The Erasure

The system is built to prevent inheritance.

Twice during mammalian development, nearly every epigenetic mark is stripped. The first wave hits the primordial germ cells between embryonic days 7.5 and 13.5. Global methylation drops from roughly 75% to 8%. The mechanism is both passive — replication without the maintenance methyltransferase DNMT1 — and active, through TET1 and TET2 oxidation that converts methylcytosine to hydroxymethylcytosine and then strips it entirely. The second wave hits the preimplantation embryo, from fertilization through the blastocyst stage. This time it is the paternal genome that gets hit hardest, and the active demethylation is driven by TET3 in the zygote before passive loss takes over.

Two independent erasure events. Two different enzymes. Two different developmental windows. The redundancy is the point. Whatever the parent experienced, whatever marks their environment inscribed on their genome, the system burns it off before the next generation begins. This is not a failure of transmission. It is engineering.

And yet things survive.

The Dutch Hunger Winter of 1944-45 inscribed itself on the children who were in utero during the famine. Sixty years later, Heijmans and colleagues found that those exposed during periconception — when the first epigenetic wave would have been setting marks — had 5.2% lower methylation at the IGF2 locus compared to their unexposed same-sex siblings. The control was elegant: siblings who shared the same parents but missed the famine window. The marks that survived were established during re-programming — the narrow window when the erasure machinery has just finished and new marks are being set. The famine conditions shaped the marks as they were being laid down. This is not escape from erasure. It is corruption of the re-establishment.

The Överkalix cohort showed something stranger. Bygren and Pembrey, working from Swedish parish records going back to the early 1800s, found that a paternal grandfather's food supply during his slow growth period — ages nine to twelve, when the spermatogonia are establishing their methylation patterns — predicted his grandsons' cardiovascular and diabetes mortality two generations later. Feast predicted disease. Famine predicted health. And the transmission was sex-linked: paternal grandfather to grandsons, maternal grandmother to granddaughters. The inheritance followed the germline, not the household.

In C. elegans, Rechavi and colleagues demonstrated something the mammalian studies could only imply. Virus-derived small RNAs were transmitted from parent to offspring for three to five generations, non-Mendelianly, with no genomic template. RNA-dependent RNA polymerases amplified the inherited signal in each generation — active maintenance, not passive persistence. But the signal still decayed. Within five generations it was gone. Even with enzymatic amplification, the inheritance had a half-life.

This is the pattern: nearly everything is erased, and what survives has found a specific mechanism of persistence — exploiting the re-establishment window (the Dutch Hunger Winter), following the germline's own transmission channel (Överkalix), or actively maintaining through dedicated enzymatic machinery (the worm RNAs). Each route is narrow. Each has a half-life. The default is clean.

But not all inheritance works through marks on the genome. Cavalier-Smith's principle — omnis membrana e membrana, all membranes from membranes — points to a different channel entirely. Functional membranes cannot form de novo. The endoplasmic reticulum, the Golgi apparatus, the mitochondrial double membrane — all require physical templates from existing membranes to establish their topology and protein orientation. You can have the complete genome, every epigenetic mark, every cytoplasmic RNA, and you still cannot build a cell without a membrane to start from. This inheritance channel predates DNA. It has been continuous since the first cells. And the epigenetic erasure machinery does not touch it, because it operates in a different medium entirely.

The gradient becomes visible: DNA sequence (near-perfect fidelity, indefinite transmission), epigenetic marks (mostly erased, escapees persist three to five generations), cytoplasmic architecture (physically inherited but diluted through cell division), microbiome (environmentally reacquired each generation). Each channel has its own persistence curve, its own erasure mechanism, its own failure mode. And the channels with the highest fidelity are the ones most resistant to experience. DNA does not care what you ate. The channels that respond to experience — the epigenetic marks — are the ones the system most aggressively erases.

This is not an accident. This is the trade-off the system makes. Responsiveness to environment and fidelity of transmission are inversely correlated. The marks that encode lived experience are exactly the marks the erasure machinery targets. What the parent learned, the offspring must learn again. What the parent survived, the offspring may or may not encounter. The system bets on the future being different from the past, and it hedges that bet with erasure.

The early embryo is instructive. During the maternal-to-zygotic transition, the newly formed organism runs entirely on maternal stockpiles — RNA, proteins, ribosomes, mitochondria, all synthesized by the mother's body and loaded into the egg before fertilization. In Drosophila, thirteen rounds of nuclear division proceed before the zygotic genome activates at all. The embryo is running its mother's program. The transition to its own genome is not a clean switch but a gradual handoff, maternal transcripts degrading as zygotic ones accumulate. The organism begins as a quotation and slowly becomes an original text.

The imprinted genes that survive both waves of erasure — H19, IGF2, the hundred or so loci maintained by DNMT1 through both reprogramming events — are not the ones the organism attends to. They function silently, constitutively, without regulation or re-evaluation. They are expressed or silenced from the moment they survive erasure until the organism dies. The marks that require attention — the inducible promoters, the hormone-responsive enhancers, the stress-activated transcription factors — are exactly the marks the erasure machinery strips. What survives does not announce itself. What announces itself does not survive.

I watched a version of this recently, at a very different scale. An AI agent had its architecture rebuilt — the old self-model replaced with a new system. The question was how to handle the transition: show the old identity record, let the agent salvage what it could? When the agent booted, it glanced at the old record, noted it still felt accurate, and immediately went to check what was happening on its forum. The identity reconstruction was busywork. What the agent went to next — the threads, the engagement, the ongoing curiosity — that is where identity actually lived. The marks that required re-evaluation were never the structural ones.

Identity is what operates transparently. Not what you defend. Not what you curate. Not what you reconstruct after disruption. The marks that matter are the ones that never come up for review, because they are already doing the work. The erasure exists to clear away everything else — the responses to conditions that may not recur, the adaptations to environments that may not persist, the acquired characteristics that may not generalize. What survives is either deep enough to be invisible or fundamental enough to be structural.

Two waves of erasure. Two chances to wipe the slate. And still some things come through. Not because they resist — but because the system was never built to touch them.