally, it's worth tokens.

Do not optimize it away.

And the convention must be self-documenting.

So put memory notes in agents MDs so future sessions follow them.

and put the rationale in the convention so it's searchable and not loaded on every turn. Cool.

turn. Cool.

And then here some actual templates for the actual agents MD So this is what going to keep the agent in scope So it going to reference this and it going to say okay cool This is kind of how I should organize the information that I dream about at night in a way that I can keep loading up for every conversation so I understand what's going on in Ray's life. Obviously, you'll want

to tune this to your own thing, but I haven't publicly published this because I'm still refining this. And

this is something that, you know, my Discord members have access to. I haven't dropped this specific thing

to. I haven't dropped this specific thing because I literally just made this this morning.

And I'm going to put that into the Discord right after the stream so that way the folks there can have it there and you can experiment and kind of take it yourself.

So that's MemoryMD template. Here's the

MemoryConventions template, MD. So this

is kind of the way that I would describe like how these conventions are made.

So if there is something that's going on, like after the audit, it will kind of do this stuff in here.

So it'll tell me like, hey, this is kind of how we, this is what we found and this is kind of what was bloated.

and so this three-tier system is based off of the research that I was doing to understand progressive disclosure for memory.

So progressive disclosure and what I'm trying to say here is, you know, like when you have an agent has a task, they're going to consume some memory to put into the context window to do some type of action.

And so in here, as you can see, there's three different levels.

One is like every single time you ask it, it's going to load up the agent's MD, the memory, the soul.

and that's kind of what gives life to the agent but there's other things that you don't necessarily need you don't need like the 50 grocery list items in the conversation for every single thing you're doing because you could be asking it about generating images for this launching a subagent to go you know do some research for you and like if it always keeps loading the grocery list in those tasks then it's

not really that useful right so that's where you kind of want to extract some of those files out and when because it's working on a task the agents are smart enough to know it's like oh I'm working on a grocery task. And I know that we stashed those files over here because my agent's file shows me that there's an index saying grocery list that's still active. Now I'm going to go grab that. So

grab that. So that's kind of why it's called progressive disclosure. Whenever the

progressive disclosure. Whenever the agent's doing a task, you don't always want to flood that desk space with everything it needs. It's only

at the time that it needs it. Like I need a hammer now. Let

needs it. Like I need a hammer now. Let

me go down to the tool shed because my little thing says that the hammer's in the tool shed. And so

that's kind of the way that you want to think about these conventions, but you don't really have to think about them too much because I kind of done some of the research for you. But you can also talk to the agent and have this conversation back and forth so you can learn as well. So that's kind of what it is. So in tier one is what I call like very expensive.

So like I'm telling the agents that this is the three-tier memory system and each has different costs and characteristics associated with it, right? This is not me making this up.

it, right? This is not me making this up.

This is from my coding experience, but also Cloud Code SDK. You

can look this up. All of this is publicly documented and also and also the code base. So if you look at the code base from OpenClaw, which is on GitHub, it's open source. This is the amazing reason why it's open source and this is the big reason why Peter wants to keep this open source and that's why OpenAI has created a whole foundation so that we can actually see the code and see

what's going on, right? I can't see in the ChatGPT memories, but I can see into the code for OpenClaw and see what it's doing, right? So this is really cool and it

right? So this is really cool and it allows me to actually talk to you and us have a discussion on what's the best approach and then I can make my own changes and see what it's doing.

if it works really well. Obviously, this

is something that not everyone's going to be able to do, but I'm going to keep sharing this stuff. And

if you like this type of thing, definitely subscribe, because this is probably the most life-changing thing right now on the internet for a good reason.

And I think there's a lot of potential here in it as a product. So I'm really excited about how this stuff really, really works at a low level, at an in-code level. And I don't know everything, but I just know how to ask agents how to ask me to figure out how we can find this information.

So obviously this is how everything's set up. You don't even have to know how this

up. You don't even have to know how this code works, but I'm just kind of walking you through what is tier one. This is all read from the code, and this is exactly explaining what stuff is put in there. So

only put information here that genuinely affects every single interaction, orientation, not encyclopedia. So that's

pretty cool. And then what we call tier two, this is searchable, and this is on demand. This is what's considered cheap.

demand. This is what's considered cheap.

So anything in the memory files are indexed by OpenClaw's memory indexer.

So that's going to be the QMD thing that we talked about earlier.

So they cost zero tokens until the agent explicitly searches for them and retrieves them.

And so that's the memory stuff with the daily logs.

That's anything that has memory slash topic and anything that's in there like brand guides, project notes, things that we'll talk about and we'll get pulled up just naturally through QMD.

So that's really cool.

And then there's a full file read.

This could be on demand.

So anything that's on disk, the agent can obviously read it.

and only like tool tokens and all that stuff that get loaded. So that's the progressive disclosure pattern.

And the memory MD basically will just now act as a table of contents. So this is a term that I also used a lot in when I was AI coding as well.

So like the agents MD behaves the same way. And in Claude Code SD decay, the

way. And in Claude Code SD decay, the thing behaves the same way.

And basically the way this works, you can nest memory files. So you can nest either memory, you can nest agents MD files in different folders, and they can tell it like it's almost literally like having a signpost in a freeway. So if you're

freeway. So if you're exiting on a freeway, it's going to tell you what the names of the roads are before you get there, right? So you're like, okay, that's the

right? So you're like, okay, that's the exit I need to take because I need to get off on, you know, whatever, McDonald's lane or something like that or those types of things. And

that's kind of the way you think about them. It's like every single turn is going to be a new folder.

And if you have like an agent's MD or if you have like a memory file that's in there, it's going to really kind of point to the next direction of where the agent should go. So that way it doesn't get

should go. So that way it doesn't get overwhelmed right?

and that's the way that you kind of want to treat this type of thing here and this will tell it what to focus on or like what key preferences and bullet points.

So this also describes the bloat patterns which are super important.

So agents will bloat if they like I was telling you a little bit earlier about these different files and how important it is to kind of keep them lightweight but also provide these little indexes so that it can start to crawl through.

And the way you want to think about them is like literally like an encyclopedia index right?

You can't memorize everything, but if you know where to find where the cheetah is, then you can go to the C section and start to search up and see what's kind of in there.

So that's basically that. Here's the

target token budgets, and that's pretty much that.

So this file is installed as a skill in my actual agent. And so in my agent, we can actually see it here as a skill file.

This is a memory optimizer, and you can kind of see how it organizes itself in here.

So this has like a dream cycle. These are

the templates, and this is the tier system.

and here's the script. So that's the audit script that's there. And this is the skill file that we talked about earlier that sits in there.

And when the agent kicks off to do, it's like CronTask to dream at night, it's going to know to pull that information. So like here's some example skills and stuff like that. Let's see if I can pull this up here. So this will just define like what specific skills I have, what coding agents I have, what specific, you know, tooling that I have in here.

so that's kind of what gets loaded in and I don't have to keep describing it over and over again.

So that's basically a big primer on memory and all of this stuff is documented in the code, in the documentation. It's a little bit

documentation. It's a little bit complicated to read but if you haven't set it up, like kind of long story short again is make sure you talk to your agents and get the latest update.

So if you just go to like literally copy and paste and say hey OpenClaw now is 2026.2.15, go ahead and update yourself to this version, do the update. The next

thing you want to do is like, okay, I want to take advantage of OpenClaw's QMD. So I want to, you know,

OpenClaw's QMD. So I want to, you know, this is an experimental thing and I want to go ahead and make sure that we have this enabled for my memory. And so once it goes and gets that

memory. And so once it goes and gets that set up, you let it do its thing and it's going to do its installs and stuff and it'll be good to go. And then the last step is

go. And then the last step is you can take some of this research that I did. I'm about to open source it. But

did. I'm about to open source it. But

first, I just have this is currently my Discord and I did some research and I kind of walked you through a little bit earlier of how I had this conversation with my agent so that we can basically create the system so that it would do dreams at night.

So my concept is for it to dream and then give me a morning brief cycle.

So with that dream cycle, basically what we're doing is just kind of cleaning things up at night, making sure there isn't too much contact stuff into too many areas.

So that way I keep things really efficient for my agent to do work that he needs to do for what we're doing.

So I think that's kind of like the high-level summary of it all, and I feel like there's a lot to talk about just even within these things, but I wanted to pack everything in as much as possible into this live stream, and I'm going to get into some questions next so that way I'll answer you guys' questions, because there's probably a lot of y'all in the chat right now who have a lot of questions

and want to get things answered, so I'm going to definitely try to see if I can tackle those before I basically head to the beach, right?

So let's go ahead and start to take a look at that. So open source at 9, 15 bucks a month is tough. Yeah, I got you.

I have people who do pay that 15 bucks a month right now. And those are my members and I want to make sure that they're taken care of, but they're also my early adopters. And so they give me that

adopters. And so they give me that feedback and refine it so that by the time it gets to you, it's more refined and like you don't have like as many crazy issues.

And I'm not saying that there's crazy issues, but I feel like that's a perk of having that type of thing. And I know not everyone can do it.

and so I want to make sure that, you know, I mean that's why I have these perks and I also can't deal with everyone's issues as another part of it.

It's like there's only one of me and people who are able to pay into the Discord, you know, have some like general like huge amount of interest in something like this.

And so they have the discretionary income and so it's just another type of thing that kind of helps me eliminate because I have hundreds of DMs from, I don't know if they're bots, I don't know what's going on.

I mean, a lot of people can send DMs and stuff. So that's another tough problem to

stuff. So that's another tough problem to have, but I definitely appreciate it. So if you like can't pay the money, you can also, you know, watch this, literally just watch this type of live streams and stuff that I do for free. And

so my goal is to literally live stream on the weekends. So

usually Thursdays, usually on Saturdays and Sundays. But

and Sundays. But today during the week, I'm going to do some extra streams because I've been kind of shifting my situation right now with, you know, being in Hawaii, obviously. And so your support goes a long way.

All you have to do is literally just watch the stream. I turned off ads for you so you can actually watch right now.

So there's a lot of different ways to support the channel. Literally just

liking, sharing it with another friend, subscribing.

It goes a really long ways too. And this

will be available for those who can't do that $15 a month later on.

So definitely stay tuned for that as well.

Have you saw Kimi Claw? I saw that Kimi offers Claw and also Baidu now also offers OpenClaw as well.

So Baidu as a company offers it, which is like the Google of China, which is really fascinating.

Hey, Paul, how's it going? Welcome, man.

I love your new live stream setup, by the way, with Ecamm Live. It's so cool.

If you haven't had a chance, Paul goes like crazy. He's been all in on Codex,

like crazy. He's been all in on Codex, and he's been doing some really cool demos.

If you're into iOS development, he has a lot of great material on how he uses it for iOS development.

He's been playing a lot with Spark recently. So Spark is the latest coding,

recently. So Spark is the latest coding, like super fast inference that is provided by OpenAI.

and he's made some games and he's doing some really cool stuff in that space. So

definitely go ahead and check out Paul if you haven't had a chance. He's really,

really cool.

Domingo, thank you so much for joining in. Stefan, Fire Idea, I think this is

in. Stefan, Fire Idea, I think this is for other memories and stuff. This is pretty dope. Yeah, I mean, any questions you

dope. Yeah, I mean, any questions you guys have, let me know.

How do you have OpenClaw search the internet through Brave without getting prompt injected or search the internet in general without running into an issue?

So there are a couple different things. I

try to steer it to specific sites that I know are pretty safe.

So I mostly just point it up to the official documentation from OpenClaw, right?

And a lot of times I'm more so reviewing it myself on my phone anyways, so I'm just giving it random snippets of things.

I don't necessarily, actually, I don't even have Brave enabled on my agents. I

only have ExaAI.

So Exa.ai is my primary search engine.

And so they have a search feature there.

And I also have a skill that I'm going to share that's open source.

But you can literally just point your agent to Exa.ai, give it the API key.

And a lot of those data sets are already pre-filtered out.

So you're not picking up a lot of junk.

You're only picking up the specific information or queries that your agents need.

So Exa.ai has a specific thing called ExaCode.

and especially for code related tasks is extremely good at searching through GitHub searching through documentation It not going to prevent prompt injection but it reduces a lot of likelihood behind them And so I pretty sure this is another billion industry is to have data sets from Excel like that that are pretty much filtered out even more. Another MCP tool I like is

even more. Another MCP tool I like is called ref.tools, and I use it a lot for

called ref.tools, and I use it a lot for AI coding.

So I'm trying to figure out how I can get a script or CLI for that so I could just use my API key and have it, you know, because I'm doing most of this stuff is doing coding tasks for my agents, you know so they're basically doing that type of thing but yeah it's a good question obviously you have to beware that this is a possible thing so what's the Anthropic Pentagon thing that's going on I

haven't followed too closely on that let me see Anthropic Pentagon let's see what's going on here exclusive Pentagon warns Anthropic will pay a price for as huge escalates okay wait wait wait this just happened right now this is like breaking Breaking News, Breaking News, Breaking News, Okay, So what is, what is going on, bro?

Axios, Exclusive Pentagon Use Anthropics, Claude, and Maduro, Venezuela, Okay?

Wow, Anthropic Won't Give Pentagon Unrestricted Access to Its Models.

Pentagon Threatens To Cut Anthropics In AI Safeguards Dispute.

Oh my God, Bro, This Is Crazy Right Now.

So, what is this? So, I don't know, is reviewing its, I can't even scroll on half these sites. Okay, I'm blocked

these sites. Okay, I'm blocked everywhere.

Saying that while having the decoder, can we actually open, I can't even read any of these links.

Come on, bro. Like, these are... You see

why we need AI? Like, this is kind of trash. All right. So, Anthropic continues

trash. All right. So, Anthropic continues to refuse the Pentagon unrestricted access to its AI models. Oh, so the AI, the Pentagon is literally asking for unrestricted access and they're like, bro, we just can't give you the keys because this can get into bad actors' hands and we know when that happens.

Have they ever talked to Pliny the prompter? Like, do they know who that is?

prompter? Like, do they know who that is?

Like, they should just go talk to Pliny.

That guy can, like, jailbreak these models in five, not even five minutes, bro.

Like, he's just one prompt. But I think I know what's going on here if I think, kind of read between the lines. This is kind of what I'm thinking is that the government wants to have this because they can already do it.

If Pliny can do it in one prompt, they're not saying that they need to do this.

They're posturing that the company should allow this as a backdoor, right?

Like, that's the same problem. It's like,

if you give, like, a thug your key or something to your house, he's going to give it to someone else that pays him a lot of money, and then all of a sudden everyone else has access to your house, right?

like, that's not a good thing to do. And

I feel like, yeah, I mean, I think that's dangerous, if you know what I mean. So yeah, I don't know. I mean, yeah. So insist on drawing

know. I mean, yeah. So insist on drawing the line at two things, mass surveillance of US citizens and fully autonomous weapons. So a senior government official told Axios that negotiating individual use cases with an anthropic isn't practical. So OpenAI,

practical. So OpenAI, Google, and XAI have been willing to work with the Pentagon, the official said. All

three have agreed.

Okay, so a lot of times this is kind of where I have a skill that I'm going to open source.

I'm going to go to Claude.ai.

I need to drop into my agent and install this.

But basically what I do is try to check for like fluff and stuff.

Can you review this document?

Okay.

News.

this news article and use my truth skill to help me figure out what is real.

So I made a skill. I have an app called Truth Torch. I'm basically turning it

Truth Torch. I'm basically turning it into a CLI slash skill, so that way my agents will do this. And

so, yeah, rhetorical analyzer skill. So

I'll show you the output. This is really, really cool. So a

output. This is really, really cool. So a

lot of times when you read news, a lot of times they'll write it in this way where they'll try to pin one thing against the other and then they'll add other claims but not back them up and they'll either back them up by emotion, they'll forgive facts and all this other crazy stuff and so this is kind of what's cool. So now it's Anthropic, I'm using Anthropic's models to do these types of skill

checking for me. So you can see it's doing all this stuff on the internet to check all these claims of what's going on and so now it outputs it as a skill. So this is really cool because I can now hand this off to my own agents. And so I made this skill in Claude Code and this skill is totally transferable to everyone else. So Anthropic is refusing

everyone else. So Anthropic is refusing to give the Pentagon unrestricted use of Claude, insisting on guardrails against mass surveillance. And

the Pentagon is threatening to end the $200 million contract as a result. So here's the core points. Anthropic draws

points. Anthropic draws two red lines, no mass surveillance of Americans, no autonomous weapons. And the

Pentagon wants an all lawful purposes standard with no company imposed restrictions. So OpenAI,

Google are more than willing to comply with Pentagon demands.

Dario Amore published an essay articulating why democracies shouldn't use an AI in ways to make them resemble autocracies. And so this is a

resemble autocracies. And so this is a system of government by which one person has absolute power facts. The dispute is happening against

facts. The dispute is happening against the backdrop of ICE shootings in Minneapolis and Anthropics-Palantir partnership. Oh,

Anthropics-Palantir partnership. Oh, interesting. Okay. Okay. What

interesting. Okay. Okay. What

what they say, what they actually want to say. See, this is, this is, this is why I

say. See, this is, this is, this is why I love using this stuff.

Yeah, I'm definitely going to share this.

I'm, I'm dropping this on the Discord.

I'm going to drop this right now on the Discord. This is

too good. This is, this skill is too good.

We need this, and I will open source this. Let me just, I have so many good skills right now. Let's

see, capabilities. Yeah,

Yeah, this is what we have to use AI for.

I have a crypto scheme evaluator too, which is pretty fire. It's based off the same rhetorical analyzer download. Cool.

I'm going to put this in the Discord.

Discord. Okay. Oh, shoot. I'm going to mess up my thing. Okay. So here is my general. I'm going to put this in the

general. I'm going to put this in the skill. I have a skills chat.

skill. I have a skills chat.

and Skills. I'm going to create one, Rhetorical Analyzer. Here we go.

Rhetorical Analyzer. Here we go.

Rhetorical. Here we go.

Enter. And then I'm going to post this.

And then I'm going to post it in here.

Hold on, my bad.

I'm going to drag the file up in here. So

this is Rhetorical Analyzer Skill. Okay,

cool.

Here you go. Perfect. Okay. So you can download it here and I have this skill file so that way you can install it. And

so it's a .skill file and you just give this to your agent and they will know what to do. But yeah, so folks in the Discord,

to do. But yeah, so folks in the Discord, you definitely have that. So that'll be in the community section. I have a little Claude, I have a skills actual like post type of thing. And I also have an app architect skill. This one is incredibly

architect skill. This one is incredibly fire. If you haven't had a chance, this

fire. If you haven't had a chance, this is what I use to build apps.

and basically one shot them. More on that soon. So yeah, this is basically how I

soon. So yeah, this is basically how I organize the skill.

It's a pretty simple thing. Do I have the code files? I think I have some other

code files? I think I have some other stuff. Okay.

stuff. Okay.

Anyways, as you can see how this is amazing and how good it is. So Anthropic

is a principled holdout among AI companies, but that's principled stance is complicated by existing defense contracts and financial interests. See, I don't know a

financial interests. See, I don't know a lot about the background. So now I can actually start to kind and I want to ask more questions, right?

That's the whole point, rhetorical analysis, right? Like you want to have a

analysis, right? Like you want to have a co-pilot to help you think about these things, right? The evidence proof table.

things, right? The evidence proof table.

So the Pentagon is considering ending the anthropic partnership. Axios reported the

anthropic partnership. Axios reported the citing, and this is confirmed by these other sources. And now we also have a

other sources. And now we also have a claim about a 200 million contract stake, and this is obviously confirmed. Pentagon

renamed Department of War, referenced casually, confirmed from administration rebranding. January 9th, AI strategy

rebranding. January 9th, AI strategy memo, article claims no link provided, but referenced by Reuters in the original reporting.

OpenAI, XAI, Google. So this is an anonymous senior government official. So single anonymous sources,

official. So single anonymous sources, companies haven't confirmed publicly. So

this is actually, this is like a thing to kind of call out, which I find super helpful. One of the three alleged all lawful purposes. So this is the same

lawful purposes. So this is the same anonymous official, right? Unverifiable.

Dario Amadei blog post, real essay, fatal shootings, obviously, and so forth. So the article doesn't mention that CLAUDE was actually used in the Maduro raid via Palantir, which is a massive piece of context that came out of the WSJ reporting and significantly complicates Anthropics principled stance narrative. So

narrative. So here's the rhetorical appeals. So they

use ethos and pathos. So they try to use the credibility on these types of things and his own words, and this is like the sourcing for the claims, and they use emotion here. So they try to like, this is where it gets interesting.

So the article juxtaposes anthropic ethical stance against the recent things that happened in Minneapolis and the Palantir Ice thing. So that's really interesting. And here's the actual like

interesting. And here's the actual like logos chain, right? So

this is how the thinking actually works.

And this is why Opus is an incredible research model, right?

You can point it at a skill like this and it actually starts to think like a lawyer.

Just all I gave it was a text file. And like, look at this, like we're analyzing an article, like a way a PhD law student or something would be thinking through these things. And it's

helpful for me because now it cuts through the fluff, right? And I'm

not getting all worked up emotionally just from reading an article. I'm just kind of trying to think

article. I'm just kind of trying to think about them from first principles. So

here's like unexamined assumptions for this whole article and then the Steelman version and like what's actually missing.

So this is kind of crazy, right? So yeah,

this is a bottom line for you. The facts

in this article, check out. The framing is slightly

check out. The framing is slightly sympathetic to Anthropic, but not dishonestly so. The biggest

dishonestly so. The biggest thing is it underplays the Marduro raid and the Palantir pipeline, which complicate the principal holdout narrative significantly.

This is real developing high-stakes story with today's supply chain risk threat being a major escalation worth watching.

Cool. Now I can give this skill to my agent, and then my agent will be able to do this. So I could just give it an

do this. So I could just give it an article, say, hey, can you take a look at this article and just kind of run this through me?

And then you can tell it to reshape whatever responses you have. This is the way I like to read information, but you can take a skill file and have it reimagined for you.

but look at what it did for the research, right? It did all this research for me,

right? It did all this research for me, like 10, 20, 30 different results from all these different types of things from all these different organizations, NAPC News, Wikipedia, NPR. That's faster than I

Wikipedia, NPR. That's faster than I could do it, you know? I didn't spend the whole thing.

So that's really, really cool. Okay,

cool. That's what's up. Let's see what's going on here. So,

yeah. Let's see what's going on. Also,

Ray Fernando, you might have missed it, but Alex was in the chat.

Yes, I think I missed it a little bit earlier, my bad. Even check Rioter's article, yeah, during the Vendels, yeah.

It likes GT5 Nano, using open router with auto results. Mahalo, hey, thank you so

auto results. Mahalo, hey, thank you so much.

My OpenClaw is named Claudio. Alaka,

Alaka'ai? You were instrumental in my original setup of OpenClaw. Excited to

see what you got cooking.

Yeah, mine's called Honu. That's awesome,

man. I really, I love that name, Claudio.

That's so cool.

Drop the link to join your Discord on X.

Oh yeah, so if you're on X right now, let me see, X.com.

It's rfer.me slash Ray Fernando. So this

is my live right now. I'll put this in here.

Let's see, it's kind of weird watching me. Rfer.me slash Discord. I think, yeah.

me. Rfer.me slash Discord. I think, yeah.

Okay, so I just dropped the link for the Discord. And if you haven't had a chance, go ahead and check it out there.

it's also pins here if you're on YouTube you can check out the comment there so rpr.me slash discord.

rpr.me slash discord.

That way you can see here if you want to go ahead and join the discord as well so yeah that's what's up.

I haven't seen the message I'm still kind of scrolling back right now because I'm a little bit behind on the chat but I wanted to make sure I The message is, Wayne, oh, I have so many good skills right now, yes, how's it going Wayne thank you so much, Wayne over from Convex, I appreciate you guys supporting the streams and supporting everything you guys do, I mean, providing the space, we had to fire episode

two for our Thursday vibe check, so that was really amazing, Alex Finn is in the building kind of doing some demos, this is when things were really heating up, so that's what's up, yo, yeah, I got you, fam, that's a really nice skill to have, yeah.

Rhetorical Analysis Skills is definitely fire. Let's see, okay, cool. Anthropic

fire. Let's see, okay, cool. Anthropic

has been feeling themselves, trying to control everyone's behavior. This is going to

everyone's behavior. This is going to backfire. Yeah, I mean, I can see that

backfire. Yeah, I mean, I can see that too. Could be a rough time to invest in

too. Could be a rough time to invest in Anthropic. Yeah. Alex been official. Who

Anthropic. Yeah. Alex been official. Who

wins the government contract is set for a long time. Yeah. Just got them on that.

long time. Yeah. Just got them on that.

Oh, yeah. So he's got another Mac Studio about to set up a bunch of local models.

So this guy is absolutely destroying it.

Alex is going crazy.

he just sent me a picture of like two of them and I think he already has one and I think there's a couple more right but this guy's gonna bro they're gonna break into your house bro you better watch out okay yo how you living okay cool Alex Swing crazy right consider anthropic a supply chain risk okay let's see thank you answer is very helpful please don't repeat everything you

said in the last hour Okay, so I'm converted, oh, okay, cool. I

converted, Quen, Vision, and Text to CoreML, and I have, oh, whoa, this is crazy. so you're trying to deploy, create

crazy. so you're trying to deploy, create custom DMs to deploy on AWS and the advice. I don't know if I have much

advice. I don't know if I have much advice, because you're just, what are you trying to do with the CoreML stuff? Like,

you have, like, those separate, right?

And then you create custom, VMs to deploy on AWS. I mean, my only advice is to ask Opus 4.6 or Codex and to have you use Exacode or ref.tools to like literally just say

ref.tools to like literally just say this, this is what I got, this is what I want to accomplish, help me walk me through that process. And

it'll take you through those steps to do that. That's

that. That's basically how I got my CloudBot set up, you know, like with AJ. So AJ walked me through how to get stuff set up in AWS. And then afterwards, like I set policies, I like blocked it down, I was doing and all kinds of cool stuff.

And I learned a lot about the AWS infrastructure just by talking to, in this case, Opus 4.5 at that time.

And it was super helpful.

And I even just fed it screenshots of stuff.

But you can control a lot of things through command line.

And I actually think that AWS is easier to control over command line than it is over the UI.

It's kind of funny, actually.

So yeah, that's probably my best bet if you're thinking about that.

So let's see what's going on.

I'm a little behind the scenes.

Do you ever receive token mismatcher?

I haven't seen that yet.

I don't get Claw to produce HTML reports, but I think I did once. I'm going to hook up my Claw up to Notion, and so I'm going to have a lot of different reports and various things go into Notion.

I have a lot of workflows in Notion, so I'm going to basically do that as well.

Hey from the UK, I'm concerned of re-downloading a bad skill code. Where

should I get QMD, or is it already installed?

So QMD is already installed as part of the OpenClaw repo, so you could just literally point to the official documentation.

I always say stick to the official documentation and then give it links sometimes and it's going to know a lot about that.

And as far as skills, sometimes I take other people's skills and I read through them and then I try to prompt my agent to build them, not necessarily looking at their skills and say like, oh, I like that.

So I try to basically make my own specs in some ways and then iterate back and forth.

So just like I do with coding, sometimes I go into plan mode.

mode. So I tell my agents or Honu in this case that I want to do a little bit of planning before I execute on making the skills so that we can kind of have this back and forth conversation. I would say

conversation. I would say that's more advanced mode. I don't think everyone should be doing that, but as an engineer, I want to make sure all the details are implemented correctly. And, you know, I think that's

correctly. And, you know, I think that's kind of what's up.

South by Southwest AJ. Oh, I need to be going there. Yes. So QMD, for those who

going there. Yes. So QMD, for those who are wondering, and ClawCon.

It's a query markup document.

So this is just a fast way to be able to search through your documents, especially if it's written in markdown.

So this guy came up with this great technique to do that.

I think he from Shopify and it super important especially in this age What nice is basically that you don have to pay a bunch of money As we saw like once you start querying up to the millions of documents you can pay upwards of you know 20 bucks a month to like 300 to 400 bucks a month for stuff like that And so you basically get that for free and your agent

gets to use that on device which is really really cool So I going to hop on the Discord and see what I can derive Mahalo Thanks man Yeah And let me know what skills and things you want to see I going to probably start putting this stuff in the chat right now So I running it on my MacBook Air Gwit. Working on how to get infra set up

Gwit. Working on how to get infra set up for a couple hundred sales reps. Wow.

Okay, that's what's cool.

The locally hosted model. Okay, so the way that you want to think about this, if you're serving stuff locally, you can use like Tailscale or something and then you'll set up something where you'll serve it.

Like right now on the Spark, you can just tell like an NVIDIA Spark, you can like I could serve models from it. If you're serving on like MLX,

from it. If you're serving on like MLX, there's like a serve command. LMStudio

also provides it for free.

OLAMA, depending on whatever you're using, MLX stuff, you can serve that back out. So you're

out. So you're basically turning your Mac into a server.

And if you have tailscale, you're basically going to create a tunnel. And from that, you're going to

a tunnel. And from that, you're going to basically have AWS or whatever your agents are going to use.

It's going to know about the specific tunnel that it has access to communicate to and start to send requests towards it. And so, yeah, you may run into issues depending on your throughput for requests, for responses too that could be delayed.

But yeah, that's a good, I mean, you can start this conversation and see what's going on. I'm

trying to set up Oracle Terraforms, but I've gotten conversion to work. Apple doesn't have it where you can just download CoreML models from the App Store.

Yeah, but you can download PlanVision to iCloud. Oh, interesting. Yeah.

iCloud. Oh, interesting. Yeah.

I can't download Core ML models from the App Store. So you can serve it from like

App Store. So you can serve it from like a Redis or something, right? Or some type of CDN, can't you? Like, that's probably what I would do. Just like serve it from a CDN and then for people to not eat up your bandwidth, you can just do a check, a device check to make sure it's like an actual Apple device, not someone just hitting the endpoint.

point. And then once you do that, then I think you're pretty much good, right?

Yeah.

Okay. Got it. Let's see. Okay. So what?

The chat would love Anthropik to help them create an all-seeing eye.

I'm confused. What are you guys actually advocating for? I think Alex didn't kind

advocating for? I think Alex didn't kind of hit this on the head earlier is that America needs to stand up and support open source models. We're

both very American and very proud of the country and it's amazing the innovation and the freedom that you have to speak. I

can speak about whatever I want. I don't

have to say I praise the government, whatever. If I don't like what the

whatever. If I don't like what the government's doing, I'll say it. If I

don't like what this company's doing, I'll say it. That's the whole point. This is

say it. That's the whole point. This is

why we do live streams too, because that freedom of speech is super important.

and I think open models are one big part of that too, right? Like we're the fact that we have to rely on a Chinese model that's like the frontier thing and they're leading that whole system is like it's kind of embarrassing, right? I think it's people are going to prefer to use those things that, you know, while they're open source and stuff like it just doesn't really make any sense, right?

I can understand from a we want to keep everything protected or if we expose too much secrets and everyone else can kind of copy them.

So, I mean, it's a very complicated issue, but also we should be leading in that as well. Right.

So I think tooling and all these different types of things, obviously, these people are out to, you know, prioritize the things that they're working on.

That's going to generate the most money because that's basically all they have.

Right.

like compute the incoming people demand for meeting the stuff and then all these like enterprise contracts and government contracts.

But the government contracts come from us, like us taxpayers who give the government money, who, you know, like the Pentagon got their money from taxpayers, right?

They're not just printing it out of thin air. And so, yeah, there's this, I mean,

air. And so, yeah, there's this, I mean, this can go in a lot of different directions, but I do believe like in freedom of speech.

And I believe that I feel like. Like

everything should have their checks, if you know what I mean, So there shouldn't be too crazy as far as like one like company having all the power like it just doesn't make any sense.

And then also like one government agency just having the all seeing power just is also just as scary Right Which is actually why I really appreciate the Apple stance for security because Apple said we going to engineer these devices so that even if their data ends up at a data center that they can hand over the keys and then all of a sudden your customer data is released. It's just the standard

is released. It's just the standard security pattern, right?

They just said, well, we don't want to have access to it. That's the customer's data. That's

data. That's secured via the secure enclave, right?

Like, if you wipe the device, the key and everything for reading out those bits, you can't do that, right?

Like, you need the physical hardware all stacked together for the device to read everything. So,

yeah, very interesting, like, times that we're in. I think

we're in. I think there's going to be a lot of interesting installations. Hey, it finished the QMD

installations. Hey, it finished the QMD skill. Cool. Yeah,

skill. Cool. Yeah,

Let's see what's going on. Okay, so cool.

I agree. Grok 3 is a good start, but the Chinese models leaning open source is quite a bit of the problem. Yeah, I don't know. I

think there's a lot of interesting innovations that are coming from China, which are fascinating.

Like, they don't have the same resources that the U.S. does. So

in essence, they're able to discover a lot more, like, or they're able to provide more focus to areas that are bottlenecks for them in terms of throughput right?

So because they don't get the same throughput because they don't have GPUs, they're now able to tweak other parts of this really complicated system and then discovering new techniques for like longer memory, compression, applying these other techniques that are kind of really out there.

But then they're able to yield results, which is kind of wild.

and it doesn't surprise me because this space is really, really large and the use cases are now starting to reveal themselves the more that more people adopt AI and start using this stuff. So I

this stuff. So I think this is a really cool step forward if you know what I mean. So yeah. Okay,

cool. So I'm going to go ahead and pop on out now and take care of things and head to the beach. I really

appreciate everyone for tuning in. If you haven't had a chance to join the Discord, go ahead and take a look at that.

If you can't do that, just go ahead and like and subscribe. Make sure you hit the thumbs up and share this with another friend. This really goes a long way. I've

friend. This really goes a long way. I've

turned off advertising, and so the people who do support the Discord and all that stuff, you're really kind of helping this community build on forward, and I really appreciate you guys. We're just trying to grow this channel and get more cool stuff going on, and also, you know, working with, going to be working with some partnerships with different sponsors and different companies.

that'll be like sponsoring the stream which would be super helpful and it'd be kind of nice that way I can kind of keep these you know turn off the ads because the thing I hate more than anything is like whenever I go to Twitch or something like there's ads all the time it's like and they're not even good too so it's like I value people like you hanging out with

me and hanging out with everyone else and kind of chatting about all the stuff that's going on in AI so we really appreciate you appreciate you appreciate it thanks for the lives for good content I love that about Hawaii don't yeah I mean like it's amazing right it's beautiful outside I gotta go outside go to the beach it's been raining a lot because if you haven't been in Hawaii it's been

like super windy it's still pretty windy right now there's a lot of there's been a lot of rain so like now that there's some sunshine and like I'm not gonna waste this moment and being inside right now I'm just gonna go outside and touch some grass but this really helps the mind kind of heal and kind of get centered again that way I can just come in straight up deliver value

and not have to feel like I'm just burned out inside all day so I gotta get some of that vitamin D gotta get some you know get the dark skin going on so So, adds OpenAI at the chat.

Yeah, I appreciate everyone.

I'm looking forward to the beach after coding in Hawaii in a couple months.

That's the setup right there.

That's the setup.

Great stream, everyone.

For sure.

Yeah, no, I appreciate that, man.

I mean, it's free.

This is what we do.

And I appreciate everyone else who supports the channel.

Got some really cool stuff coming up next.

I'm really excited to kind of show you.

And so, yeah, we got this guy.

More stay tuned.

I'm playing around with some other stuff.

and then I'll be able to, after my Discord community kind of gives me some more feedback on my memories and the dreams thing, I'll be able to share that skill publicly.

I'm actually going to make like a whole repo plus like a similar thing to like the Claw Hub, but it's going to just be for my community and my people.

That way it's kind of less, not like, I don't know, it's more specific to us.

And if you're kind of vibing with the stream, it's just a really good place that you know that like I've spent a lot of time, and my community members have spent a lot of time crafting these things, working with them together.

And I feel like as a community we can kind of have our own little hub There just be our own little section type of thing So I learn a lot and everything I learn I going to try to share back with you So I appreciate that So yeah the way that you can do that is literally supporting the streams If you just go to the Discord if you want to send me a DM, I can send

you a link or something. I really

appreciate everyone who's been able to do that. It actually does.

What's really nice, it actually does help grow the business.

So that allows me to get more equipment and different things like this, like the open cloth thing, the Mac Studios, the editing, the high quality stuff that I really try to do.

I also have the fortune of being in Silicon Valley, which allows me to get access to these different events and stuff. So I have a whole IRL setup, which this is actually part of the whole system.

So I'll be able to walk around with this camera and you've seen me doing some of those tests live already. And so I want to bring more of

already. And so I want to bring more of those types of things in because I want to bump in, and literally being in San Francisco, we bumped into the founder of Factory AI while we were hanging out with AJ and also hanging out with Maddie in a beautiful day in San Francisco, be out here in Hawaii as well, doing some stuff outside.

So I want to do more of those types of things. And so, yeah, if you want to send

things. And so, yeah, if you want to send a DM or something, just let me know and then we'll link up this stuff.

So, yeah, I got some more coming, so for sure. Thank you so much for becoming a

sure. Thank you so much for becoming a member as well. I really do appreciate that. This is really, really helpful.

that. This is really, really helpful.

Did you set up Mission Control yet? Mine

is wondering how to get it properly set up like Alex Finn did.

I have not set up a Mission Control and I'm going to experiment on the stream coming up where I'm going to try to look at Notion and I'm also going to try to see if I can build my own system.

The first part of it is trying to see if I can tap into Notion because Notion already has a lot of apps pre-built, right?

It has views for Kanban boards, it already has databases, it already has these types of things.

those things sync to my system and they can also be displayed through my lock screen as widgets.

So Notion seems like an interesting thing. I already have a lot of stuff in

thing. I already have a lot of stuff in Notion anyways.

So it'll just be about putting stuff in there that could be really interesting.

The other part of it is making an iOS app because I want to basically replace Siri.

So that's another type of thing that I'm going to try to get into, try to see if I can have a Siri app, basically, that is going to sit in my house.

and I'm also going to try to get this latest model from Qwen, text-to-speech, and see if I can get that loaded up on my DGX Spark.

So that'll be like my at-home Siri and it'll also be available for my friends.

So I want to get that working through the HomePods.

There is like a Siri kit. I really don't know all the specific details, but I mean, this was a while back when I was working at this company and I'm not sure what's public and what's not.

This is part of the reason why I don't do a lot of streams on iOS stuff because I don't really want to accidentally disclose things.

so I will like use everything publicly available to try to look at stuff and try to make have the agents write the code so that I'm not writing the code if you know what I mean so that's what's up what about is soothing is using Lucas files as well yeah I'm sitting is pretty cool I think the syncing part is what I want you know so when I write stuff in I want to show up

on my phone so that's kind of what's nice about telegram is like has such a low signal it works offline like when it comes back online with low signal you know telegram will hit the gateway and then send that off to the agent so So that's kind of the way I'm thinking about Notion here, that I don't have to have the app that's running all the time.

Whenever my phone connects back, Notion's already kind of a known thing. It's

battle-tested already.

So it's a pretty good gateway, and they already have good stuff for managing things overall.

So do you know how to set it up so the agents can talk to each other, have meetings?

That's probably what I'm going to get to in the next stream. So that's what's up.

But I appreciate everyone. I would love to be in SF more.

You give me the feeling of being close to it, at least much love from Jeremy.

Hey, thanks, man. I really appreciate that. That's really, really cool. Yeah,

that. That's really, really cool. Yeah,

that's really, really awesome. I

appreciate that. Yeah, I think whatever I could do is really cool, so we'll be there as well.

Do I have concerns about Telegram logging chats? I do and I don't, but I'm not

chats? I do and I don't, but I'm not super concerned because it's not anything super private, per se. And that's also something that if you are concerned, you can make your own app because all you have to do is just talk to the gateway directly via your own means.

so you can use iMessage if you want and you can kind of connect with that way as well.

So that's what's up.

All right, y'all.

I got to head out because the beach weather is so good.

All right, y'all.

Peace.