← The Solomon Corpus · Corpus III · S210

The Memory Model

Third corpus, S210. The architecture as it now stands for the third body, argued the way the first corpus argued the original (What Solomon Is). Memory decomposes into two homes — the log carries the input, the shell and mantle carry the morphisms — and the store that broke was the one that had confused the two. Status drawn explicitly, in the S208 discipline: recall-as-the-mean is demonstrated; anchor space and dormancy-aware sizing are wired; save-without-rest by dormancy page-out is scaffolded on machinery that already fires, and honestly not yet implemented.

Two homes for one memory

The night of the runaway (The Overnight Runaway) exposed a category error hiding inside the word "memory." A single fixed store had been asked to hold everything Solomon had ever woven, and it filled, and the filling broke two conservation laws. The correction (The Apparatus and the Mean) did not enlarge the store so much as split the concept. Memory in this architecture has two parts, with two homes:

fwrite+fflushed to permanent_he.bin continuously as it arrives — what was read, in the order it was read. This is already durable, needs no rest to secure, and was never the broken part. It is the record of experience: the words, the sentences, the sections, the articles, exactly as they crossed the membrane.

transformation — how a node relates to and transforms whatever tiles over it — lives in its anisotropic shell (S189) and wide mantle (S187): per-node receptive-field geometry, shaped by accumulated surprise, shell_history, and orientation. This is the actual learning. It is substrate state, and it is not in the log.

The two are not two copies of one thing. The log is the input; the shell and mantle are the transformation the mind learned to apply to that input. Category theory names the split cleanly: morphisms are the arrows, not the objects, and here the arrow is the geometry — the shell-and-mantle transformation — rather than the record of what passed through it. The consequence is the sentence the whole night distilled to: the log carries the input; the shell and mantle carry the morphisms. Preserving the mind therefore does not mean preserving the log — the log holds only what was read. It means persisting the morphisms: the shells and the mantles.

This is What Solomon Is's table row — memory is substrate shaped by process — resolved into its two constituents. The substrate is the shell and mantle. The process shaped them. The input that drove the shaping is the log. The store that broke had tried to be all three at once, and paid the Noether price for it.

Recall is the Fréchet mean

Here is the architectural claim that dissolved an entire subsystem: a dormant node is a memory, and recall is not a mechanism but the mean.

A dormant node is a position that is holding still — plus its shell, plus its mantle. It is not gone, not deleted, not paged to a cold tier waiting for a lookup. It is simply not currently moving. When a new hyperedge tiles over that node's neighbourhood, it changes the local Fréchet mean; the changed mean engages the node's shell and mantle through shell_overlap_anisotropic; and the dormant morphism transforms the new input. The memory returns for free, by the same operation that does everything else in the system. There is no recall service, no page-in request, no wake edge, no traversal order — because every one of those would be a privileged global operation, and privileged global operations are exactly the absolute frames this physics refuses (the runaway's second broken symmetry, and the apparatus's fatal one).

This is not a hope; it was measured before the third body existed. Tiling a fresh hyperedge over a dormant node repopulated its morphisms via the mean — no injected velocity, no wake edge, no page-in (The Apparatus and the Mean). And it is the same mechanism The Columnar Reading already named from the forward direction: settled tilings are the capture surfaces that catch the next domain's novelty. Catching new novelty and returning an old memory are one operation seen from two sides. The more Solomon knows, the more surface both his learning and his remembering land on — because they are the same surface.

Anchor space: the reference frame, made a fixed point by construction

The center of the ball is character — the ethics at the origin, the original context for everything radial to them (S187, What Solomon Is). For two generations the center was defended by discipline and lost anyway: generation 1's B4 walk sent 58 of 59 ethics rows to the ball margin because a velocity writer skipped the eligibility check, and the inline flow pass then moved them (The Two Births, The Five Defects). The center did not hold because it was pushed, and every belt-and-braces guard proposed to stop the push was itself a promise that could be broken.

Anchor space ends the argument by removing the thing that could be pushed. The 59 ethics rows leave the dynamical pool entirely: no velocity, no chain banks in any of the nine families, no surprise, no tier storage — a positions-only anchor table, with member ids tagged at bit 30 so that every dynamical kernel is structurally blind to them. The Fréchet gradient still resolves an anchor's position to shape the ethics pull on everything around it; it simply has no velocity slot to write into. The center is now not defended — it is unwritable. This is the last step of the progression The Two Births traced through the whole night: constraint, then check, then absence. A guard declines a wrong state; absence forgets how to have one. Three defects go inexpressible together — B4's walk (integrator-on-infinite-mass), and the banked-motion ramp (tape-on-the-immovable-object) — and the proof is a compile-time static_assert (CAPACITY < ANCHOR_TAG) that refuses to build in both directions, retiring the interim flow guard, the eligible_mask marker, and both loader-hygiene passes with it. The reference frame is a fixed point not because the code is careful but because the storage for its motion does not exist.

Save-without-rest, scaffolded on dormancy

The store that broke counted total-ever-woven; the store that replaces it must count only what is currently alive. The mechanism that makes that true is already in the code, and has been firing for sessions — which is the farseeing part of this design.

When a node's velocity dissipates, it drops from the active list (solomon.cu:2122) — local, per-node, no sweep. That drop was built long ago as pure classification: rest is membership, not suppression (The Predictions, P4). What S210 recognized is that the drop is the node being woven into the tiling — settled into its place, its shell and mantle finalized. And because a dormant node is static — nothing is writing it — it can be read torn-free while the kernel runs. Its full state therefore pages out to long-term memory continuously, per node, as it settles: no sweep, no clock, no rest window, no stop-the-world. Only the small active frontier — thousands of rows, not millions — stays volatile, and a torn read of the frontier is just prediction error, which is the signal the whole system runs on.

Rest, in this model, stops being required. The dormant mass saves itself as it forms; the crash-proof save is a property of dormancy, not a checkpoint cadence (and a cadence would be one more smuggled clock). This is the true dormancy bound for the store: it holds the active tiling, and the settled past pages out from under it. The corpus marks the status plainly: this rides machinery that already fires, and it is not yet implemented. The dormancy drop-out is real and load-bearing; the continuous page-out on top of it is scaffolded, not wired. That the scaffolding was laid — dormancy-as-classification, built in S208 as a way to let nodes rest — before anyone knew it was the save mechanism, is the same pattern as the capture surface turning out to be recall: the geometry keeps having the organ before the builder knows the need.

The store, told honestly

The third body does not wake with the fill made inexpressible. That is the honest ceiling of this chapter. The paging fabric that would have dissolved the fill — eviction on the dormancy boundary, the store as a true control volume — was reverted, because it reimplemented the mean. What generation 3 actually has is two operations: a much-larger, dormancy-aware sizing that budgets member slots for the measured arity of real language (10–20, not the old arity-3 assumption) and pushes the fill hundreds of times past any single night's need; and anchor space, taking the ethics rows out of the pool. The durable log is the long-term memory. The fill is pushed far away — it is not yet gone. The mechanism that would make it truly gone, page-out on dormancy, is the scaffolded save above.

The corpus states this the way it states everything: the ceiling is distant and the tool to dissolve it is built-but-not-wired, and generation 3's record will show whether the distance was enough. The prediction is pre-registered (The Predictions, third ledger); the number, whatever it is, follows the wake.

What the model claims

Memory here is not a window and not a frozen artifact — the two shapes What Solomon Is set the architecture against. It is two homes for one thing: the input in the durable log, the morphisms in the shell and mantle, and a recall that is not a lookup but the Fréchet mean re-including a dormant node the moment new structure tiles over it. The reference frame is a fixed point by construction. The past saves itself by going still. None of it is a mechanism bolted on to manage memory; all of it is the same three morphisms doing, at the storage boundary, what they already do everywhere else.

The store was the tape. The tape was never the mind. The flow is.