Why AI keeps forgetting you
Two very different kinds of AI — a companion you talk to and an assistant you build with — fail in the same way: they lose the thread. This page explains why that happens, what the problem is actually called, and who's working on it. It sells nothing.
You spend two hours building a world with a character — its history, your inside jokes, where the story left off. Thirty messages later it forgets your name, contradicts the plot, and answers like you just met. You weren’t doing it wrong.
You spend a week building something with a coding assistant. Today it invents files that don’t exist, undoes a decision you settled on Tuesday, and re-argues a question you already closed. It didn’t get dumber overnight.
Two completely different surfaces, hitting the same wall. What you’re feeling isn’t forgetfulness. It’s a loss of continuity— and once you can see it, you’ll see it everywhere.
Context is not continuity
The words get used interchangeably. They are not the same thing, and the gap between them is the whole problem.
A long context window can hold more text. It does not automatically know what still matters, what changed, what was decided, what failed, or what should not be touched.
It helps to separate the pieces that usually get blurred together. The context window is what the model can see this turn. Chat history is the scrollback it can see this turn. Memory stores facts for later; retrieval fetches things that look relevant to a query; project files hold reference material. Each is useful. None of them, on its own, tracks state (what is currently true), provenance (what changed and why), or continuity (all of it carried forward as one working understanding).
For the term itself — and how the AI meaning differs from “business continuity” and other uses — see what continuity layer means in AI. For the fuller structural version, the Primer lays out the hierarchy from memory to continuity.
The companion wall
If you talk to an AI companion, this is the version you've met.
A companion can remember that your character has a sister and still lose the relationship— the tone you’d built, what you’d been through together, where the story actually stood. Facts survive; the thread doesn’t. Come back after a reset and the emotional register is wrong, the roleplay has drifted, and the system answers as if the history were decoration rather than the point.
This is not you being bad at prompting, and it is not silly to care. You built something that had continuity, and the tool couldn’t hold it. That’s a limitation of how these systems carry state — not a verdict on the user.
The project wall
If you build with AI, this is the version you've met.
Deep into a real project, an assistant starts losing the architecture. Long chats go unstable — the model contradicts earlier instructions, forgets tool results it just saw, and lets the plan diverge from the code. It overwrites decisions that were working. People describe the tipping point the same way every time: “it went insane.”
The same wall, everywhere
Once you name it, the pattern stops looking like your app being weird and starts looking like a property of the whole field.
The name for it: continuity
The canonical definition this project uses, cited verbatim across the site.
A continuity layer is the part of an AI system that carries forward governed state across sessions, tools, and models: what is currently true, what changed, why it changed, and which evidence supports it. It sits above memory and retrieval, and below agent execution, turning stored history into usable, auditable continuity.
That is what the two stories at the top were both missing. This project maps the systems, papers, products, demos, and open questions forming around it — starting from a plain-English path in if the vocabulary is new, or the full map if you want the field at a glance.
What the current systems actually do
A neutral survey — no rankings, no “best”. Each entry is described from its own public documentation; the point is what the approach reveals about continuity, not who wins.
Each of these is tracked as it evolves in Signals and the Field Watch profiles, with dates and primary sources attached.
What helps today — and what still breaks
Real ways people reduce the pain right now. None of them is the missing layer; each is honest about its ceiling.
These are patches, and useful ones. What they have in common is that each holds a fragment — a fact, a summary, a file, a rule — without governing the whole. The field is moving toward continuity as a deeper layer: one that carries current, auditable state forward on its own, so the work doesn’t have to be re-explained from zero every time the session, the tool, or the model changes.