Transcription
Your AI agent probably doesn't have a
brain. And what I mean by that is it
doesn't have a system that allows it to
read and think through context that you
have developed over months and years and
reliably come back and be proactive
with. I published a whole guide on the
second brain last month. It was super
popular. A lot of people built it. A lot
of people improved on it. You can use
Zapier. You can use Notion. You can use
N8N. You can use an MCP server. You can
use Obsidian. I have all of those
pieces. But what I don't have is the
agent piece and that matters because in
the intervening period in the last few
weeks we are now at a point where agents
are becoming mainstream. Anthropic is
working on one. OpenAI hired Peter
Steinberger the inventor of Open Claw.
Open Claw itself passed 190,000 GitHub
stars and spawned over one and a half
million autonomous agents in just a
couple of weeks. We need a second brain
system that is agent readable. And so
what I'm going to lay out here today is
the architecture for what I am calling
an open brain. A databasebacked AI
accessible knowledge system that you own
outright with no SAS middlemen that can
break or repric or disappear. One brain
that every AI you use, Claude, Chat,
GPT, Cursor, whatever ships next month,
can plug into via MCP. You can type a
thought in Slack and five seconds later
it's embedded. It's classified. It's
searchable by meaning from any AI tool
you touch or any AI agent that wants to
touch it. The total cost and yes we've
benchmarked this. It's roughly 10 to 30
cents a month. I'm publishing a
companion guide on the Substack to
handle the step by step. This video is
about why the architecture of an agent
readable system matters much more than
the individual tools you choose and why
the memory problem we're talking about
here is secretly the bottleneck in
everything you're doing with AI today
and why people who solve it for agents
and themselves will have a compounding
advantage that whitens every single
week. So, first let's talk about the
memory problem that is hiding inside
your prompting. If you've been following
my videos for a while, you know I keep
coming back to one idea. The quality of
AI output depends entirely on the
quality of your ability to specify.
That's not a nice to have principle
anymore. That is the whole game. I laid
out the full framework I see for
prompting in 2026 in a video I did last
week. From prompt craft through context
engineering to intent engineering to
specification engineering, that
hierarchy is real. And the people who
are 10x more effective than their peers
have built context infrastructure that
does the heavy lifting on all of those
pieces, the context engineering, the
specification engineering before they
have to type a single prompt. And what I
want to talk about in this video is how
you take that abstract skill set and how
you turn it into a memory problem that
gives you a leg up on everybody else. In
other words, if you're going to do
context engineering, if you're going to
do specification engineering, seriously,
you need to invest in a memory system
that is yours, that is agent readable,
that makes calling and retrieving that
context, that makes specifying easier.
The best prompt in the world cannot
compensate for an AI that does not know
what you've been working on, what you've
already tried, what your constraints
are, who the key people in your life
are, or what you decided last Tuesday.
And by the way, that is also the
constraint working with agents. They
need that context, too. And right now,
that's exactly what most of us are
struggling with when it comes to AI.
Every single time we open a new chat, we
often start from zero. Every single time
we switch from claw to chat GPT to
cursor, we tend to lose things, which is
why we gravitate toward one of those
systems more than another. Think about
how much of your prompting is asking AI
to catch up on what you know already.
The background here is you're burning up
your best thinking on context transfer
instead of real work. A Harvard Business
Review study found that digital workers
toggle between applications nearly 1,200
times a day. I get tired saying that
sentence. Every switch seems really
small but collectively this is
devastating our attention. I have
watched this context switching issue
play out over and over and over again in
my own life in the lives of others and
what I keep coming back to is the
insight that our desire to specify to be
clear with AI is only getting higher and
it's demanding more of our memory
systems and our memory systems and
memory structures are not keeping up.
Memory architecture determines agent
capabilities much more than model
selection does. That's widely
misunderstood. And when you construct
memory incorrectly, you're stuck
reexplaining yourself forever or you're
stuck in a world where you know how to
access memory and the agent doesn't. I
believe we can make a stable memory
system that is reasonably futureproofed
that enables us to plug in new tools via
MCP server very efficiently. So we don't
have to keep updating our system. And
yes, I want to acknowledge something.
Claude has memory now. Chad GPT has
memory now. Grock has memory now. Google
has memory now. These features are
getting better all the time. But think
about what they give you and what they
don't.
Claude's memory doesn't know what you
told Chad GPT. Chad GPT's memory doesn't
follow you into cursor. Your phone app
doesn't share context with your coding
agent. Every platform has built a walled
garden of memory and none of them talk
to each other. There's a whole new
category of products emerging in early
2026 specifically because platforms
refuse to solve this products like
memcync
one context. The problem is real enough
to spawn an entire VC backed industry.
So what you've really got is multiple AI
tools getting upgraded all the time,
adding AI tools all the time to
experiment with them, and you have a
thin siloed layer of context that only
works inside each of those individual
tools. You know what? That's not really
memory. That is five separate piles of
sticky notes on five separate desks. And
now let's add autonomous agents into the
picture. The agent category has
absolutely detonated in the last few
weeks, but the use cases that are
shining, like the guy who got thousands
of dollars off a car purchase, they're
shining because the agent has the
ability to securely and safely access
relevant memories, relevant context from
the user. Whereas agents that just guess
contacts or have to fill in the dots
because you aren't able to provide them
secure access to all of your systems,
they're not going to be nearly as useful
for you. And whether we're talking about
agents or we're talking about tools, the
part that should bother you even more is
that these systems that corporations are
designing are all designed to create
lock in. Memory is supposed to be a lock
in on chat GPT, ditto on other systems.
So you've spent a long time building up
history with a tool and now if you want
to try the latest other model, let's say
you're on chat GPT and you want to try
Gemini or you want to try Claude or you
want to try another model, you lose all
of that context, not because the new
model is worse, but because your context
is trapped in the old one and oh by the
way, all of that memory in those
individual tools, that is not agent
readable. And so as we get to a world
where autonomous agents are becoming
more and more and more a thing, the big
corporations are betting that if they
can trap you with memory, you will only
use their agents and they will get to
keep you and your attention and your
dollars forever. But your knowledge
should not be a hostage to any single
platform. And for most of us right now,
frankly, it is. And that's shaping our
entire AI future. We don't necessarily
have a free choice between tools right
now because the product strategy of
these large businesses is to keep you to
keep you engaged to keep you
entertained. I've talked about how in
many cases you're pushing for engagement
with these models. One of the reasons
why chat GPT40 was so mourned and so
grieved was because it was an engagement
optimized model and people liked the
engagement. It works. Ditto with memory.
Memory is engaging. Feeling known is
engaging. It works. It's smart product
strategy. But you're smart, too, and you
don't have to go along with that product
strategy. And you might be thinking at
this point, Nate, you made a video on
second brain. I can just connect it to
my open claw and I'm fine. Absolutely,
you can try that. But you're going to
run into a structural mismatch that most
people haven't noticed. That explains
why the current generation of notetaking
tools needs a different more structural
memory layer underneath. The internet
right now is forking. I've talked about
that. There's the human web with fonts,
with layouts, with what you're reading.
And there's the agent web that's
emerging with APIs, with structured data
that's built for machine to- machine
readability. That fork is happening to
your memory architectures and your notes
as well. Your notion workspace, for
example, is built for human eyes. It's
built for pages, for databases, for
views, for toggles, for cover images.
It's beautiful for you. It's useless for
an AI agent that needs to search by
meaning, not by folder structure. Your
Apple notes are locked into an
ecosystem. Your Evernote has a decade of
accumulated clutter with no semantic
structure. Your bookmarks are a
graveyard of things you've meant to
read. These tools were built for the
human web back in the 2010s. They were
designed for you to browse, to organize,
to read. They were never designed
fundamentally with the expectation that
AI agents would query them. That got
bolted on later, much more recently. And
the apps adding AI features today are
mostly doing it as bolt-ons, like chat
with your notes. Great. You have one AI
that can kind of search one app. What
about the other five tools you use every
week? We're still in a world of separate
sticky notes on separate desks. You've
traded one silo for another. Every
second brain app has been reaching for
something that required a different
layer entirely. Infrastructure built for
the agent web, not the human web. And
that's what I want to focus on here.
Because if you can build infrastructure
for the agent web, you are suddenly in a
position to make a lot more
human-friendly decisions with how you
plug into that infrastructure. The
infrastructure is yours. It's something
your agent can plug into. It's something
your chat bots can plug into, but you
control and manage it. This frees you
from having memory that only lives with
one of these corporations and their
clouds AI systems. You don't have to
depend on chat GPT memory anymore. It
also frees you from having to depend on
an individual SAS company not changing a
setting in order to keep your own second
brain working. And ultimately, as agents
get better, it frees you from having to
do as much manual work to retrain a
second brain. And so, this is me
essentially giving you a sense of how
agents unlocking are changing our
perspective on memory and changing our
perspective on prompting and changing
what we need to be digital citizens.
Just as we needed a personal computer to
be digital citizens over the 2010s, over
the 1990s, over the 2000s, we need our
own memory architectures to be
responsible AI citizens now. But we
haven't really had a way to do that. And
until very recently, until the last few
weeks, we haven't had AI agents that
would make that really practical. Now we
do, and now the world has moved, and now
it's time to talk about it. So, let's
get specific. What am I proposing here?
Instead of storing your thoughts in an
app designed for humans, you should
store them in infrastructure designed
for anything. A real database, vector
embeddings that capture meaning, not
just keywords, a standard protocol that
any AI can speak. I'm calling it open
brain because the architecture is what
matters and you should not be forced to
choose any given model. This is all
possible because of MCP, the protocol
shift that I talked about briefly above.
It started as Anthropic's open- source
experiment in November of 2024, but it's
since become the HTTP infrastructure of
the AI age. It's the USBC of AI. It's
one protocol. Every AI, your data is
yours. It stays in one place, but every
tool that speaks MCP can read it. So, at
a high level, I don't want to make you
go and click somewhere. Let me show you
what this actually looks like.
Your thoughts live in a Postgress
database you control, not somebody
else's proprietary format. This is the
most boring battle tested technology you
can imagine. Postgress is not exciting.
It's not deprecating. Postgress isn't
chasing a growth metric. Postgress isn't
VC backed and needing to hit a billion
dollar unicorn valuation. It's just a
standard way of storing data. And you
want that boringness because everything
else needs to plug into it. The nice
thing about the database is that if you
construct it properly, if you vectorize
it, every thought you capture gets
converted into a vector embedding, which
means it's a mathematical representation
of what it means that is immediately
natively AI readable. So when you ask
what was I thinking about career changes
last month, it can find your note about
how you were considering moving into
consulting or how you were considering
moving into product even if you never
used the word career in the original
thought. is called semantic search and
it's a whole different universe from F.
So what this looks like when you have
Postgress hooked up with an MCP server
is you can type into a Slack channel,
hey I was talking with Sarah. She
mentioned she's thinking about leaving
her job to start a consulting business.
She's been really unhappy since the
reorg. 5 seconds later, the system has
stored the raw text, generated a vector
embedding of the meaning, extracted the
metadata, the people, the topics, the
type, the action items, and filed all of
it in a real database. Now, any AI that
you're working with can go see that. If
you're in Claude working on a coaching
framework, hey, search my brain for
notes about people considering career
transition. Found it. If I'm in chat GPT
drafting an email, same search, same
result. If I'm in cursor building a tool
and I need to remember a decision I made
last week, hit the MCP server, it's
right there. One brain, every AI
persistent memory that never starts from
zero. Even if you start a new tool
tomorrow and you've never touched it
before. So this has two basic parts,
right? Capture runs through any tool you
have open. You type a thought, it hits a
superbase edge function that generates
an embedding and it extracts the
metadata in parallel and stores both in
a Postgress database with PG vector and
it just replies in thread with a
confirmation showing what it captured.
The whole round trip takes under 10
seconds. Retrieval runs through an MCP
server that connects to any compatible
AI client. You have three tools.
Semantic search, which is finding your
thoughts by meaning, listing recent,
which is browsing what you captured this
week. and stats. See your patterns,
right? You can hit this from Claude,
from Claude Code, from Chad GPT, from
cursor, from VS Code, from anywhere you
can query your brain through an MCP
server. If all of this sounds like Greek
to you, the companion guide walks you
through a complete setup. Copy paste, no
coding, about 45 minutes to set up. And
you know how I tested this? I asked
someone in my life to follow this guide
before I showed it to you. And she has
no coding experience whatsoever. And I
said, "Can you get to a point where you
can set this up?" And she could. And it
took her about 45 minutes. And I'm not
kidding about the cost because the total
running cost on the free tiers of say
Slack and Superbase, which is what I'm
talking about here, it's roughly a dime
to 30 cents a month and API calls for
about 20 thoughts a day. So you're going
to spend more on coffee this morning
than you're going to spend on the system
this month. Here's why getting memory at
the fundamental architectural level
matters beyond the nice feeling we get
from building a cool tool. I love to
build. You can probably tell people who
love to build will love to build anyway,
but it matters for everybody. It doesn't
just matter for those of us that like to
experiment. We are in the middle of a
massive shift in how AI integrates into
our daily work. The models keep getting
better at a terrifyingly fast pace and
you don't want to fall behind. Opus 4.6
6 shipped just a couple of weeks back.
The agent market is growing probably in
triple figures this year. Threeperson
engineering teams are routinely
outproducing teams 10 times their size.
And we're finally seeing this explosion
in AI productivity show up even in
economywide metrics. Eric Bjornson wrote
in the Financial Times last month that
US productivity grew roughly 2.7% in
2025, which is double the decade
average. And frankly, Eric attributed a
fair bit of that to AI agents and AI.
But the key is, as I've called out
before, AI adoption is not the same
everywhere. If you're just talking with
a single chatbot, I've said it over and
over, you're not really adopting and
working your workflows around AI in the
way you need to. And the people getting
those outsized results are not depending
on better models to get there. They're
actually restructuring how they work
with AI as a primary collaborator. But
you cannot collaborate with something
that has no memory of you. Think about
the difference between these two
workflows. Person A opens up Claude,
spends four minutes explaining their
role, their project, their constraints,
and the decision they're trying to make,
and they get a good answer. Person B
opens up Claude. It already knows her
role, her active projects, her
constraints, her team members, and the
decisions she made last week because all
of that lives via MCP server in Open
Brain.
All of it is loaded up before she types
a word. She asks for a question, she
gets an answer informed by six months of
accumulated context. If she wants to
switch to Chad GPT for a different
perspective, she'll get a different
model, but she'll get the same brain,
the same context, and the same answer
quality. Every single tool will have the
full picture for her. And the key is
that advantage will keep compounding.
Every thought person B captures makes
the next iteration better. Every
decision logged, every person noted,
every insight saved as another node to
what's a growing knowledge graph that
every AI in the system can access. So
person A is going to start from zero
every single time. The gap between I use
AI sometimes and AI is embedded in how I
think and work is the career gap of this
decade. And it comes down to memory and
context infrastructure. And the gap is
going to get wider as person B continues
to accumulate knowledge every week. The
people who build persistent, searchable,
AI accessible knowledge systems will
have AI that gets better at helping them
over time because it has more context to
work with. Every thought you capture
makes the next search smarter, the next
connection more likely to surface. And
that is a compounding advantage that you
own, that the big companies don't own.
Whereas the people who keep reexplaining
themselves in every chat window are
going to wonder why AI still feels like
a party trick. It's the same tech. It's
just wildly different outcomes. And the
variable here is your infrastructure.
And one thing I want to call out here,
I've given you a simple example where
you can retrieve a clear answer in text
in any AI tool you want with an MCP
server. But MCP servers are not just for
retrieval. And if you construct an open
brain, your MCP server can work in a lot
of different directions to give you
advantages you might not think of if you
are just used to using memory in a
single tool. MCP means you can write
directly into the brain from anywhere. I
really meant that. You can write into
Claude on the phone. You can use Chad
CPT on the desktop. You can use Claude
code in the terminal. You can rig it up
uh to talk to a messaging app. any MCP
compatible client becomes both a capture
point and a search tool. You're not
locked into Slack or any other system.
That's what open means. And then think
about what you can build over the top.
It's easy to use MCP to build a
dashboard that visualizes your thinking
patterns over time, a daily digest that
surfaces forgotten ideas based on what
you're working on. And do you know that
you don't need to use code to do that
because you can just ask the AI tool of
your choice to retrieve from the MCP
server the relevant slice of context and
build something because the data is
stored in a way that is easy to plug in
and easy to store and easy to access
from any tool out there. The ceiling is
wherever you decide to stop building.
Now I want to be honest the metadata
extraction isn't always perfect. The LLM
makes its best guess to classify with
limited context and it will sometimes
mclassify a thought or miss a name. It
doesn't matter as much with semantic
embedding because the embeddings handle
so much of the heavy lifting with
retrieval. Semantic search works even
when the metadata is off. The one real
requirement for this to work is that you
actually use it because the system
compounds. Every thought you capture
makes the next search smarter and the
next connection more likely to surface.
But it needs input. You need to build
the habit. You need to be dumping your
thinking into the system and let it do
the rest. Now, if you're a subscriber on
the Substack, I've put together four
prompts that cover the full life cycle.
And I actually want to describe them in
the video because even if you're not a
subscriber, you should understand how we
can use prompts in the architecture of
this system to think more deliberately
and make the memory architecture fit our
needs. The memory migration is the first
thing I'm going to suggest. You want to
run this right after setup. It extracts
everything your AI knows about you
already from Claude's memory, from Chad
GPT's memory, from wherever you've
accumulated context, and it saves it
into your open brain. Every other AI you
connect then starts with that foundation
instead of zero. So you want to run it
once and let it pull that stuff down.
I'm also building what I call the open
brain spark because I sometimes get
writer's block. So you want to have an
interview prompt that discovers how the
system fits your specific works. It asks
about your tools, your decisions, your
reexlanation patterns, your key people,
and then generates a personalized list
organized by category that suggests what
you should be putting into Open Brain
regularly. Use it when you're staring at
the Slack channel or you're staring at
your messaging app or you're staring at
Shed GBT and you're wondering what do I
type that I want to put into OpenBrain
today. I also put together quick capture
templates. So these are five sentence
long starters optimized for really clean
metadata extraction. So a decision
capture prompt, a person note, an
insight capture, uh a meeting debrief,
each one is designed to trigger the
right classification in your processing
pipeline. And after a week of capturing,
you'll find you don't need them as much
because you're going to develop your own
patterns. but they're really useful for
building that habit early without having
to think about how to sort of send the
system a coherent message where it's
likely to classify correctly.
The weekly review is another one I put
together. End of week synthesis across
everything you captured. It clusters by
topic. It scans for unresolved action
items. It detects patterns across days.
It finds connections you missed. And it
identifies gaps in what you're tracking.
So about 5 minutes on a Friday afternoon
becomes more valuable every week because
your open brain continues to grow.
If we zoom back out, when this thing
works, when you get the Postgress
database set up, you're starting to use
it in whatever messaging app you want,
you're starting to see the memory become
consistent across all your AI tools, and
you're starting to realize you do not
depend on proprietary paid for memory by
big AI companies.
something happens that's a little bit
hard to describe until you experience
it. Your AI in every single part of the
system, whether you're using Claude or
Chad GPT or both or Cursor or Grock,
whatever it is, it starts to know you.
Not in the creepy corporate surveillance
way, in the hey, we were thinking about
this last week and it's relevant to what
you're asking me now kind of way. The
way a great colleague remembers what
matters. So every AI you use gets
better. You're less afraid of trying a
new AI because you can just plug it into
MCP and it finally has the context.
This is what an agent readable world
makes possible. And I want to call out
something really special here. When I
suggested the original second brain
guide, I built it before the agent
revolution went mainstream, which again
was only about a month and a half ago, a
month ago.
And it was useful for humans and it was
designed to solve a fundamental
cognitive problem that we've had which
is that we have trouble holding stuff in
our head and we need to see patterns
over time. LLMs can help us assess
patterns. That's all still true and you
can use this open brain in that way. But
when the agent revolution came through
in the last few weeks because again AI
is moving that fast. What we need to
move to is a second brain system that is
more foundational. Something that
enables both us and our agents to
reliably read from a system that isn't
SAS controlled, that isn't proprietary
company controlled, that is frankly
open- source LLM friendly. And when we
have that, we get two benefits. Yes, the
agent can read it. And that is in line
with where we're going with agents and
how quickly agents are going mainstream.
And that's the reason I'm making this
video. But second, look at how much
cleaner and clearer the human readable
part of this gets. We get downstream
benefits that we did not get when we
think about the system from only a human
readable perspective. Because if you
think about the system from a human
readable perspective, you get something
like what I described. You focus on
SASfriendly solutions with graphical
user interfaces that humans can easily
read because you want to make it easy
and accessible to build the system. And
that's what I did originally. But if
you're willing to get slightly technical
and follow a clean step-by-step tutorial
to get to something that is a true
database, what you get is a
futureproofed system that unlocks the
human benefit of touching any AI system
in the future that you may want to try
without doing any additional effort. And
so we humans reap a tremendous amount of
value from the clarity that comes from a
truly foundational architected memory
system. This reminds me of one of the
larger lessons I've been meditating on
in the AI revolution which is that AI is
forcing a clarity of thought in our work
in our lives that has a tremendous
amount of human benefit. Toby look has
said that he thinks a lot of corporate
politics amount to bad human context
engineering which is a very provocative
take and I think that that is something
that pops out here because we need
extraordinary clarity to work with AI
agents and when we develop that
extraordinary clarity through memory
architectures that are foundational
through good databases through a clean
MCP server We get the benefit of cleanly
and clearly being able to plug in and
work with that memory system anywhere.
We do good context engineering for our
human brains when we build the right
context engineering for AI, which is
kind of Toby's point about politics.
When we do good context engineering for
agents, we happen to do good context
engineering for people. And that makes
people less likely to play politics. So
the second brain you built, if you were
one of the thousands of people that
built it when I talked about it, was
always reaching for this. It was
reaching for a place where your thinking
lives, where it's searchable by meaning,
where it's accessible to any tool you
use. And those tools solve the capture
problem. They solve the organization
problem. But what they didn't realize
they needed to solve because it wasn't
really there yet was the agent readable
problem.
Open brain adds that foundational layer
not by replacing what you built but by
giving it an infrastructure underneath a
database, a protocol, your thoughts,
every AI you'll ever use. So you can
build it in a morning over coffee this
weekend. Yes, really you. And your
future AI, your future self as a human
will thank you for every thought you
start to capture. Now, if you have
already built a second brain, I'm also
including a special migration guide so
that you can figure out how to not lose
the thoughts you've been capturing and
make sure you get them into a system
that is more agent readable going
forward. Best of luck. Don't be afraid
of how this is slightly technical. There
have been lots of visuals all the way
through this YouTube helping you to see
what I mean. And you'll see more guides
in the substack if you're interested.
And honestly, I put enough visuals into
this video that if you are not ready to
hop into the Substack, totally fine. You
should still be able to get there. You
should be able to show this video to an
AI and say, "Help me build this." And it
should be able to do it.
Cheers.