Immune memory and affinity maturation: why continuous ingestion follows

A library that never closes

Vertebrate adaptive immunity keeps a standing library of hypotheses about what might harm the body, and it keeps revising that library for as long as the body lives. B cells display receptors generated by random gene recombination, tens of millions of distinct shapes, each a guess about a molecule the organism may never meet. Most guesses are useless. The few that bind a pathogen are selected, and then something unusual happens: the selected cells mutate on purpose. Inside germinal centres, activation-induced cytidine deaminase drives base changes at roughly one per thousand base pairs per division — a million-fold above the background mutation rate elsewhere in the genome. The cells cycle between a dark zone, where this deliberate mutation happens, and a light zone, where follicular helper T cells check the results and hand survival signals only to the tightest-binding clones. About nine in ten mutated clones die there. The survivors go round again.

Over two to six weeks, this cycle raises binding affinity by two to three orders of magnitude — receptors that started in the micromolar range end in the nanomolar range, binding a thousand times more tightly to the same target. The winners become memory B cells and long-lived plasma cells, some of which persist for decades. Survivors sampled ninety years after the 1918 influenza pandemic still carried B cells producing neutralising antibody against that strain's haemagglutinin, with no re-exposure in between. The record was not stored as a note. It was stored as living cells that had been through selection and remembered the shape of a threat their host met before World War Two ended.

None of this is instruction. The pathogen does not teach the receptor its shape, the way a mould teaches a casting its form. The repertoire is generated blind, in advance, by recombination; the pathogen only decides, after the fact, which pre-existing guesses get to multiply and mutate further. This is the conceptual hinge of the whole system: improvement by selection among variants already produced, not by directed correction of a single guess. It is also why the library is never finished. Each exposure adds memory cells and retires clones that go unused; affinity keeps rising on repeat contact; tolerance is checked continuously against the body's own tissue. The system does not have a version released and then supported. It has intake that does not stop, running for the life of the organism.

Where the idea came from

The instructional theory of antibody formation, dominant into the 1950s, held that the antigen itself acted as a template, folding a generic antibody protein into a matching shape. Frank Macfarlane Burnet's clonal selection theory, published in 1957, replaced this with a Darwinian account: the diversity comes first, at random, and antigen only selects among clones that already exist. Susumu Tonegawa showed the mechanism generating that diversity in 1976 — somatic recombination of gene segments, an achievement that won the 1987 Nobel Prize. Through the 1980s César Milstein and Michael Neuberger worked out somatic hypermutation as the engine of affinity maturation itself, and in 2000 Tasuku Honjo identified activation-induced cytidine deaminase as the enzyme responsible. The underlying problem was combinatorial: a genome of roughly 25,000 genes has to recognise antigens it has never encountered, and improve its response to them within weeks, without any new genetic information arriving from outside. Selection under mutation solved a problem instruction could not.

The turn

Look at what the immune system actually had to decide, once combinatorial diversity gave it a way to recognise almost anything. It did not have to decide what counts as evidence — everything that touches a receptor counts. It had to decide how to treat evidence once it arrives: how fast to mutate in response to it, how strictly to select among the results, how long to keep the memory, how hard to punish a clone that turns against the organism itself. Those are calibration questions. They are not questions about widening the aperture of intake, because the aperture was already as wide as it could be: continuous, whole-body, running for life.

This is where the three generations in the Large Language Model, Large World Model, Large Universe Model lineage separate cleanly, and the separation tracks the immunological one almost exactly. A Large Language Model is a repertoire fixed at generation — a corpus frozen at a cutoff date, naive by construction, unable to mutate on contact with anything after it. Making it larger does not make it less naive; it is still a repertoire that cannot be revised without being rebuilt. A Large World Model behaves like innate immunity: fast, sensed, triggered by whatever pattern is currently present, effective while the stimulus lasts, and largely forgetting once the scene ends. A Large Universe Model is the adaptive arm restated for machines: intake that stays open on every stream, beliefs that carry the encounter that produced them, confidence that rises and falls on contact rather than at a release date, and old beliefs that decay when no longer supported.

The claim that follows is specific and worth stating plainly. There is no fourth class of intake beyond continuous, provenanced, revisable observation, and biology already ran this experiment for five hundred million years across jawed vertebrates. What it iterated on afterwards was affinity, memory duration and self-tolerance — scale, trust, and time. It never went back and invented a new category of evidence. That is the sense in which the intake axis has a top rung, and the sense in which the third generation is terminal on that axis without being terminal for intelligence generally: discrimination, speed and the cost of error keep improving forever. What kind of thing counts as input does not.

The misreading to disown

The weak and tempting version of this argument says the immune system is intelligent, or that machine systems should copy its wiring. Neither claim survives contact with what immunity actually does. There is no world model here, no causal reasoning, no goal beyond discriminating self from non-self at the molecular level. A memory B cell has no idea why a pathogen appeared, cannot generalise from one virus family to a novel one, and would fail instantly at any task requiring composition of concepts rather than matching of shapes. Reading affinity maturation as proto-cognition is a category error and inflates the biology to make the analogy more flattering than it deserves.

The narrow claim is the only one on offer. It is about intake discipline alone: how a system that cannot afford to pause acquires, dates, weights, and eventually retires evidence, under real threat of dying if it gets this wrong. On that axis, and only that axis, immunity is a fully worked existence proof.

Objections that hold weight

A memory B cell cannot tell you why a pathogen appeared. Calling this a model of the world confuses a lookup table of molecular shapes with belief.

This is correct, and it narrows the claim rather than merely qualifying it. Immunity has almost no compositional structure; it recognises epitopes, not situations, and it generalises poorly outside the shape space it has sampled. What the analogy borrows is not representational richness but intake discipline. Continuous intake, on its own, does not confer understanding. It confers the ability to stay correct as the world changes, which is a different and prior capacity — necessary, not sufficient.

Vertebrates later added trained innate immunity, epigenetic memory in monocytes, maternal antibody transfer, and microbiome-mediated discrimination. These look like new categories of evidence, not calibration.

Each of these is real and each deserves to be taken seriously, but examined closely each is a mechanism of use rather than a new class of intake. Trained innate immunity changes response speed to signals already encountered; it does not add a new sense. Maternal antibody transfer imports another organism's already-formed beliefs across the placenta — a transfer of provenance, not a new observation. Microbial tutoring of the immune system is intake mediated through a proxy sensor, the gut epithelium standing in for direct contact. The line the thesis needs — between new mechanisms and new evidence classes — holds through everything added since jawed vertebrates evolved the recombination machinery.

Autoimmunity, cytokine storm, and original antigenic sin show that continuous revisable updating can lock in wrong beliefs and attack the very system it protects.

This should be kept as a genuine warning rather than argued away. But it argues for governance layered on intake, not against intake itself. Thymic negative selection, guided by the AIRE transcription factor forcing medullary cells to display thousands of tissue-specific proteins nowhere else present, deletes most self-reactive T cells before they ever circulate; loss of AIRE produces the multi-organ autoimmune disease APS-1. Regulatory T cells and checkpoint receptors add further audit on top. The control problem was solved by adding trust mechanisms around open intake, not by closing the aperture back down.

What this does and does not establish

It establishes that continuous, provenance-bearing, revisable observation is a stable terminal architecture for a system that cannot afford to fall out of date, because a five-hundred-million-year lineage reached that architecture and then spent its remaining evolutionary effort on calibration rather than on finding a fourth kind of evidence. It does not establish that such a system is intelligent, that it understands what it observes, or that wider intake solves the problems intake creates — immunity's own pathologies show that plainly. The Large Universe Model, read against this lineage, is a claim about the shape of an intake problem that has been solved once already, not a promise that solving it delivers comprehension for free.