The last decade gave us infinite storage and instant retrieval. We have more of our own data than we have ever had. We use less of it than ever. And the systems we hand that data to — the language models that can reason about almost anything — can't actually remember it across the conversations we have with them.
Walk through the apps on a competent person's phone and you'll find a strange thing. Notes from three years ago they cannot find. A reminder system they have given up on. An inbox they declared bankruptcy on twice. A calendar that reflects only what other people have asked of them. Three different to-do apps, none of them current. A journal they kept for eleven days. The information is all there. The system is not.
The problem is not that we lack tools. The problem is that all of them are waiting for us. They store what we file, retrieve what we search, generate what we ask. They are responsive in the literal sense — they respond. None of them notice. None of them hold a model of what matters to a particular person at a particular time and act on it without being asked.
That layer — the one that pays attention so the human doesn't have to — is the missing piece. I think of it as the attention layer, and I think it is the most important consumer software primitive of the next decade.
What is missing, exactly
Consider what a thoughtful human assistant does for the person they work for. They don't store information. They notice it. They notice that the cover letter is due Friday and the user keeps deferring it. They notice that two weeks have passed without a follow-up to the school. They notice that the user mentioned feeling stretched in three different conversations and brings it up gently, once, in private. They keep a continuously updated model of what is alive, what is unresolved, what is quiet but not dead, what matters this week versus what mattered last quarter.
No software does this. Not Notion. Not Obsidian. Not the personal CRMs. Not the AI chatbots. Not the meeting transcribers. Each captures a fragment — relationships, notes, conversations, calendars — but none of them holds the whole picture of a person's operational life and acts on it.
Two technical primitives are missing, and the work of building this layer requires building both. The first is memory that actually persists with salience, consolidation, and provenance — not the shallow "memory features" Claude and GPT have started bolting onto chat interfaces, which forget what matters and remember what doesn't. The second is attention itself — the layer that uses that memory to notice without being asked, to hold the model of what matters, to act on it on its own clock. Inference is cheap. Embeddings are commoditized. Native push is solved. What's missing is the substrate underneath the substrate — and the willingness to build it as the foundation, instead of bolting AI features onto an existing app shape.
What an attention layer looks like
If the attention is the product, then the home screen is what the system has noticed. Not a list of tasks the user filed. Not a graph of notes the user wrote. A short, generated paragraph in the system's own voice about what matters to this person right now, drawn from everything they have told it across all the time they have been using it.
The user did not write that. The user typed eight scattered things over the last two weeks. The system kept a model. The model is the product.
The conversation is a continuous thread, not a chat with discrete sessions. Memory carries across days. When the system mentions something from last Tuesday, it says so, and the source is tappable. The notifications, when they come — and they mostly do not come — are about specific things, with specific referents, that the user would want to know about. No streaks. No engagement loops. No marketing.
This is not a productivity app. Productivity apps make the user better at managing a system. Attention layers remove the system entirely.
Why now
Three things are simultaneously true for the first time. Inference is cheap enough to run continuously per user. Models are good enough at reasoning over personal context to be trusted with the synthesis. And the consumer is exhausted enough by the proliferation of single-purpose apps to be ready for something whose explicit promise is you don't have to manage this.
The attention layer is what gets built when a developer takes those three facts seriously instead of shipping another wrapper around a chat completion endpoint — and is willing to build the missing memory substrate underneath it, rather than waiting for the major labs to provide it as infrastructure.
What I'm building
Lila is the first attention layer — an iOS client whose entire surface is the system's continuously updated model of what matters in your life. It is built on lila-core, an open-source persistent-operator runtime that builds both missing primitives: a memory architecture with salience scoring, source-ID receipts, and nightly LLM-driven consolidation that produces a generative working-memory layer; and the attention layer on top of it — scheduled cognition, proactive reach-out, the model itself as the product. The runtime is the substrate; the iOS app is the surface; the category, I think, is new.
The thesis is not that the world needs another note app, or another assistant, or another AI feature. The thesis is that two layers between a person and the rest of their tools — the layer that holds the model of what they care about, and the layer that acts on it without being asked — have been missing this whole time, and it is finally possible to build both.
The category is new. The need has been there the whole time.