Blog

The memory is the moat

Last updated: April 2026

Every few months a new frontier model arrives and the benchmarks shuffle, but if you've been watching closely the pattern is hard to miss: the gap between the leading models is narrowing. Prices are falling. The model that was state-of-the-art eighteen months ago turns out to be good enough for the vast majority of tasks people actually need done, and the one that replaces it in another eighteen months will cost a fraction of what the current generation does. The engine is becoming a commodity.

But the context, everything a system knows about you specifically, your work, your history, your organisation, the particular way your team uses certain words when they mean certain things, is a different story entirely. Context is hard to build, impossible to fast-forward, and enormously valuable once you have it, and it's increasingly where the real competitive separation lives.

The zero-context trap

Most people and most companies still use AI the way they use a search engine. Open a window, ask a question, get an answer, close the window. Every conversation starts from nothing. The model knows an extraordinary amount about the world in general and absolutely nothing about your situation in particular.

This is roughly equivalent to hiring a brilliant consultant who happens to have amnesia. Every morning they arrive knowing everything about their field and nothing about your company, so you spend the first stretch of every engagement re-explaining your context, your constraints, your history, and the things you've already tried. By the time they're properly oriented, you've burned a meaningful portion of the session, and tomorrow the whole cycle starts again.

The amnesiac consultant is still useful, sometimes impressively so, but they're operating at a fraction of what they could deliver if they carried forward the understanding from last week, had absorbed your existing research, and could reason about your specific situation without needing to be briefed from scratch every time.

This is where most AI usage sits today, at both the individual and enterprise level, and it's worth understanding why.

The gap between access and understanding

The enterprise world spent 2024 and 2025 solving the access problem. Connectors, MCP servers, API integrations, tool-use frameworks. The idea was that once the AI could reach your data, it could work with your data. And in a narrow sense that's true: you can now point a model at your Slack, your email, your CRM, your project management tool, and ask it questions.

What companies are discovering, often after significant investment, is that access produces retrieval but not understanding. The AI can search your Slack history and find a message. It can scan your Google Drive and return a document. But ask it who owns the relationship with your biggest customer and it pulls a name from the most recent email thread, which may or may not be the right answer. Ask it from a different tool and you get a different name. Ask about your top priorities this quarter and it surfaces whichever strategy document it finds first, even if that document is six months old and three pivots behind.

The system can find information perfectly well. The trouble is that it finds too much, can't tell what's current, can't resolve conflicts between sources, and presents a fragment as the whole truth with considerable confidence.

Understanding requires synthesis. Knowing which information is current and which is stale. Knowing which sources are authoritative when they contradict each other. Recognising that "lisa.chen@acme.com" in email, @Lisa in Slack, and "Lisa from Acme" in a meeting transcript are the same person, and that everything she's said across every channel should be unified under one identity. Combining the project tracker, the calendar, the hiring pipeline, and last week's standup notes to arrive at an insight that none of those sources contain individually: the infrastructure migration is at risk because the lead engineer is out next week, the dependency on the payments team is unresolved, and the original timeline assumed a new hire who hasn't started yet.

A competent employee could tell you all of this after a few months on the job. No AI system reliably can, because almost nobody is building the synthesis layer that would make it possible.

The compounding problem

What makes this particularly interesting from an investment perspective is that contextual understanding compounds in a way that model capability doesn't.

A better model gives you a step-function improvement on the day you switch to it. Better context gives you an advantage that grows every day the system runs. On day one, the context layer knows a little. On day thirty, it has absorbed thousands of messages, hundreds of documents, dozens of meetings. It has resolved conflicts, built identity maps, tracked what changed and what didn't. The understanding on day thirty is qualitatively different from day one, and the understanding on day 180 is different again.

Every new data point makes the existing context more valuable, not by adding one fact but by potentially confirming a project status, updating a relationship, revealing a priority shift, and resolving a conflict between older sources. The graph gets denser, more connected, and more useful at an accelerating rate.

You can't fast-forward this. You can't write a cheque and skip to month six. The understanding has to accumulate, which means the cost of waiting isn't "we don't have it yet" but "we're falling further behind every day we don't start." A company or individual that begins building their context layer today will have six months of compounded understanding that no amount of spending can buy later.

The moat, when you look at it clearly, is the accumulated, synthesised, conflict-resolved understanding of a specific company or a specific person's knowledge. Models can be switched and connectors are increasingly standardised, but that accumulated understanding is proprietary, it compounds, and it only grows with time.

The personal version

The same dynamic applies at the individual level, and it's where the argument gets interesting for anyone thinking about personal productivity tools.

Consider two researchers who both use frontier models and are both capable people. The first uses AI in the zero-context way: opens a chat, asks a question, gets an answer, moves on. The second has spent six months building a personal knowledge library: capturing articles, annotating papers, writing literature notes, saving sources, recording voice memos about ideas, and running all of it through a system with semantic search that finds things by meaning rather than filename.

When the second researcher asks a question, the AI draws on six months of accumulated context, and the answer that comes back is grounded in their specific reading, their specific thinking, their specific research trajectory rather than in whatever the internet happens to surface for that query. On day one, the difference between these two researchers is barely noticeable. By day 180, the second researcher has a thinking partner that knows their work while the first still has a very fast search engine that knows nothing about theirs.

This is the consumer version of the enterprise context graph. Your knowledge library is your personal context layer: the accumulated notes, articles, annotations, and original writing that represents what you know and what you've been thinking about. Through MCP, that library becomes the memory layer for any AI agent you use, decoupled from any specific model or vendor. You can swap models freely because the library persists regardless of which one you're running.

Depth over breadth

There's a related observation from the enterprise side that's worth drawing out: the consumer AI market rewards breadth while the enterprise market rewards depth.

Consumer AI can tolerate false positives because there's a person in the middle to catch them. You ask Gemini to write an investment memo and a wrong line or two doesn't matter because you're going to read it and edit it before it goes anywhere. The value proposition is speed, not accuracy.

Enterprise AI, and particularly agentic AI where the system acts rather than advises, has essentially zero tolerance for false positives. An agent that processes invoices and gets one wrong creates a real problem. An agent that screens CVs and mischaracterises a candidate's experience creates a legal one. The depth of domain-specific understanding required to operate without a human in the loop is closer to what it took to build Waymo (tens of billions of dollars of edge-case training on data that exists nowhere on the internet) than what it took to build a chatbot.

The same principle applies to personal AI, though the stakes are lower. A generic AI writing a literature review for you produces generic output that reads like someone spent twenty minutes on Google Scholar. An AI grounded in your personal library of annotated research, documented thinking, and accumulated notes produces something connected to your actual work. The depth comes from the context you've accumulated rather than from the model's general capability.

Where this is going

The transition from retrieval to understanding is the defining shift for the next phase of AI, at both the enterprise and personal level.

2024 and 2025 were about giving agents access to tools. Connectors, APIs, MCP servers. That work was necessary and it's largely done. The infrastructure exists for AI to reach your data.

2026 and beyond is about building the synthesis layer: continuously updated, source-grounded, conflict-resolved representations of reality that agents can read from rather than searching from scratch every time. At the enterprise level, this looks like what some people are calling a "company brain." At the personal level, it looks like a knowledge library that grows with you, learns what you know, and makes every AI interaction more useful because the context is already there.

The companies and individuals who build this layer now are making an investment that compounds in a way that model selection never will, because the engines are converging while the fuel, the accumulated context that makes those engines useful for your specific work, takes time to build and can only grow through use.

The people building context today will have a kind of advantage in six months that the people who start then can't catch up to by spending more, because this particular advantage has to be grown rather than purchased.

Frequently asked questions

If models keep improving, won't they eventually work without personal context? Models will improve at general reasoning, but general reasoning applied to a specific situation will always be less useful than contextualised reasoning grounded in specific knowledge. A brilliant doctor who doesn't know your medical history is less useful than a good doctor who does, and the same principle holds regardless of how brilliant the doctor becomes. Context doesn't become less valuable as the model improves. It becomes more valuable, because a better model can do more with better context.

How is a personal knowledge library different from ChatGPT's memory? ChatGPT's memory stores a paragraph of facts about you derived from conversations. A knowledge library contains thousands of documents, notes, annotations, and connections accumulated over months of actual work. The difference in scale changes what's possible entirely: a paragraph of instructions tells the model who you are, while a library spanning thousands of items shows it what you know and how you think. The ownership difference matters too, since with a library you control the context rather than renting it from a conversation history.

Does building a context layer mean being locked into one vendor? The opposite, if designed correctly. When your context lives in your own library rather than in a provider's conversation history, you can switch models freely. Open protocols like MCP make the context portable. The library persists regardless of which model you happen to be running, which is precisely the point: the context should outlast any individual model generation.

What about privacy and data ownership? This is one of the strongest arguments for building your context layer in a system you control rather than handing it to a provider. The context is the moat, which raises the question of who should own it. A private, encrypted library that belongs to you, with API access on your terms, means the context compounds for your benefit without being extracted, retrained on, or held hostage by anyone else.