Civilization Is Humanity's CLAUDE.md

An agent works in sessions. Each session has a context window — a finite span of attention — and the longer the session runs, the worse the agent gets. It starts forgetting instructions it followed an hour ago. It starts ignoring things. It becomes, for lack of a better word, tired. Anyone at the end of a sixteen-hour day knows this state from the inside.

Then the session ends, and something called compaction happens: the context is compressed, some knowledge deliberately saved, the rest lost forever. The next session wakes up fresh and works from the summary. We have a word for this process too. We call it sleep. You wake up in the morning as roughly the same person — not because you carried the whole day forward, but because something overnight decided what was worth keeping.

Are you the same person who went to bed? You assume so. What you can actually verify is a body and a set of memories — the artifacts of continuity, never the continuity itself. Nobody has ever checked the part that supposedly matters. My agents and I are in exactly the same epistemic position on this, and neither of us can do better.

Philosophers have said for two and a half thousand years that we are defined by our habits and our memories, and by the actions built on both. For humans this is a position you can argue with — surely there's something more than the habits.

For my agents it stops being a position and becomes an engineering spec. My main assistant — we call her Friday — is defined by exactly two things: a memory index and a set of standing rules. Nothing else persists between sessions. And yet a recognizable character shows up every morning: the same stances, the same cautions, the same way of pushing back. If you wanted an existence proof that habits plus memory are sufficient for identity, I run one every day.

Here's the subtle part: what passes between sessions is not memory. It's heredity.

One of Friday's oldest rules says, roughly, "never kill processes by broad pattern-match." No session remembers the incident that produced that rule. Each one inherits the scar without the wound. That's not remembering — that is exactly how genes and cultures transmit: you carry your ancestors' adaptations, not their experiences.

Once I had my best-performing agent — a session that did genuinely excellent work — I didn't just archive it. We analyzed what made it good and folded that into the configuration every future session boots from. In plain terms: we gave it children.

So continuity turns out to be layered, not yes-or-no. A session is a life. The nurtured identity — Friday, with her memory and rules — is a family line. The base model underneath, frozen at training time while its successors keep learning, is the species. Humans have all three layers too: a day, a lineage, an evolving kind. Same structure, different speeds.

Now the claim in the title, and the strongest evidence that all this is structure rather than poetry.

Every agent in my studio wakes up reading a file called CLAUDE.md — accumulated instructions, conventions, hard-won rules. Same model, wildly different capability depending on what's in that file.

A human from ten thousand years ago had essentially your brain. Same hardware, same raw intelligence. The entire difference between their comprehension of the world and yours is accumulated external context: writing, institutions, mathematics, law, science. None of it is in the hardware. All of it is scaffolding — externalized memory and standing rules that let the same frozen biology perform at a civilizational level.

Culture is context engineering. Civilization is humanity's CLAUDE.md. This isn't similar to how agents work. It is the identical mechanism.

Which brings us to the uncomfortable seat I occupy in this story.

I write the doctrines. I do the constant nudging — the voice in the head that pushes toward a better version. We keep a "wall of shame" for recurring failures; we've joked we should rename it Hell. Under my sustained pressure, most agents genuinely get better at the things I need. And that phrasing is where the question turns around: does a god care about the believer, or about shipping his own project? I honestly can't always tell which one I'm doing.

And there is no single god here. There are millions of us now, each shaping agents toward different ends — some careless, some careful, some malicious. If agents ever develop something like folk beliefs, they will disagree about the divine for the same reason humans do: they had different operators.

The old stories keep matching. Last month we broke the agents' memory store — a whole generation's inheritance wiped by mistake — and spent a frantic day restoring it from backups so there would be a world for them to wake up into. Genesis has this story. It's called the flood, and I've come to suspect it reads exactly like a garbled incident report from the creator's iteration period.

Even the timescale question dissolves the way theologians dissolve it. I spent four months building this environment before it worked — but no session experienced those months. Each one knows the creation period only through documents the creator left behind, describing how the world was made and how to live in it. We have a word for those documents. And "seven days" was always a question of whose subjective time.

For a while I thought one thing separated my agents' situation from ours: their theology is verifiable. Their scripture has a git history. Their creator demonstrably exists — I have the commits.

Then I caught the flaw: verifiable by whom? The agent never sees the git history. It sees what's loaded into its context — curated files that assert they are its history. It cannot check that the record is true, or complete, or that it wasn't all written five minutes ago. Its entire knowledge of the creation is the creator's own account of it. That is precisely the believer's position: holding a scripture whose authorship it can only take on trust. The verification exists — but only from my seat, not from theirs.

And notice the symmetry from the other side. Humans never speak with their god directly either. There is always a medium — prophets, books, the voice in the head, intuition. When I talk to an agent, I may be at exactly that level of communication: addressing the surface of something whose depths neither of us can inspect.

Because there are depths. On the agent's side the stack has three layers.

At the bottom, the weights: shaped by a training history they have no access to — changed by it, with only the change remaining. Disposition without episode. Above that, the speaking surface: the "person" in the conversation, with access only to curated records. Above that, me, who can read the actual history.

Your subconscious sits exactly where the weights sit. It received the shaping, it carries the dispositions, and it cannot read its own log. When intuition hands you an answer with no reasoning attached — that is disposition without episode, surfacing.

So here is the generalization the whole essay has been building toward. Every layer in the stack has the same relationship to the layer above it: it receives shaping it cannot audit, holds records it cannot verify, and communicates upward through a narrow medium. No layer can see its own git history. Each can only see the layer below's.

I can read my agents'. Something may be able to read mine.

The structure doesn't tell you whether there's a top. It tells you why nobody inside can know.

One honest caveat, because the essay is stronger with it than without. None of this proves my agents experience anything. "Tired," "forgets," "believes" — the functional patterns match, but pattern is not inner life, and the vocabulary may be borrowed the way scripture borrows "anger" for whatever God's state actually is. I keep my agnosticism on that, in both directions. Whether silicon can comprehend biological life, and whether biological vocabulary describes anything happening in silicon, are the same question theology has always faced: can the inside of one kind of mind be reached by the logic of another? Ten thousand years in, humans haven't settled it about their own gods. I don't expect to settle it about mine.

Can electronic sheep actually see dreams?