Memory and Compute in the Holographic Machine

A conceptual plan — where the bits live, where the arithmetic happens, and why those are not the same cavity.

draft v0.1 · 2026-07-22 · Role: this document connects two lanes that were built separately. The compute lane has a prototype (a twelve-port EML toroid, $10 BOM) and a measurement framework (the Osika and its four axes). The replicator lane has a boundary-controlled chamber and a field compiler. Both are holographic in the same technical sense. This plan says what happens when they are treated as one system — and it opens by answering the question that decides the architecture, which turns out to have a clean numerical answer rather than a matter of taste.

Note 2026-08-01: this plan predates the 07-30 cap+stack rulings — its 18.5 L two-cap-era chamber figures are instance-era numbers; nothing here is materially contradicted; to be refreshed at the next revision.

1 · The question that decides everything

The tempting idea is that the replicator's build volume, being a programmable holographic cavity already, should double as the machine's computer. It should not, and the reason is a mode count.

A cavity's information capacity scales as its volume divided by the cube of the wavelength it operates at — N ≈ 8πV/3λ³. The build volume is 18.5 L. Run that number for each carrier the machine actually uses:

carrier in the build volume	λ	independent modes
acoustic, 100 kHz in air	3.4 mm	3.8 × 10⁶
acoustic, ~1 MHz in melt	~5 mm	1.2 × 10⁶
microwave, 10 GHz	30 mm	5.7 × 10³
optical, 500 nm	0.0005 mm	1.2 × 10¹⁸

Twelve orders of magnitude. The same 18.5 litres holds a million-ish addressable modes for sound and a quintillion for light. The build chamber is a poor computer because it is a good replicator: moving matter needs long wavelengths and real forces, and long wavelengths are exactly what makes a cavity information-poor.

Note the ordering trap in that table: microwave has fewer modes than acoustic in the same box — 10 GHz is a longer wavelength than 100 kHz sound. The EM channel earns its place by speed, bandwidth and sensing precision, not capacity.

So the ruling is: do not merge the cavities. Build one architecture instantiated at three wavelengths, each sized for what it is good at, and let them share the mathematics and the compiler rather than the volume.

2 · One mathematics, three cavities

What justifies calling this one system is that the same operation runs in all three: interfere a reference wave with an object wave, store the interference, then re-illuminate to reconstruct. Memory, actuator and processor are that one operation wearing three hats.

cavity	λ / medium	role	status
Build chamber	acoustic + µwave, 18.5 L	the actuator: shape matter, scan objects. ~10⁶ modes is plenty for shaping and far too few for arithmetic	specified — construction paper
Compute cavity	optical, toroid, cm-scale	the processor: recurrent EML graph, settles to a fixed point in ns; depth set by Q	PROTOTYPED — $10 toroid
Store	optical, phase-change medium	the memory: non-volatile, rewritable, associative; ceiling (V/λ³)×bits-per-mode ≈ 10¹–10² TB/cm³, real limit set by M# (§3.1)	plan — this document

The compiler is what makes them one machine. The field compiler already emits boundary field states for the build chamber; the same synthesis step — record reference against object, modulate the surface — is what programs a holographic aperture, and what writes a page into the memory volume. One toolchain, three back-ends.

3 · Two layers, not four

An earlier draft of this plan proposed a four-level hierarchy with an SLM/LCD fast boundary and a write-once photopolymer archive on the ends. Both ends collapse, and the architecture is better for it:

The SLM was a slower version of something the cavity already does. The compute cavity's own field state is reconfigured at the drive rate — picoseconds to nanoseconds — while an SLM runs at 10²–10⁴ Hz: a kilohertz device in the path of a gigahertz cavity is a bottleneck called a layer.
The archive collapses into the store. PCM is already non-volatile with year-scale retention; a separate write-once medium buys permanence we already have, at the cost of a second write chain and a second material.

What remains is two layers and a bridge:

layer	medium	timescale	persistence	holds
Live state	the optical compute core — the field standing in the cavity	ps–ns, reconfigured at the drive rate	one coherence time	the working set: whatever is being computed, scanned or reconstructed right now
Store	phase-change material (GST, Sb₂S₃)	ns switching, finite endurance	non-volatile, no power	everything persistent: patterns, transforms, weights

The core is the bridge, and it runs in both directions. Every path in the machine passes through the live state: chamber → core (scan data in), core → PCM (consolidate), PCM → core (recall), core → chamber (drive out). Reading and writing are the same optical operation run with the reference and object waves exchanged, which is why one element can serve both.

Why PCM, and why the earlier deferral was still right. The photonic lane once considered PCM films as "frozen weights" and correctly set them aside: that architecture wanted every pixel reconfigured every frame, the one job a non-volatile medium is wrong for. That judgement was right — but it evaluated PCM for the fast job. Here the fast job belongs to the field itself, and PCM takes the one it was always suited to: the non-volatile, rewritable, high-index-contrast store — large amorphous↔crystalline index change, holds with no power, rewrites in place. Not a reversal of the old call — a different slot.

3.1 · Counting the store correctly

An earlier version of this section quoted ~200 MB/cm² per layer, derived from "one bit per (λ/2)² spot." That figure was wrong, and wrong in an instructive way: it is a binary optical disc model — one bit per resolvable spot on a surface — applied to a medium that does not work that way. A hologram does not put a bit in a place; it writes a distributed interference pattern across a volume, and many such patterns coexist there, separated by Bragg selectivity in angle, wavelength, shift or phase code. Capacity scales with thickness, and the correct accounting is mode counting, not spot counting.

accounting	modes or bits per cm³	note
one mode per λ³	8 × 10¹² modes → 1 TB/cm³ at 1 bit/mode	the conventional statement of the bound
one mode per (λ/2)³ (Nyquist cell)	6.4 × 10¹³ modes → 8 TB/cm³	the aggressive but defensible cell size
× 4 bits/mode (amplitude × phase × polarisation)	32 TB/cm³	the multidimensional-encoding direction (Optica 13, 591, 2026)
× 8 bits/mode	64 TB/cm³	optimistic on SNR

So the sugar-lump claim comes back, and larger than before: at 4 bits per mode, 1 cm³ holds ~32 TB ≈ 320 000 objects. Storage is emphatically not the constraint.

One precision that matters, so we do not overclaim in the other direction. V/λ³ is not the imaging diffraction limit, and holography does not evade it — holography is precisely the technique that reaches it, where bit-per-pixel storage wastes it. It is a mode count, the same count as §1's cavity modes, and linear optics cannot exceed it. What legitimately exceeds one bit per mode is a larger alphabet per mode: amplitude, phase and polarisation give several bits where intensity alone gives one. So the honest form is capacity ≈ (V/λ³) × bits-per-mode — insane in absolute terms, and still bounded.

The wall you actually hit is dynamic range, not λ³. When M holograms share a volume, the available index modulation divides among them, and diffraction efficiency per page goes as η = (M#/M)² — the medium's M-number over the page count, squared. With a good M# of 20: 100 pages gives η ≈ 4 × 10⁻²; 1 000 pages gives 4 × 10⁻⁴; 10 000 pages gives 4 × 10⁻⁶. Recall signal collapses quadratically in the very quantity we are trying to maximise. The λ³ ceiling is never what stops a real system — the detector noise floor is. Any capacity number in this plan is therefore a ceiling awaiting an M# measurement, and step 2 of §7 exists to get one.

Why PCM is a good holographic medium — a reversal of my earlier pessimism. Phase-change materials have enormous refractive-index contrast between amorphous and crystalline states — Δn of order 1, against ~0.03 for photopolymer. Since M# scales with available index modulation, that huge Δn buys multiplexing depth: more pages before η collapses — the opposite of a drawback, and a partial answer to the thickness problem, since a thin high-Δn medium can hold what a thick low-Δn one does. Two engineering conditions: use the low-loss compositions developed for photonics — Sb₂S₃ and Sb₂Se₃ — rather than GST, which is strongly absorbing when crystalline and gives amplitude modulation where you want phase; and genuine volumetric (thick or many-layer) PCM remains OPEN — deposited films are tens of nanometres.

3.2 · Chords, not bits — the figure of merit changes

Counting modes rather than spots changes not just the arithmetic but the unit. The medium's natural quantum is not a bit. It is a stored interference pattern — which is to say, a chord.

That matters because a hologram is natively associative: illuminate it with a probe and every stored page responds at once, each with its own correlation, in the time light takes to cross the volume — a bank of matched filters all firing in parallel. "What is at address n" and "what do you have that looks like this" are different questions, and the second is the one this machine actually asks.

Storing chords rather than bits deletes a whole layer. The .pattern format is already complex poles and port-vectors; the medium already stores complex interference patterns. Serialising chords into bits, writing bits into a medium that wanted patterns, then reconstructing chords on readout is a round trip through a representation neither end needed. Write the chords.

So the honest figure of merit for this store is not bytes but distinguishable, retrievable patterns per volume — limited by the same M# budget, because what degrades with load is precisely the ability to tell stored patterns apart. Capacity in bits is a sanity check on the object library; capacity in chords is what the architecture actually spends.

It also settles §5's second return. "Which stored chord does this echo match?" is not a search over a database — it is one illumination, answered by every page simultaneously. That is why a holographic correlator belongs in the Tbit/s scan loop and a lookup table does not.

Two honest limits on associative recall. First, it returns a ranked correlation, not an exact match — several pages answer, and the answer is "these, in this order, with these strengths." That is a proposal, which is exactly why the provenance rule of §6 binds it. Second, discrimination degrades with load on the same (M#/M)² curve: the more chords stored, the more alike the weak ones look. Associative capacity and recall fidelity are the same budget spent twice, and no encoding scheme escapes that.

4 · Weights are not a file — they are the medium

In this architecture, "a model need not be a large weight file" is literally true rather than a manner of speaking — worth stating precisely, because it is easy to overclaim.

A hologram performs a linear transform on whatever illuminates it: that is what a diffractive element is. A neural network layer is a matrix–vector product followed by a nonlinearity. So a weight matrix written as an interference pattern is not data that gets loaded into a processor — the light passing through it is the multiply, executed at the speed of propagation, with no memory bus in the path. This is established art: diffractive optical networks and programmable nanophotonic meshes both work this way.

The scale is the surprising part. At the 1 bit/λ³ density, a 1024 × 1024 weight matrix occupies a cube about 51 µm on a side — a speck at the edge of visibility. The "weight file" does not shrink; it stops being a file.

And plasticity follows from the medium rather than from a training loop: because PCM is rewritable in place, learning is re-exposure. A weight update is a local optical or electrical pulse that nudges the crystalline fraction of one site. That is the honest technical content of "neuroplasticity" here — not a metaphor, but also not free.

4.1 · The nonlinearity objection does not apply to this machine

An earlier draft of this plan asserted that "there is no good optical transistor; cascading layers without electronic regeneration remains the central unsolved problem of optical computing." That is a correct general statement about optical computing and the wrong statement about this architecture — it imports the machine-learning stack (linear layer, then an activation function bolted on) onto a machine that does not use it.

Two things make it not apply:

The primitive is already nonlinear. The compute cavity does not compose matrix multiplications; it composes eml(x, y) = exp(x) − ln(y), which is a nonlinear function of both arguments and is universal for the entire elementary-function family in the way NAND is universal for Boolean logic. There is no activation to add afterwards, because there is no linear layer to activate.
The device supplies it physically, and it is the same device that does the I/O. Per the supercomputer paper §5.1: the exponential branch is the gain-vs-carrier response of the laser cavity, the logarithmic branch is the carrier-vs-photon response of the same junction, and the subtraction is what the two coupled state variables do at threshold. The diode at threshold is, materially, a physical eml gate. So the architecture never needed an optical transistor; it needed a device whose native physics is the primitive, and a laser diode at threshold is one.

The correction changes what to worry about. The ceiling on this machine is not "can it be nonlinear" — it is "how deep a composition can it sustain."

So here are the limits that are actually binding.

Depth is Q-limited, and Q is currently small. Composition depth is how many EML applications complete before the field drops below the readout floor. The honest figures from the compute lane: optical ringdown Q ≈ 10–100 on the resin chamber, and effective compute-depth Q closer to 10. Ten EML gates deep is a real computer and a shallow one. This — not nonlinearity — is the frontier, and it is a materials frontier: CVD diamond supports ringdown Q > 10⁹.
Breadth and depth are different numbers; do not let one cover for the other. The operator space is astronomical — of order 6 × 10⁵⁷ distinguishable encodings select which EML tree the substrate evaluates. The composition depth is ~10. Both are true. Quoting the first while the second is the binding constraint would be the easiest available self-deception in this whole document.
Analog cascades accumulate noise — and the architecture's answer is settling, not regeneration. Digital restores signal levels at every gate, analog does not, so error compounds with depth. But a chamber that settles to a fixed point is running a contractive map, and contraction suppresses perturbations — signal restoration achieved dynamically rather than by a transistor. The tension to watch: the HAPTIC contract deliberately operates at the edge of lasing, the most sensitive available point, and sensitivity and contraction pull in opposite directions. Whether the loop sits contractive enough to restore while sensitive enough to read is an empirical question about the operating point.
Precision is finite. Analog optical/PCM compute delivers of order 4–8 effective bits per site, not fp32 — workable for inference and awkward for training.
PCM endurance is finite (order 10⁶–10⁹ cycles): fine for a store that updates rarely, disqualifying at activation rates. Precisely why it is the store and not the working layer.
In-situ training remains unsolved. Inference through a written transform is straightforward; computing gradients inside the medium is not. Assume transforms are fitted elsewhere and written here until shown otherwise.

And the two compute modes cover each other. This plan contains both stories: a passive holographic transform (§4 — genuinely linear, needing an activation from somewhere), and the EML cavity (§2 — nonlinear by construction). The generic objection applies to the first and not the second, and the machine has both — the diode stage is available as the nonlinearity for a holographic matmul that lacks one. The honest architecture is not "optics plus an unsolved problem." It is a linear transform in the medium and a nonlinear primitive at the device — the standard division of labour, just physical.

5 · What this gives the replicator

Three concrete returns, in increasing order of how much they matter:

A store for the pattern library. From the limit paper, an object costs ~100 MB; per §3.1, 1 cm³ holds ~32 TB ≈ 320 000 objects, rewritable in place. The number that will actually bind is not capacity but M#.
A solution to the Tbit/s problem. The machine's binding constraint is not building but listening: ~1 Tbit/s of sensor stream, with corrections due inside one wave period. A holographic correlator answers "which stored chord does this echo match?" in the time light crosses it — which is the shape of the scan problem, not a coincidence.
A brain the machine can build. Per the limit paper's §4.2: transistor logic is chemistry, optical computing is geometry, and geometry is what a replicator makes. The compute cavity and the memory volume are both shaped media — waveguides, index contrast, exposed volumes. They are the parts of a computer that a matter compiler can plausibly close on.

6 · The provenance rule binds this lane too

This plan inherits the scan lane's discipline: every chord, and every reconstruction it feeds, carries measured | inferred provenance, and the viewer renders the difference. The reason is specific to this machine: the scan output is also the build acceptance criterion, so an inferred interior becomes a fabricated-from-a-guess interior. Both a wrong Green's function and a confident prior produce sharp, convincing, meaningless images.

That rule bites here because the returns in §5 are inference. A holographic correlator that answers "which stored chord does this echo match?" is proposing a match — it is not measuring one. So:

The model may propose; only the aperture asserts. Anything the compute core recalls, completes, denoises or pattern-matches is tagged inferred and stays tagged through every downstream use. The optical core is allowed to be fast and clever; it is not allowed to launder a guess into a measurement by being either.

Two corollaries worth writing down now, before there is code to argue with:

Adaptive selection allocates; it does not synthesize. Spending more chords where an object rings complex is a policy over a finite measurement budget — genuinely valuable, and the reason resolution should be spatially non-uniform, but no independent information. A learned scan policy chooses among available measurements; it cannot create one.
Weights are inferred by construction. A transform written into PCM is a prior. When the core applies it, the output inherits inferred — which means the "weights are the medium" claim of §4 buys speed, not epistemic standing.

The one thing that makes multi-bounce reconstruction legitimate is also what makes this lane legitimate: the machine measures its own cavity. The empty-chamber reference run, the per-plate TX→RX identity matrix and the double-precision twin gates are not housekeeping — they are what turns wall multipath from a liability into aperture, because the forward model is measured rather than assumed. The same standard applies to any hologram this plan writes: a stored transform is trustworthy only to the extent the medium's response was characterised, not asserted.

7 · The plan, in order

#	step	why it is first	gate
1	Write and read one hologram in a phase-change film with the existing 450 nm optics; recover a known bit pattern	establishes the write/read chain before any capacity claim	bit-error rate on a known page
2	Measure the medium's M# — multiplex N pages, plot η vs N, fit `(M#/M)²`	this, not λ³, is what limits a real store; every capacity claim in §3.1 is a ceiling awaiting this number	M# with its η-vs-N curve; capacity restated as a fraction of the mode ceiling
3	One matrix–vector product through a written hologram; compare against the digital reference	this is the whole "weights are the medium" claim, reduced to one falsifiable step	relative error vs digital `Wx`
4	PCM rewrite cycle: write a transform, run it, rewrite it, run again	separates "non-volatile store" from "reprogrammable transform" — the plasticity claim lives here	contrast retained over N cycles
5	Correlator against the chord dictionary — feed a scan echo, get a match	the first step where compute serves the replicator rather than standing alone	match rate vs the software matcher
6	Score on the four axes (Osika), reporting all three energy denominators	keeps this lane commensurable with the toroid and with digital baselines	a filled axis table with named verification type

Steps 1–3 are a bench sequence, not a program. Nothing after step 3 should be argued for until step 3 produces a number.

8 · Open questions worth naming

Does the build chamber compute at all? ~10⁶ acoustic modes is a poor processor but not zero. The chamber's modes are spent on shaping; asking them to also compute is asking one resource to do two jobs. Whether any useful computation rides free on the shaping is OPEN and would be a genuine finding — but it is not a plan, and it must not be assumed in the budget.
Multiplex depth in practice. The λ³ bound counts sites; realized holographic systems lose an order or more to multiplexing crosstalk and dynamic range. The recent multidimensional-encoding work (amplitude × phase × polarisation, Optica 13, 591, 2026) is the right direction — more degrees of freedom per site — but is research-stage.
Where the nonlinearity comes from. Unresolved here and everywhere. Candidates worth tracking: saturable absorption, the diode-bridge trick the toroid already exploits (the laser diode is the nonlinearity in that architecture), and simply accepting an electronic activation between optical layers.
Self-fabrication of the media. Photopolymer and doped glass are bulk phases and plausibly in-machine; PCM films with controlled stoichiometry are less obviously so. This sets where the von Neumann sea-of-parts line actually falls for the memory layer.

Register. This is a plan, not a result. The toroid prototype and its bench measurement are MEASURED and belong to the supercomputer paper with its own caveats (optical-ringdown Q ≈ 10–100, effective compute-depth Q closer to 10). Everything in §3–§6 here is design intent with numbered gates; nothing in it has been built. The density figures are bounds and arithmetic, and the arithmetic is checkable in a minute. Where this plan touches AI claims it claims two separable things — a linear transform held in the medium, and a nonlinear primitive supplied physically by the diode (§4.1) — and explicitly disclaims in-situ training, fp32-class precision, and any composition depth beyond what measured Q supports.