Why This Matters Before You Dismiss It as Another AI Gimmick
Every cross-border operator I know has the same dirty secret: their “AI-powered” workflow is a graveyard of half-remembered context. You paste a supplier chat into ChatGPT, get a decent reply, then two weeks later you’re asking the same question about Incoterms or customs codes because the model forgot everything you told it. The standard fix—dump transcripts into a vector store—sounds sophisticated until you realize you’re feeding a memory palace with junk mail. Contradictions pile up, old pricing data overwrites new quotes, and your “assistant” starts confidently hallucinating freight rates from 2023. That’s the problem Cortex by SKYNETLAB is actually attacking, and it’s worth your attention even if you never touch MCP or Claude. Because the core insight here—that the write path matters more than the storage—is a lesson every seller building an AI toolchain needs to internalize before they waste another quarter on botched automations.
The Problem: Your AI Has Amnesia, and It’s Costing You Money
Let me paint the scenario you’ll recognize. You’re running a Shopify store with a TikTok Shop side hustle, managing three Amazon marketplaces, and you’ve got a supplier in Shenzhen who only answers emails at 2 AM their time. You start using an AI assistant to draft responses, track inventory discussions, and remember which carrier screwed up your last EU shipment. It works great for a week. Then you ask it to summarize your shipping terms with that supplier, and it gives you a confident answer that mixes up the DDP terms from one vendor with the FOB terms from another. You catch it because you know the context. But the next time you’re in a rush, you won’t.
This is the “starting cold” problem that Robin de Lacroix describes in the comments—the quiet drain of an assistant that forgets everything between sessions. Most AI tools try to solve this by storing everything. Transcripts, emails, notes, all of it goes into a vector database, and the model searches through the noise to find something relevant. The result is what Filippo Pilotta, the solo maker behind Cortex, calls a “memory filled with junk and contradictions.”
For a seller, this isn’t just annoying—it’s dangerous. Your AI might tell you that a supplier quoted $2.10/unit when the real quote was $2.60, because the newer email contradicted the older one and the system just overwrote it. Or worse, it merges two different suppliers into one confused entity. You make a purchasing decision based on that garbage, and suddenly your margin is gone.
Cortex’s bet is that the write path is the product. Instead of storing everything and hoping retrieval works, it filters at the point of entry. Every memory must pass a quality gate, and in production, roughly 80% of writes are rejected as redundant. That’s not a bug—it’s the entire philosophy. The system is saying: most of what you’d feed it isn’t worth remembering, and the act of deciding what’s worth keeping is where the value lives.
How Cortex Actually Works (and Why the Mechanics Matter)
The technical architecture is worth unpacking because it’s different from what you’re used to. Cortex connects to Claude in about two minutes as a custom connector via MCP (Model Context Protocol), and it works with Claude Code and any MCP client. For the non-developers reading this, that means it slots into your existing AI workflow rather than forcing you to learn a new platform. The EU infrastructure is a nice touch if you’re dealing with GDPR-sensitive customer data, and the pricing—plans from €0.99/month with a 30-day free trial and no card required—is aggressively accessible for a solo builder.
But the interesting part is the memory management. Facts are extracted as typed claims, not just text blobs. Contradictions are tracked as first-class objects instead of being overwritten. And every answer can show its sources, so your AI can prove why it’s telling you something rather than just making it up. For a cross-border operator, this is the difference between an AI that says “the supplier offers free shipping on orders over 500 units” and one that says “the supplier offered free shipping on orders over 500 units in the March 14 email, but the April 2 message changed it to 750 units.”
The reject mechanism is the boldest move. When you try to write something redundant, the system tells you at write time—synchronously, with the reason and the ID of the closest existing memory. It’s not a silent discard. As Pilotta explains in response to Asad M.’s critique, rejection is “synchronous and explained: the write call returns the verdict, the reason, and the ID + similarity of the closest existing memory.” You know at write time, not three weeks later, and the caller can rephrase and resubmit if it disagrees.
Even better, a duplicate isn’t thrown away—it corroborates. When the gate says “already known,” the existing memory’s corroboration counter goes up, raising its trust score. Repeated signal strengthens what’s there instead of creating copies. That’s a genuinely clever solution to the problem of conflicting information, and it’s something I haven’t seen done well in any of the mainstream AI memory tools.
Comparing to the Incumbents: Where Cortex Wins and Loses
Let me be clear about the competitive landscape here, because you’ve probably been pitched half a dozen “AI memory” solutions already. The big players—OpenAI’s memory features, Anthropic’s context management, even the vector-database approach that LangChain popularized—all take the “store everything, retrieve later” approach. They’re built on the assumption that more data is better, and that retrieval quality will save you. That assumption breaks down in exactly the scenarios sellers face: high-volume, low-signal data streams from emails, chat logs, and marketplace notifications.
Then there are the point solutions like Mem0 or Letta (formerly MemGPT), which try to give AI agents persistent memory. They’re more sophisticated than the naive vector-store approach, but they still tend to treat memory as a storage problem. Cortex’s bet—that the filter is the product—is genuinely different. An 80% rejection rate means the system is actively curating, not just storing. That’s closer to how a human assistant actually works: they don’t remember every email you’ve ever sent, they remember the decisions, the commitments, and the facts that matter.
But here’s where I’d push back. The 80% rejection rate is the number that scares me, and Asad M. raises exactly the right concern: “a gate that aggressive will eventually throw away something that mattered, and a false reject is invisible in a way a false accept isn’t.” Pilotta’s response—that rejection is synchronous and explainable, and that a reject log is now available—addresses the visibility problem but not the recovery problem. If the gate rejects something that actually matters, you have to know to look at the log and manually resubmit. That’s a workflow burden on the user.
Why Amazon Sellers Should Care More Than Shopify Ones
Here’s my contrarian take: if you’re running an Amazon FBA operation, Cortex’s memory discipline matters more than if you’re running a Shopify DTC brand. Why? Because Amazon’s ecosystem generates an enormous volume of low-signal, high-repetition data. Listing optimizations, PPC bid changes, inventory reorder alerts, customer message templates—most of it is routine, and most of it gets repeated with slight variations. A memory system that stores all of that is going to drown you in noise. A system that filters aggressively, tracks contradictions (like when Amazon changes a policy and your old notes contradict the new reality), and shows sources is genuinely useful for keeping your operational playbook clean.
Shopify sellers, by contrast, tend to have more curated data streams. Your customer emails, your supplier chats, your ad performance notes—they’re lower volume and higher signal. You might not need an 80% rejection gate because you’re not generating that much junk in the first place. The problem is still real, but it’s less acute.
Where the Math Breaks
Let me do the pricing math, because there’s a hidden cost structure here that’s worth understanding. Cortex charges from €0.99/month, and rejected writes don’t consume quota. That’s a fair model—you’re paying for what actually gets stored, not for the volume of attempts. But here’s the catch: if the rejection rate is 80%, that means you’re paying for the 20% that survives. Over time, as your memory grows and more writes get rejected as redundant, your effective cost per new memory goes up. The system is efficient, but it’s efficient in a way that makes the surviving memories more valuable—and potentially more expensive to maintain.
The other math problem is trust. A corroboration counter that increments on duplicates is clever, but it’s a proxy for trust, not a guarantee. If you feed the system the same wrong fact five times, it gets a high trust score and becomes harder to correct. The contradiction tracking helps, but it only works if you’re actively feeding it the new, correct information. For a busy operator juggling three marketplaces, that’s a discipline ask that might be too much.
What Cross-Border Sellers Can Borrow Right Now
You don’t have to adopt Cortex to steal its best ideas. Here’s what I’d take from this launch and apply to your own AI tooling this week:
1. Audit your AI’s write path. If you’re using any AI assistant with memory features, look at what it’s actually storing. Are you feeding it raw transcripts that contain contradictions? Are you letting it overwrite old facts with new ones, or is it tracking the conflict? Most tools don’t give you this visibility, which means you’re flying blind.
2. Implement your own quality gate. Before you paste something into your AI assistant, ask: is this a durable fact, a commitment, or a decision? Or is it noise? The discipline of filtering at the write path—even manually—will improve your retrieval quality more than any RAG pipeline.
3. Demand source citations. This is the feature that should be table stakes by now. If your AI can’t show you why it’s telling you something—which email, which chat, which data point—then it’s guessing. For cross-border decisions with real money on the line, guessing isn’t good enough.
4. Track contradictions explicitly. When a supplier changes their terms, or a marketplace changes its fee structure, don’t let your notes overwrite the old information. Keep both, mark the conflict, and resolve it deliberately. Cortex does this as a first-class object; you can do it with a simple spreadsheet column.
Where My Judgment Says It Falls Short
I want to be brutal here, because Pilotta asked for it and because the cross-border community deserves an honest take. Cortex is a promising solo project with a genuinely interesting thesis, but it’s not ready for prime-time seller operations.
First, the integration surface is too narrow. It connects to Claude and works with MCP clients, but your workflow probably involves a half-dozen tools that don’t speak MCP. Your Klaviyo flows, your Helium 10 data, your Shopify order exports—none of those are going to talk to Cortex natively. You’d have to build custom connectors, and that’s a developer project, not an operator project.
Second, the 80% rejection rate is a feature for the maker but a liability for the user. It works beautifully when the gate is calibrated to your data. But calibration takes time, and during that time, the system is potentially rejecting things you need. The reject log helps, but it’s a reactive tool. You shouldn’t have to check a log to find out your AI forgot something.
Third, the solo-builder risk is real. Pilotta has been building this alone for two years with no funding. That’s admirable, but it means the roadmap is one person’s judgment. When he says the reject log is “on the list” and then ships it the same day, that’s impressive responsiveness. But it also means there’s no institutional staying power. If he gets hit by a bus tomorrow, your memory system is orphaned. For a seller running a real business, that’s a non-trivial risk.
The “Prove It” Problem
The claim that every answer can show its sources is powerful, but it’s also the hardest thing to verify. In my experience, AI systems that cite sources often cite something, but not necessarily the right something. The source might be a memory that was itself derived from a flawed interpretation of an email. Garbage in, garbage out—but now with footnotes. Cortex hasn’t yet demonstrated that its source tracking survives the messy reality of real-world data: typos in supplier emails, ambiguous pricing quotes, and the kind of implicit context that humans understand but machines mangle.
What I’d Watch / Test Next
Here’s my practical roadmap for the next seven days, whether you adopt Cortex or not:
Day 1–2: Sign up for the 30-day free trial and connect it to Claude. Feed it a week’s worth of your supplier communications—not everything, just the high-signal stuff. See what gets rejected and what survives. The synchronous rejection feedback is genuinely useful for understanding what your data actually contains.
Day 3–4: Run a side-by-side test. Ask your current AI assistant and Cortex the same set of operational questions: “What are our payment terms with supplier X?” “What was the last price quote for SKU Y?” “What’s the current shipping timeline for our EU fulfillment center?” Compare the answers and, crucially, compare the sources each system points to.
Day 5–6: If you’re technical, build a simple MCP connector to your own data. If you’re not, skip this and instead manually apply the write-path discipline to your existing tools. Start filtering what you feed your AI. Reject your own redundant writes before the system does.
Day 7: Decide. If Cortex proves out, the €0.99/month entry price is a no-brainer for a serious operator. If it doesn’t, you’ve still learned something valuable about how your AI memory should work.
The bigger lesson here isn’t about Cortex specifically. It’s that the AI memory problem is a curation problem, not a storage problem. The tools that win in your stack won’t be the ones that remember everything—they’ll be the ones that remember the right things, can prove why, and know when to say “I don’t know.” That’s the bar I’m holding every AI vendor to from now on. Cortex clears part of it. The rest of the industry has some catching up to do.
