10:57Leveling Up with Eric Siu · Talking Head
How Loops Will Make You Way More Money
An 11-minute case for why manually prompting AI agents is already dead — and what building business loops looks like in practice.
June 12thThe argument in one line.
Stacking AI tools without a shared memory layer produces zero compounding — the missing piece is a single, queryable company brain that captures every call, doc, and decision so agents and humans can finally execute as one.
Read if. Skip if.
READ IF YOU ARE…
- You run or work at an agency, consultancy, or small company that has already adopted several AI tools but finds they do not talk to each other.
- You have sales call recordings, Slack threads, SOPs, or CRM data sitting in silos that never feed your AI workflows.
- You want a practical architecture — not just a concept — for building persistent AI memory into your operations.
- You are trying to understand how to give AI agents the right context without drowning them in irrelevant history.
SKIP IF…
- You are looking for a no-code or plug-and-play AI tool recommendation — this is an architecture framework, not a product review.
- You have no existing data assets (calls, docs, SOPs) to feed a memory layer; the value compounds from what you already have.
The full version, fast.
Most AI deployments fail to compound because the tools never share context. The fix is a company brain: a six-layer architecture that captures raw inputs (calls, Slack, CRM, SOPs), retrieves the right context at runtime, enforces a freshness-and-authority hierarchy for source of truth, controls access by workflow, turns human corrections into system rules, and finally deploys execution agents inside real workflows. Single Grain runs this on 500K tokens of persistent memory, 90+ daily crons, and 2,800 sales-call transcripts — one person now does the work that used to require many, and the system keeps getting smarter.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:43
01 · Why Most People Buy AI Backwards
Problem framing: copilots and agents with no shared memory produce no compounding. The missing piece is a single company brain.
01:43 – 02:47
02 · Inside Single Grain's Company Brain
Social proof: 2.3M X views/month, 500K tokens of persistent memory, 90+ daily crons, 2,800 Gong call transcripts turned into operational playbooks.
02:47 – 03:38
03 · The Execution Architecture in Layers
Five-layer overview: Capture, Retrieval, Source of Truth, Permissions, Feedback, Execution. Cake analogy — each layer is the foundation for the next.
03:38 – 04:10
04 · Layer 1: Capture Everything
Capture is the foundation most teams skip. Raw inputs: call transcripts, Slack threads, docs, briefs, wikis, CRM activity, SOPs, content decisions, agent outputs, daily logs.
04:10 – 04:30
05 · Why Compounding Wins
Warren Buffett statue visual — compounding is the eighth wonder, never interrupt it. Eventually you will need a dedicated full-time role just for the capture layer.
04:30 – 05:13
06 · Layer 2: Retrieval
The right context for the right task at runtime. Identify the task, retrieve and rank context, assemble context pack. Agent needs 6 relevant pieces, not the whole company history.
05:13 – 06:45
07 · Layer 3: Source of Truth
Freshness hierarchy: live greater than recent, recent greater than historical. Authority hierarchy: sales leader outranks product leader on sales questions. Time decay and trust score govern what the brain trusts.
06:45 – 07:14
08 · Layer 4: Permissioning
HR details, M&A details, financial context — not everyone should see everything. Wrong permissioning equals legal risk. Right brain for the right workflow.
07:14 – 07:55
09 · Layer 5: Feedback Loops
Corrections become rules. Every connection strengthens the system for everyone. Humans set the standard and will stay in the loop. Book rec: Thinking in Systems.
07:55 – 09:45
10 · Broken vs. Working: Layer by Layer
Side-by-side comparison for all 6 layers. Broken = scattered nodes, manual retrieval, stale docs, open access, corrections disappear, isolated AI drafts. Working = auto-capture, runtime chunks, freshness hierarchy, workflow-scoped access, corrections become rules, agents inside real workflows.
09:45 – 10:39
11 · Layer 6: Execution and CTA
Execution agents inside real workflows: integrated, accountable, measurable outcomes. Pitch for singlebrain.com. Closing: each layer depends on the one before — together they learn, protect, and compound.
Lines worth screenshotting.
- Most teams buy AI backwards — they stack agents and copilots, then wonder why nothing compounds; the missing piece is shared memory, not more tools.
- One sales call produces six downstream artifacts: objection library, sales training input, positioning signal, content idea, CRM risk flag, and future agent instructions.
- An agent does not need the entire history of the company — it needs six pieces of context that matter for the specific task in front of it.
- The freshness hierarchy for source of truth is: live data beats recent, recent beats historical — and trusted sources accumulate higher weight over time.
- Permissioning is product-level work: if everyone sees everything, you have a legal problem; the right brain must serve the right workflow.
- Feedback loops are what separate a system that learns from one that stagnates — every human correction must become a rule, not a one-off fix.
- Compounding is the eighth wonder of the world; a company brain compounds knowledge the same way Buffett compounds capital — never interrupt it.
- 500K tokens of persistent memory, 90+ daily crons, and 2,800 Gong call transcripts are the inputs that produced 2.3M X views per month with minimal headcount.
- The broken version of retrieval is a human manually finding context over several days; the working version is an agent pulling the right chunks at runtime.
- Execution agents without a memory layer produce isolated, one-off outputs with no ownership — agents with one produce integrated, accountable, measurable outcomes.
- If you are starting a company today, begin capturing call transcripts, Slack threads, CRM activity, and SOPs on day one — the earlier the compounding starts, the bigger the moat.
- Obsidian is recommended as a primary knowledge store for individual decisions; the principle is one canonical home for structured notes, not scattered files.
Six layers that make AI actually compound.
WHAT TO LEARN
Individual AI tools do not compound — a memory layer that connects them does, and building that layer is an engineering and organizational choice, not a tool purchase.
- Buying more AI agents without shared memory produces isolated outputs that disappear — the compounding only starts when everything feeds one persistent brain.
- Every sales call recording is undervalued if you treat it as just a call; structured ingestion turns it into an objection library, positioning signal, content source, and agent instructions simultaneously.
- The capture layer is where most organizations stop before they start — call transcripts, Slack threads, SOPs, and CRM activity must be systematically ingested, normalized, and indexed before any agent can use them.
- Retrieval is not search — it is assembling a curated context pack of six or fewer relevant pieces so the agent has exactly what it needs for the task at hand, not the whole company history.
- Source of truth requires a defined hierarchy: live data outranks recent, recent outranks historical, and more authoritative roles carry higher trust scores — without this hierarchy, agents will confidently cite six-year-old playbooks.
- Permissioning is a legal and operational concern, not an afterthought — HR data, M&A details, and financial context require workflow-scoped access or you create liability the moment an agent routes the wrong data to the wrong person.
- Feedback loops are what separate a learning system from a frozen one — every human correction must be captured as a rule update, not a one-off fix that disappears.
- The broken-vs-working comparison is the most useful diagnostic: if your capture is scattered, your retrieval is manual, your source of truth is stale, or your corrections keep disappearing, you are in the broken state regardless of which AI tools you have bought.
Terms worth knowing.
- Company brain
- A single unified intelligence layer — typically an LLM with persistent structured memory — that connects all of an organization's tools, call transcripts, SOPs, and data sources so agents and humans share the same context.
- Gong
- A sales call recording and intelligence platform; transcripts from Gong calls are the primary raw material fed into Single Grain's company brain.
- Cron / cron job
- A scheduled automated task that runs on a defined time interval; Single Grain runs 90+ daily crons to ingest and process new data into their company brain continuously.
- Context pack
- The curated set of relevant documents, facts, and instructions assembled and passed to an agent before it performs a task — the output of the retrieval layer.
- Permissioning
- The access-control layer that determines which team members or agents can read which data; wrong permissioning in a company brain creates legal and security risk.
- Eval
- Short for evaluation — a test or benchmark used to measure an AI system's output quality, used here to describe the feedback mechanism that lets the company brain improve over time.
- Single Brain
- The commercial product built on top of this architecture — a unified intelligence layer that connects Slack, Google, Meta, SEO data, and specialist agents via a chat interface.
Things they pointed at.
07:55bookThinking in Systems by Donella Meadows
Lines you could clip.
00:00
“Most people are buying AI backwards.”
04:57
“An agent doesn't need the entire history of the company. It just needs six pieces of context that matter for the task in front of it.”
10:07
“Each layer depends on the one before. Together, they create the company brain that learns, protects, and compounds.”
02:15
“A call is no longer just a call.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
00:00Most people are buying AI backwards. So they're buying copilots or buying agents or buying all this different tooling. And what's really the thing that matters the most is the ability to combine everything together.
00:12This concept of having a company brain, a single brain, that's missing. And in this video, what I'm gonna do is I'm gonna show you how you could go about doing this and how we went about doing this as well and how it's gonna change everything about how you go about doing things yourself or how you go about doing things with whatever organization that you're working with because I guarantee you, it's going to make you a lot more money.
00:31Let's get into it. So you can see over here, the whole idea here is is, look, you're disconnected before was that you had a lot of different files, you had a lot of different disparate tools. So you might have Slack.
00:40You might have Gong. You might have, let's say, your CRM. Right?
00:44You might have Salesforce or HubSpot. You might have your your different you might have different GitHub repos. You might have different analytics tools to plug in.
00:51So everything was was this situation before where things weren't talking to each other. But the idea that you have now is that you have this company brain where everything's connecting to it. Okay?
00:59So you have your different Slack channels, and then this brain is then feeding these agents that you have over here, which will then lead to the different output that you can you can get out, which I'll show you in a second. I literally, what our company brain is doing for us is that it's publishing x posts that are getting hundreds of thousands of views.
01:15And I've I think last last month, you know, for for us at least, it's it's for for my personal account, we did break a record in terms of getting about 2,300,000 views or so, which is great because our x account is continuing to to compound. Whereas, like, months and months ago, we were maybe getting maybe a 100,000 views or something like that.
01:30So that's going up quite a bit, and it's making it where one person can do a lot more. And that's because we have the ability to to have these company brain agents talk to our people, and that make thing makes things a lot easier. So here's what I'll call it here.
01:43So at Single Grain, which is my ad agency, we have we have this running across, like, 500,000 tokens of persistent memory. You have 90 plus daily crons, multiple specialized agents, and thousands of sales calls that are feeding the system. And so you basically have one sort one source that that the one source set includes 2,800 Gong call transcripts turned into operational playbooks.
02:03So this is how we do things in general because we've been doing these calls for such a long time. And and so you can see 15 calls produced, 390 insights, 470 facts, 125 frameworks. So this whole idea of of creating a searchable company, a queryable company, you can only do this if you have a company brain, a single brain.
02:19So what what matters is that, you know, you don't just look at a call as just a call. One one call artifact produces way more. It produces an objection library.
02:28You get sales training input. You get a positioning signal. You get a content idea source, CRM risk flag.
02:32You get future agent instructions. So you get way more output from your input. Okay?
02:36And that that that to me is very high leverage. And so for me, everything I think about doing is how do I get a four to one return on investment? A company brain, it should then convert knowledge into better execution.
02:46Alright? So if you look at the this this execution architecture in five layers, you can see that capture, that's at the very bottom. Right?
02:52So your calls, your CRM, your your SOPs overall. So you wanna make sure, again, that's the foundation. That's why people are talking about building this queryable company, and you you you want this stuff to be searchable.
03:02Right? So your your agents can find it. Then retrieval.
03:05So you want the right context at the right time. You want the source of truth, you know, what to trust. You wanna make sure that you're you're feeding good guardrails or you're providing good guardrails in terms of judgment for your agents, and then it'll it'll learn with you over time.
03:16And then permission, who can use what? So who gets what in your company. Okay?
03:20And then feedback, corrections, and rules, and then execution agents and workflows. Right? So that's all great.
03:25But you you gotta make sure that, again, you wanna think of this as a as a cake. So at the very bottom, again, you start with capture. Again, most people aren't aren't thinking about even the the the capture part.
03:33So if you're thinking about that right now, you're already way ahead of the game. Okay? So we've talked about the layers over here.
03:38If we click on this over here, you can see that what what actually gets captured. Like, what are we talking about? So call transcripts, Slack threads, docs, briefs, wikis, CRM activity, internal SOPs, content decisions, agent outputs, daily logs.
03:50If you're starting a company right now, you should start doing all of this. Just day one, you should be doing this. Okay?
03:56You wanna make sure you capture the raw material. You wanna normalize it, label it, right, store, and then index. And you you wanna make sure that you you continue this go you you continue, yeah, you continue to move the ball forward.
04:06Okay? So what ends up happening is as you continue to do this over time, it compounds with you. So we like to talk about how, you know, compounding is the the eighth wonder of the world.
04:13And you can see over here that that's actually that. That's Warren Buffett. Okay?
04:17That's a statue of Warren Buffett. Compounding, you never wanna interrupt. Compounding, you wanna keep building here.
04:21Right? And so that's what it is. You you wanna make sure that you're managing this over time.
04:24Eventually, you're gonna need a full time role just to managing just dedicated to managing the the capture portion. And then the next level is that you're moving to retrieval. Right?
04:32You So wanna identify the task, retrieve and rank context, assemble the context pack, and then go from there. So you wanna have the right context for the task.
04:39And so you can see what's happening here. By the way, all these artifacts that you're seeing from this post, this post was AI generated, and this this this artifact was AI generated. And this wouldn't have been possible without that company brain where initially I fed context to it, and then I fed context, and then someone else took it over from me and and went from there.
04:58Okay? So agent doesn't need the entire history of the company. It just needs six pieces of context that matter for the task in front of it.
05:04Okay? So, anyway, I would just say is that all I have to say is that, look, I'm not gonna read through all of this.
05:10I you you should definitely go to my ex, check out the article over here. And then, also, you wanna figure out, you know, what is the the source of truth. Okay?
05:17So for example, if maybe I'm pulling from a Gong call that is talking about sales objections that that that maybe it's from six years ago versus something that's maybe more recent, I'm probably gonna want the more recent thing just so you should think about there's a time decay of content as well. You should also consider that someone that has more authority in the space should be given a higher trust score.
05:39So, for example, if it's coming from the sales leader versus someone that is a, let's say, a product leader when it comes to the sales process, you'd probably trust the sales leader over the other person. So you would need to continue to train your company brain on what makes sense and what doesn't.
05:53And I would also even recommend that the fact that you're watching this video right now, take the transcript from this video. You can go to, like, one of these transcript websites, take it, and then figure out how you can make your own based on what I'm talking about right here. By the way, if you wanna grow faster, you need to have a single brain unified intelligence sitting inside of your chat.
06:10So could be inside of Slack or Teams, but you can see right here this bot's working and it's creating ad creatives. It is doing data pulls from Meta, from Google. You can pull your SEO data.
06:18So imagine having all these data connectors. You you can ask whatever question you want, get the data pulls a lot faster. Your team can see it as well.
06:24They can choose to execute, and then you can run your other specialist agents that you have inside. So we have ad creative agents. We have email infrastructure agents.
06:31There's only the agents that can do a bunch of things, and the whole idea is they're all playing together, they're playing with your team as well. That way you're gonna be able to just move a lot faster and then grow a lot faster. So check it out.
06:40Just go to singlebrainwithab.com. Singlebrain.com. We'll see you inside.
06:44Permissioning. So is it if we're talking about HR details, should this be available to everyone in your company? Probably not.
06:51If we're talking about m and a details, should everyone in your company be able to see it? Probably not. Client context, maybe everyone can see it.
06:57That's that's cool. Financial context, depends. Right?
07:00Depends on the client. So it's really make sure it's really important that you make sure to set up the right permissioning here. With the wrong permissioning, it can cause a ton of legal issues for your company.
07:10So you want to have the right brain for the right workflow. Alright? And then number number five, layer five over here is feedback loops.
07:17Okay? So you wanna make sure that if you're not letting your agent learn over time, you're not providing feedback, you're not having, you know, evals that it can learn from. Right?
07:25If you're not training it, it's not going to get better. And and like, you're gonna need someone that's that's on top of that. Maybe whoever's managing you you have different people managing these layers.
07:33You know, the operating rules here, you wanna capture every correction, structure the signal, update the right layers, improve continuously, and then humans set the standard ultimately. And you wanna continue that loop over time. I don't think humans are ever gonna be out of the loop here.
07:44You're always gonna want to make sure that over time maybe some loops need more checking over time, maybe some some not as much, but I I don't think that's gonna go away for a while. Right? So ultimately, you have six layers of this company brain here.
07:55So broken versus what work what's working. Okay? For So the broken version, you have nodes that are scattered everywhere under capture.
08:00Right? We're talking about getting the raw material in. For number two, for retrieval, you you have talking about finding the right context for the the right task for the task.
08:08Human finds the context slowly. So you you know this, obviously, is like, oh, I'm gonna get back to you in a couple days. Oh, that data pull?
08:13Yeah. It's gonna take about four days or so. Yeah.
08:15No. It's not gonna do that. Right?
08:17So this one, the agent pulls chunks at runtime. So you get it right now. And if we go back for number one, didn't talk about working version for number one.
08:23Calls, CRM, Slacks, SOPs, those are all automatically captured. You look at your granola calls. You look at your just you wanna make sure that, obviously, you wanna sanitize it and make sure that the sensitive things are are sensitive.
08:33And this is always growing over time. And then you can think about using Obsidian as kind of the primary. If you need another set of of notes, you wanna store all your notes in one area, Obsidian's a great place to kind of store all those those all the decisions that you've made over time.
08:45That's super helpful. And then number three, your source of truth. So deciding what what's true and what to trust.
08:50So AI, trust stale documents, old info, weak sources, outdated SLPs. I think this is where a lot of people go wrong, and I would say even recently, we have gone wrong. That was something we need to continue to watch over.
08:59But freshness and hierarchy decides what went, so you need to define what that hierarchy is. So live is greater than recent, greater than historical. Trusted sources rise over time, so just keep that in mind.
09:10Number four, permissioning. Control who can see what and when. If everyone sees everything, that's high risk.
09:15Too much access, hard to contain. In fact, this this graphic is so good that I need to share it with the AI product team on on my side.
09:24Access depends on workflow over here. So write context, write people, write boundaries. Number five, feedback.
09:29Turn human corrections into system learning. So corrections disappear, correction be becomes rules, and every connection strengthens the system for everyone. Because you wanna think about by the way, you wanna get better at systems thinking, read the book Thinking and Systems.
09:39Because everything that you're building right now is a system. It's not just a company brain. It's it's on top of that whatever plugs into it as well.
09:45And then number six, execution. You wanna deliver value inside the real work. So before, you had AI drafts, isolated tasks, one off outputs, no ownership, no con no impact.
09:53It's hard to scale that. Right? And then well, the working version is agentship inside real workflows.
09:58So integrated, accountable, measurable outcomes. You want predictability there. And again, like feedback, you don't wanna to just fix it one to once and lost forever.
10:07And that that's the problem. Right? I I think, especially with a lot of agencies, you have this issue over here where just a lot of this stuff is broken.
10:12If you can fix this, you're gonna be able to scale. So each layer depends on the one before. Together, they .
10:20That's what makes everything fantastic, and that is how you build a company brain. So hope you enjoyed this video.
10:27You can check out more videos on the channel around building this company brain or just talk AI business and marketing in general, and we'll see you on the other side.
The Hook
The bait, then the rug-pull.
The first three words land before any intro music, b-roll, or throat-clearing: most people are buying AI backwards. From there, the argument never lets up — agents that do not share context do not compound, and compounding is the entire point.
Named ideas worth stealing.
02:47model
The 6-Layer Company Brain Architecture
- Capture
- Retrieval
- Source of Truth
- Permissioning
- Feedback Loops
- Execution
A layered model for building persistent AI memory in an organization. Each layer depends on the one below — skip one and the whole stack breaks.
Steal forany agency, consultancy, or operator trying to make AI compound across their team instead of remaining one-off
02:15list
One Call Artifact Six Outputs
- Objection library
- Sales training input
- Positioning signal
- Content idea source
- CRM risk flag
- Future agent instructions
Every recorded sales call is raw material the company brain can extract into six downstream artifacts automatically.
Steal forjustifying the cost of call recording and transcription infrastructure
09:04model
Freshness Hierarchy
- Live greater than Recent greater than Historical
- Trusted sources rise over time
- Authority score by role
A tiered trust model for deciding which source wins when the company brain has conflicting information.
Steal forany RAG or retrieval system that ingests data from multiple time periods
How they asked for the click.
VERBAL ASK
06:42product
“Just go to singlebrain.com.”
Soft-woven into the product demo segment around 06:40 with a live Slack bot demo showing ad creative generation, then verbally called out at the end. Not a hard pitch — feels like a natural here is where we took this moment.
MENTIONED ON CAMERA
Visual structure at a glance.
