Morning Walk With Murty

00:00:00 Speaker 1: I want you to take a second and just think about your relationship with the AI tools you fire up every morning.

00:00:06 Speaker 2: Yeah, the ones we practically live in now.

00:00:08 Speaker 1: Right, exactly. And if you use them enough, you've probably noticed those little, like, memory features popping up lately. Oh for sure. Like the AI starts remembering your tone or you know, it remembers your job title. Maybe it remembers that you have this deep, unyielding hatred for corporate jargon and a strong preference for bullet points.

00:00:27 Speaker 2: Which is very relatable, honestly.

00:00:29 Speaker 1: Yeah, but what if the AI, um, didn't care about your personal preferences at all? Right. What if instead of trying to be your buddy, the AI just relentlessly studied its own workflow, like cataloging every single mistake it made while trying to do your job? Yeah. What if it did all the studying while you were sleeping?

00:00:47 Speaker 2: It it completely flips the script. On what the industry has told us AI memory is actually supposed to achieve, right? Up until now, we've been, you know, conditioned to expect personalization from these tools. We haven't really been given the tools to expect, like structural systemic improvement in how the machine actually executes complex tasks over time.

00:01:08 Speaker 1: Okay, let's unpack this because we have a really fascinating deep dive today. We're looking at a recent Mach Tech post breakdown of perplexities newly launched system, which is called brain.

00:01:19 Speaker 2: Yeah, brain.

00:01:20 Speaker 1: And this is an architecture designed specifically for their AI agent, which they very literally named computer.

00:01:27 Speaker 2: A very straightforward.

00:01:27 Speaker 1: Name, super straightforward. And our mission today is to figure out how perplexity is kind of reframing the entire concept of AI memory, moving the goal post from basic user engagement to hardcore agent performance.

00:01:40 Speaker 2: Exactly.

00:01:41 Speaker 1: But before we get into the actual gears and wires of how brain works, I do want to challenge this premise just a bit.

00:01:47 Speaker 2: Okay, go for it.

00:01:47 Speaker 1: Because I actually like that my AI remembers my preferences. You know, it saves me a ton of time when I don't have to prompt it to write in my specific voice. Yeah. Why is that paradigm something we need to, um, move away from?

00:02:01 Speaker 2: Well, it's not necessarily that we need to abandon it entirely, but we definitely need to recognize its limitations.

00:02:07 Speaker 1: Okay.

00:02:07 Speaker 2: Think about traditional AI memory along two axes. The first axis is, uh, what is memory about? And the second axis is what is it for?

00:02:17 Speaker 1: Okay. So about and for.

00:02:18 Speaker 2: Right. Right now, almost all AI memory is about you the user. Yeah. It stores your working style, your tastes, uh, your typical prop structure.

00:02:29 Speaker 1: Oh, yeah. My bullet points.

00:02:30 Speaker 2: Exactly. And the purpose, the what is it for is engagement. It basically produces a profile of you to reduce friction, making you feel more comfortable. So you just keep coming back to the platform, right? It's essentially, you know, a parlor trick designed for consumer retention.

00:02:46 Speaker 1: So it's kind of like a highly attentive, maybe slightly sycophantic human assistant.

00:02:50 Speaker 2: That's a good way to look at it.

00:02:51 Speaker 1: Like the kind who meticulously memorizes exactly how you like your coffee. Yeah. Two pumps of vanilla oat milk. Exactly one hundred and sixty degrees. And they have it waiting on your desk every single morning.

00:03:02 Speaker 2: Which feels amazing.

00:03:03 Speaker 1: It feels great. But then you asked that same assistant to file the weekly regional sales reports, and they completely botched the filing system.

00:03:11 Speaker 2: They put the Q3 folders in the Q2 drawer.

00:03:13 Speaker 1: They mislabel all the spreadsheets, and you have to sit down and teach them the entire corporate directory all over again.

00:03:18 Speaker 2: And the kicker is you have to do that every single week, right? So what's fascinating here is that perplexity is essentially saying, forget the coffee.

00:03:26 Speaker 1: Get the coffee, I like that.

00:03:28 Speaker 2: Yeah. Brain operates on an entirely different set of axes. Its memory is not about you at all. It's about the agents work, okay? And its purpose isn't user engagement. It's recursive utility.

00:03:41 Speaker 1: Recursive utility.

00:03:42 Speaker 2: Yeah. So brain tracks what the AI agent did to solve a problem. What specific steps succeeded, which ones resulted in errors? And this is the most important part. The manual corrections you had to step in and make.

00:03:54 Speaker 1: Oh wow. So it's an assistant who, um, might hand you a lukewarm black coffee, but they spend hours meticulously studying your company's actual filing system. They take detailed notes when they mess up a spreadsheet formula, they cross-reference the directory, and they basically build an internal rule book. So they never, ever make that mistake again.

00:04:15 Speaker 2: And that internal rule book is what perplexity calls a traceable context graph.

00:04:20 Speaker 1: Okay.

00:04:21 Speaker 2: The ultimate goal isn't a psychological profile of the user. It's a living structural map of the work itself, designed purely to enhance performance.

00:04:31 Speaker 1: A traceable context graph. I mean, that sounds incredibly dense.

00:04:35 Speaker 2: A bit of a mouthful.

00:04:36 Speaker 1: It's like pure computer science word salad. I'm going to need you to break that down for us. Sure. Absolutely. How does it actually store and process all that information about my workflow without just like, drowning in a massive, unusable text file of its own history?

00:04:51 Speaker 2: Let's look at the mechanics of it. Brain builds this context graph in the form of what they call an LLM wiki, a large language model wiki.

00:04:58 Speaker 1: Okay, an LLM wiki right now.

00:05:01 Speaker 2: This isn't a website that you or I click through on a browser. It's an internal knowledge base that's loaded directly into the agents sandbox. Got it. Imagine a network of nodes. One node is a specific project you're working on. Another node is a key client. Another node is an internal database you use all the time. The wiki maps out how all these ideas, people, and projects actually interact in your specific daily workflow.

00:05:29 Speaker 1: Wait. Hold on. How is the AI getting access to all these different parts of my workflow in the first place?

00:05:34 Speaker 2: Ah yeah.

00:05:34 Speaker 1: Because I'm assuming it's not just passively reading the text I type into a chat box.

00:05:38 Speaker 2: No, not at all. You're touching on a really crucial piece of the architecture here, which is connectors.

00:05:43 Speaker 1: Connectors.

00:05:44 Speaker 2: Yeah. In an enterprise environment, an AI agent isn't just a simple chat bot. It's plugged directly into your company's infrastructure using APIs.

00:05:52 Speaker 1: Okay. The bridges.

00:05:53 Speaker 2: Exactly. APIs act as bridges between different software. These are the connectors. So the agent has a connector to your Slack channels, the connector to your Google Drive, a connector to your Jira tickets. Wow. It's pulling in vast amounts of raw data from across your entire organization.

00:06:08 Speaker 1: Okay, so the agent has all these connectors pulling in data and it's building this LM wiki. But the article specifically emphasizes the word traceable, a traceable context graph. Right. Why is traceability the defining feature here? Like why does it matter so much if the graph is traceable or not?

00:06:25 Speaker 2: Because an AI without traceability is essentially a black box, and a black box destroys trust. Right? If an AI agent gives you a confidently wrong answer, say it pulls up a financial projection for Q4 that is just wildly off.

00:06:40 Speaker 1: Which happens all the time.

00:06:41 Speaker 2: All the time. And if you have no idea where it got those numbers, you can't fix the problem. You just end up throwing your hands up and you stop using the tool.

00:06:48 Speaker 1: Yeah, you go back to doing it manually.

00:06:49 Speaker 2: Exactly. But with a traceable context graph. Every single memory entry in that LM wiki is hard linked back to its exact source.

00:06:59 Speaker 1: Oh I see. So if it gives me a weird financial projection, I can look at the output and the AI will basically say, hey, I generated this number based on a spreadsheet found in the Q2 Google Drive folder, cross-referenced with a Slack message from the CFO sent last Tuesday.

00:07:14 Speaker 2: Perfect example. And then you can immediately spot the error. You realize, ah, the CFO actually sent an updated projection on Wednesday and the AI just missed it, right? So you correct the AI pointing it to the new source. And because the graph is traceable, the AI doesn't just like memorize the new number. It updates the actual pathway in its wiki. Wow. It notes that okay, Wednesday Slack messages supersede Tuesday Google Drive files for this specific metric.

00:07:42 Speaker 1: It updates the pathway that. That actually brings us to a part of this architecture that genuinely confuses me a bit.

00:07:47 Speaker 2: Okay.

00:07:48 Speaker 1: So perplexity states that brain updates this wiki incrementally, synthesizing all these sessions and connector results in manual corrections overnight. Yes, overnight. Here is where I have to push back. We live in an era of like instant compute, instant gratification.

00:08:05 Speaker 2: We do.

00:08:06 Speaker 1: If I correct the AI at ten zero zero a m, why isn't it updating its massive internal wiki in real time right after the task? Why on earth is it waiting until I am asleep?

00:08:15 Speaker 2: It's a great question. It really comes down to the fundamental limits of how large language models process information, specifically regarding tokens.

00:08:24 Speaker 1: Tokens, right. The currency of AI.

00:08:25 Speaker 2: Exactly. Think of tokens as the foundational pieces of words or data that an AI reads and generates every single time an AI processes tokens, it requires computational power. Okay? If you force the model to continuously restructure its entire understanding of your complex corporate world in real time, after every single little prompt.

00:08:48 Speaker 1: Oh, I.

00:08:49 Speaker 2: See you're going to burn through an astronomical amount of compute. It would be incredibly expensive, and it would slow the agent down to a crawl while you're actually trying to get work done.

00:08:59 Speaker 1: So it's kind of like trying to rebuild the engine of a car while you're actively driving it down the highway.

00:09:03 Speaker 2: Yes. It's just not feasible.

00:09:04 Speaker 1: Wow.

00:09:05 Speaker 2: Okay. And beyond just the compute costs, the overnight delay is actually necessary for the quality of the learning.

00:09:11 Speaker 1: Really?

00:09:11 Speaker 2: Yeah. The overnight process is what data scientists call the synthesize step. Imagine you're deep into a complex research project yourself throughout the workday. You hit dead ends. You try different database queries. You write all these messy, disjointed notes.

00:09:25 Speaker 1: Write a total chaotic mess on my desk.

00:09:27 Speaker 2: Exactly. If you try to write the final polished executive summary while you were still in the middle of hacking through that messy research, the summary would be a disaster. You need distance.

00:09:38 Speaker 1: You need to, like, step back, look at the entire messy day, figure out what actually worked, and compress it into something useful.

00:09:47 Speaker 2: Precisely. Overnight, the brain system has the computational breathing room to look at the entire conceptual loop of your day. It looks at the tasks you assigned, the dead ends it hit via those connectors and the manual corrections you made. Then it synthesizes all of that raw, messy interaction into clean, organized, hyper efficient updates for the MLM wiki.

00:10:09 Speaker 1: That makes a lot of sense.

00:10:10 Speaker 2: So when you log in the next day, the agent has a much stronger, clearer signal of how to do its job. Perplexity frames this brilliantly. Actually, there's a current token. Usage is an investment in more efficient token usage later.

00:10:22 Speaker 1: Okay, I really want to focus on that specific phrase, because if this system is successfully synthesizing all that messy data while we sleep, what does that actually look like for the bottom line and investment in more efficient token usage later? How does an AI simply remembering its mistakes actually save a company money?

00:10:40 Speaker 2: To understand the financial impact, we have to look at how enterprise AI operates at scale. The Mach Tech post piece highlights some early first party testing numbers from perplexity.

00:10:50 Speaker 1: Okay, what are the numbers looking like?

00:10:52 Speaker 2: On tasks that the computer agent has seen before, they're reporting a twenty five percent increase in answer correctness.

00:10:58 Speaker 1: Wow.

00:10:58 Speaker 2: twenty five percent right? They are also seeing a sixteen percent increase in recall, meaning the AI's ability to pull the exact right piece of information from a massive data set.

00:11:09 Speaker 1: That's huge.

00:11:09 Speaker 2: It is. But the metric that will really make CFOs sit up and take notice is a thirteen percent drop in cost on tasks requiring historical context.

00:11:18 Speaker 1: Okay, a thirteen percent cost drop. Let me try to visualize how that happens just from memory, because an AI obviously doesn't get paid an hourly wage.

00:11:25 Speaker 2: No, it doesn't.

00:11:26 Speaker 1: But it does cost tokens. So think about asking an AI to solve a complex, multi-step problem across your company's network without memory. It's like dropping a hiker into a dense, overgrown forest and telling them, hey, go find a hidden lake.

00:11:42 Speaker 2: Okay, I like this.

00:11:42 Speaker 1: The hiker has a machete. They hack through the brush, they take a wrong turn at a river. They backtrack. They hack through more brush, and finally, exhausted, they find the lake, right? In the AI world, every swing of that machete is a model call. It's token usage. It costs server time and it costs money.

00:12:01 Speaker 2: And in a traditional setup, what happens when you ask the AI to find that lake again the next week?

00:12:05 Speaker 1: Right. It drops the hiker back at the starting line. The brush has somehow magically grown back, and it has to hack its way through all over again.

00:12:12 Speaker 2: Exactly. But with brain, that overnight synthesis completely changes the landscape.

00:12:17 Speaker 1: How so?

00:12:18 Speaker 2: While you're sleeping, the AI doesn't just remember where the lake is. It builds a highly detailed topographical map with GPS coordinates. That's the traceability aspect. It marks the dead ends with massive red stop signs. So the next day, when you ask it to run that route, it doesn't even bring the machete. It just follows its own perfectly optimized map straight to the destination.

00:12:40 Speaker 1: Oh wow. So drastically fewer model calls, drastically fewer tokens burned, resulting in that thirteen percent cost reduction.

00:12:48 Speaker 2: Yes. If we connect this to the bigger picture, you start to see why this is so critical for enterprise scaling. Without a work focused memory, an agent relearning the same context from scratch every single time creates massive computational waste.

00:13:02 Speaker 1: Yeah, it's just burning money.

00:13:03 Speaker 2: Exactly. And the article outlines three very distinct concrete use cases where this topography map radically changes the workflow.

00:13:11 Speaker 1: Let's dig into those because I really want to see how this map works in practice. The first use case they highlight is data scientists running routine weekly pipeline audits.

00:13:20 Speaker 2: Right? A pipeline audit can be incredibly messy. You are querying multiple databases, checking for anomalies all over the place. Sounds tedious. It is. And without brain, the AI might stumble into a deprecated database. Or it might try to use a query syntax it found in a document from twenty twenty one that you know just no longer works. It swings the machete at the wrong trees, right? But with brain, it remembers the reliable sources. It remembers that last week you specifically corrected it for using the outdated twenty twenty one syntax. So it starts the audit using only the verified updated pathways.

00:13:54 Speaker 1: Okay, I can see how that saves a ton of time. The second scenario is for support teams.

00:13:58 Speaker 2: Yeah, this one is huge.

00:14:00 Speaker 1: So imagine a company with thousands of customer support tickets pouring in every day. The AI agent uses its connectors to search Jira internal wikis, old emails just to find a solution to a customer's software bug. Without memory, it scans the entire haystack every single time. But with brain, the AI actually learns which specific internal documents successfully resolved past tickets.

00:14:23 Speaker 2: Exactly.

00:14:23 Speaker 1: So when a new ticket comes in with a similar error code, it bypasses the company wide search entirely. It routes directly to the proven solution document, turning, you know, a ten minute triage into a ten second automated response.

00:14:35 Speaker 2: And the third use case is arguably the most complex out of all of them. Developers debugging across massive code repositories.

00:14:44 Speaker 1: Oh yeah. That's a nightmare.

00:14:45 Speaker 2: Because modern software is sprawling, a bug on a user facing website might actually be caused by a tiny configuration file buried super deep in a back end server somewhere.

00:14:57 Speaker 1: Finding that connection is usually a needle in a haystack for human developers, let alone an AI.

00:15:02 Speaker 2: Exactly. But brain remembers the connective tissue. It remembers that the last time this specific front end module failed, the root cause was actually discovered in that obscure backend config file.

00:15:14 Speaker 1: Oh, wow.

00:15:14 Speaker 2: Because the context graph links those two nodes. So the next time a similar error pops up, the agent doesn't have to guess. It uses its internal map to jump straight to that config file, diagnosing the root cause with a fraction of the computational effort.

00:15:27 Speaker 1: So in data science, customer support, and software development, the mechanism is fundamentally the same. The agent builds on history instead of repeating it. Exactly. I mean, it sounds phenomenal. It honestly sounds like the way we all assumed AI would work when the whole enterprise hype cycle started, right? The dream we were promised.

00:15:43 Speaker 2: Exactly. But as always, we have to transition out of the theoretical highlight reel and do a hard reality check.

00:15:50 Speaker 1: Yeah, we do.

00:15:51 Speaker 2: Because we need to look at the fine print of this rollout. Yeah. There are some massive caveats that perplexity is currently dealing with.

00:15:57 Speaker 1: There are significant hurdles. First of all, while the conceptual architecture is brilliant, the access is heavily restricted right now. Restricted how? There is no public API for brain. You can't just take this context graph system and plug it into your own startup's custom applications today are okay. It's rolling out strictly in a research preview phase and only for perplexity Max in enterprise Max subscribers.

00:16:23 Speaker 2: So it's very gated right now. And beyond the gated access, we have to look critically at those numbers you mentioned earlier, right?

00:16:29 Speaker 1: The twenty five percent correctness and thirteen percent cost drop.

00:16:32 Speaker 2: Yeah, those are fantastic. But those are first party numbers. Perplexity generated those statistics from their own internal testing environments. We don't have any independent third party benchmarks to verify those claims yet. We really don't. We have to maintain a healthy skepticism until the broader enterprise community can actually stress test the system under real world, chaotic corporate conditions.

00:16:54 Speaker 1: Definitely.

00:16:55 Speaker 2: Which, uh, actually leads directly into the most critical vulnerability of this entire concept.

00:17:01 Speaker 1: The privacy and data governance implications. Yes, this is where the whole thing starts to sound less like a super helpful assistant and more like a massive corporate liability. Let's just play out the reality of a traceable context graph.

00:17:13 Speaker 2: Okay, let's do it.

00:17:13 Speaker 1: If I'm running a fortune five hundred company and I have this AI agent building a permanent, highly detailed, interconnected web of every single project my employees touch, every Slack message they send, every dead end they hit. That is a terrifying centralization of data.

00:17:29 Speaker 2: It's basically a map of the entire enterprise brain.

00:17:31 Speaker 1: Exactly. If that LLM wiki gets compromised by a bad actor, they don't just get a list of passwords. They get a holistic, perfectly organized map of exactly how my entire business operates.

00:17:41 Speaker 2: It is the single biggest hurdle for enterprise adoption. And it goes far beyond just, you know, keeping hackers out.

00:17:48 Speaker 1: What do you mean?

00:17:49 Speaker 2: It's an internal governance nightmare known as data contamination. Traceability helps with debugging, sure, but it does not inherently solve access control. Think about role based permissions.

00:18:01 Speaker 1: Right, right. So if I'm an entry level marketing coordinator and I ask the AI agent to help me draft a strategy document, how do I know the AI isn't pulling context from the CEO's private, highly classified merger and acquisition plan simply because both projects exist in the overarching brain?

00:18:17 Speaker 2: Exactly. The agent has to understand and enforce complex access hierarchies natively.

00:18:24 Speaker 1: That sounds incredibly difficult.

00:18:26 Speaker 2: That is because if the AI learns a brilliant shortcut to analyzing financial data by watching the CFO, but that shortcut involves accessing a restricted ledger, the AI has to know it absolutely cannot use that same shortcut when helping a junior accountant, right? If the context graph cross-pollinates data across different clearance levels, it creates massive internal security breaches just organically.

00:18:47 Speaker 1: Wow. So it's the ultimate double edged sword. The more deeply the AI understands the connective tissue of your company, the more devastating it becomes if that knowledge is misapplied or leaked.

00:18:59 Speaker 2: Yes, perplexity. And really any company trying to build these autonomous work memory systems will have to prove that their compliance frameworks and permission models are absolutely airtight.

00:19:11 Speaker 1: Because an enterprise will not deploy a tool that risks leaking their M&A strategy to the marketing department.

00:19:17 Speaker 2: Never. The memory is only as useful as it is secure.

00:19:21 Speaker 1: And, you know, even if they solve the security puzzle, users will still have to adapt to the friction of that overnight schedule we talked about. That's true because if I'm working on a super time sensitive crisis at two point nine oh PM and I correct the AI, it's honestly going to be pretty frustrating to know the system won't internalize that correction at a structural level until the next morning.

00:19:38 Speaker 2: Yeah, it demands a shift in how we actually manage these tools. Yeah. We have to treat them less like instant search engines and more like, say, junior employees who need time to digest their training.

00:19:49 Speaker 1: Right? It's a longer term investment in your tooling. You are actively managing and cultivating the agents map of your world.

00:19:55 Speaker 2: Exactly.

00:19:56 Speaker 1: Which brings us to the end of our journey. Today we've explored perplexities, brain system and how it signifies a massive departure from AI that just tries to be your friend.

00:20:07 Speaker 2: We covered a lot of ground.

00:20:08 Speaker 1: We did. We've looked at the mechanics of the LLM wiki, the crucial role of connectors and traceability, and how the system synthesizes data overnight to save massive amounts of compute tokens. We've seen the potential to revolutionize data auditing, customer support, and coding by basically giving agents a topographical map of their past work.

00:20:28 Speaker 2: And we've balanced that potential against the very real challenges of gated access, unverified first party metrics, and the profound data governance risks of centralizing a company's entire workflow into a single AI's memory bank.

00:20:41 Speaker 1: But before we sign off, I want to leave you with one final, slightly provocative thought to chew on.

00:20:45 Speaker 2: Okay, let's hear it.

00:20:46 Speaker 1: It's something that builds on this idea of an ever improving AI brain. If agents like computers, successfully build these flawless self-improving context graphs of a company's entire workflow. What happens to human institutional memory?

00:21:02 Speaker 2: Oh, that is a really fascinating angle.

00:21:04 Speaker 1: Think about your own workplace right now. If a senior developer or a veteran operations manager leaves the company, a massive amount of invisible knowledge walks out the door with them, right?

00:21:14 Speaker 2: They just take it all.

00:21:15 Speaker 1: They know the quirks of the database. They know exactly which department heads to bypass to get things done quickly. But if brain has been watching them for years, learning their shortcuts, mapping the dead ends they avoid.

00:21:26 Speaker 2: The AI stays behind.

00:21:27 Speaker 1: Exactly. The AI is perfectly trained on those workflows. So are we outsourcing our own institutional knowledge so completely that eventually the AI becomes the only entity in the building that actually understands how the company runs.

00:21:41 Speaker 2: It's a wild thought.

00:21:42 Speaker 1: You might not need the AI to bring you coffee anymore, but if it's the only one who knows how to keep the lights on, who is really managing who. Thank you for joining us for this deep dive. As always, staying well informed and fiercely questioning the tools you use is the best asset you have. Keep exploring.