Marin Ivezic
Every quantum networking roadmap you will read contains a component that has no classical equivalent and no obvious reason to exist. It is called a quantum memory, and the name is misleading. It doesn’t store data in any sense a network architect would recognise. It stores one thing, briefly, and then either hands it back or loses it.
Understanding why that component is unavoidable is the fastest route into understanding quantum networking as a whole. The argument runs through one physical constraint, and once you have that constraint the rest of the architecture follows almost mechanically: why repeaters cannot repeat, why distance is an exponential problem rather than a linear one, why the specification sheet for a memory has two different time constants on it, and why a vendor who quotes only one of them is telling you half the story.
The one property that changes everything
Start with the unit. A qubit is a two-state quantum system – the spin of an electron, the polarisation of a photon, an atom that is either in its ground state or an excited one. The classical bit it resembles takes the value 0 or 1. The qubit can be in a combination of both at once, a condition called superposition, and it stays in that combination until something measures it. Measurement forces a choice: you get 0 or you get 1, with probabilities set by the combination, and the combination itself is gone.
That last sentence is the whole difficulty. In a classical network, reading a signal is free. A repeater on a long fibre run reads the incoming bits, decides whether each was a 0 or a 1, and transmits fresh clean bits onward. The read is non-destructive in every way that matters, and it is what makes distance a solvable engineering problem: chain enough repeaters and you can go as far as you like.
Reading a qubit destroys it. Not damages it – destroys the information you were carrying. If Alice sends a qubit in superposition and a repeater in the middle measures it to decide what to send onward, the repeater learns one bit and the superposition is gone. What arrives at Bob is a classical bit that Alice never sent.
So the first architectural fact about quantum networks is that the standard tool for beating distance is unavailable.
Why a repeater cannot repeat
You might reasonably ask why the repeater can’t copy the qubit instead of measuring it, and pass the copy along while keeping the original.
It can’t, and this isn’t an engineering limitation waiting on better hardware. The no-cloning theorem, proved in 1982, says that no physical process can take an unknown quantum state and produce two identical copies of it. The proof is short and follows from the linearity of quantum mechanics. There is no clever device, no future breakthrough, no sufficiently cold apparatus that gets around it. If you know what state you are making, you can make as many as you like. If the state is unknown – which is the entire point of transmitting it – you get one.
Amplification is the same problem wearing a different hat. An optical amplifier in a classical link works by stimulated emission, turning one photon into many. Applied to a single photon carrying quantum information, the process adds noise in exactly the amount required to prevent the output from being a faithful copy. The physics is defending itself.
Which leaves direct transmission. Photons in standard telecom fibre lose intensity steadily with distance; over roughly 100 km, only about one percent of what you sent arrives. That’s fine for classical signalling, where you have billions of photons and only need some of them. For single photons carrying single qubits, a one percent survival rate means ninety-nine attempts out of a hundred produce nothing at all, and the rate keeps halving as you extend the link. Push to 500 km and single-photon transmission through standard fibre becomes vanishingly rare, so even at a high attempt rate a link can run for long stretches with nothing to show for it. Distance in a quantum network is not a linear cost. It’s an exponential one.
This is where entanglement enters, and where the memory becomes necessary.
What the memory actually does
Entanglement is a correlation between two particles that is stronger than any classical description of them separately can account for. Prepare two photons in an entangled pair, send one to Alice and one to Bob, and measurements on the two will show correlations that no shared list of pre-agreed values reproduces. That correlation is the resource a quantum network distributes. It’s what quantum key distribution consumes, what teleportation consumes, and what a distributed quantum computation consumes.
Entanglement can be extended across distance by a trick that has no classical analogue. Suppose you can’t entangle Alice and Charlie directly because they are 200 km apart. Put Bob in the middle. Entangle Alice with Bob over 100 km. Separately, entangle Bob with Charlie over the other 100 km. Bob now holds two qubits: one entangled with Alice, one entangled with Charlie. He performs a Bell state measurement on his pair – a joint measurement that asks which of four possible correlations his two qubits share, without asking what either qubit individually is. That measurement destroys Bob’s two qubits and leaves Alice and Charlie entangled with each other, over 200 km, having never exchanged a photon.
This is entanglement swapping, and it is the operating principle of every quantum repeater design.
Entanglement generation over each 100 km segment is probabilistic, and that is what makes the memory necessary. Most attempts fail, because most photons are lost. Alice-Bob might succeed on the fortieth attempt. Bob-Charlie might succeed on the two-hundredth. Bob can only perform the swap if he holds both entangled qubits at the same moment.
Without somewhere to put the first success, he doesn’t. The Alice-Bob entanglement arrives, exists for the few microseconds it takes a photon to pass through, and is gone long before Bob-Charlie succeeds. Bob spends eternity holding one half of a pair and waiting for a partner that will not arrive in time. The two-segment link performs no better than one long segment, and the whole scheme collapses back into the exponential loss it was meant to defeat.
A quantum memory is the component that holds Bob’s first success while he keeps trying for the second. That’s its entire job. It converts a requirement for two simultaneous coincidences into a requirement for two sequential ones, and that single change is what turns an exponentially hard problem into a merely difficult one.
Think of it as a buffer, with one crucial difference from every buffer you have worked with: this one can’t look at what it’s holding. The moment it reads the qubit to check on it, the qubit is worthless. A quantum memory has to store something it is forbidden to inspect, hand it back unchanged, and be trusted to have done so.
The two numbers on the datasheet
Because the memory cannot check its contents, its quality is described statistically, by how long the stored state survives before the environment ruins it.
Two time constants matter, and they measure different failures.
T1, the relaxation time, is how long an excited state survives before decaying to the ground state on its own. A qubit stored as “atom is in the excited state” eventually is not, because excited states emit photons and drop. T1 governs the loss of the qubit’s energy, and with it, the distinction between the two basis states.
T2, the coherence time, is how long the relationship between the two components of a superposition survives. A qubit in superposition is not just “partly 0 and partly 1” – it is 0 and 1 in a specific phase relationship, and that phase is where the quantum information lives. Random nudges from the environment – a stray magnetic field, a lattice vibration, a passing thermal photon – scramble the phase without necessarily changing the energy. This is decoherence: the state becomes a boring classical mixture, and every quantum advantage evaporates.
T2 can never exceed twice T1, because a qubit that has decayed has certainly lost its phase. In most real hardware T2 is considerably shorter than T1, sometimes by an order of magnitude, because dephasing noise is easier to come by than energy decay.
For an architect, the operational reading is simple. T2 is the number that bounds your protocol. Fidelity of a stored state falls off exponentially with the ratio of storage time to T2, which means you want your storage time to be a small fraction of T2, not a comparable one. Store for a full T2 and you haven’t stored anything usable. Store for a tenth of it and you are in reasonable shape.
Work the numbers backwards from your topology. A 100 km segment means a photon flight time of roughly half a millisecond, and each attempt at entanglement needs at least one round trip for the classical signal confirming success. Hundreds of failed attempts before a success is normal. That puts the required storage time in the tens to hundreds of milliseconds for a modest link, and considerably longer for a chain of many segments where the memory at one node waits on successes at several others.
This is why “our memory achieves a coherence time of X” is only half a claim. The other half is what X has to be for the network you want to build.
The number vendors mention last
Coherence time gets the headlines. Retrieval efficiency decides whether the device is useful.
Retrieval efficiency is the probability that a photon you send into the memory is absorbed, stored, and re-emitted on demand. If it is 10 percent, then nine times out of ten the qubit you were trying to save is simply lost – and you’ve added a lossy component to a network already fighting loss. A repeater node with a 10 percent memory can easily be worse than no repeater at all, because you’ve introduced a new failure point in exchange for a synchronisation capability you rarely get to use.
Retrieval efficiency has improved steadily across platforms, and ensemble memories lead most of the field on it. But a memory that holds a state for an hour at three percent efficiency is a physics result, not a network component. When you read a memory specification, three numbers travel together and none of them means much alone:
- Coherence time, which bounds how long you can wait
- Retrieval efficiency, which sets how often the wait was worth it
- Fidelity, which describes how faithfully the returned state matches what went in
A vendor quoting one of the three is quoting the one that flatters the device.
Storing more than one thing at a time
There is a fourth property that changes the arithmetic of a whole link: multiplexing, meaning the ability of one memory device to hold several distinct qubits simultaneously, in separate slots distinguished by arrival time, by frequency, or by spatial mode.
The reason this matters is a probability argument. If each entanglement attempt succeeds with probability p and you can only run one attempt at a time, you wait 1/p attempts on average. If the memory can hold results from N parallel attempts, your effective success rate rises accordingly, and the required coherence time drops because you are not waiting as long. Multiplexing buys back coherence time by spending hardware complexity.
Some platforms multiplex naturally. A crystal doped with rare-earth ions contains a very large number of absorbers, and by shaping the absorption spectrum into a comb of frequencies you can park many photons in one piece of material and recall them separately. Others do not. A single trapped ion holds one qubit and that’s that.
For an architect comparing designs, multiplexing depth belongs next to coherence time in the comparison, because the two trade against each other in setting the link rate.
The interface problem
Everything so far assumed you can get a qubit into the memory and back out. That turns out to be the hardest engineering problem in the field, and it is worth understanding why, because it explains most of the platform choices you will encounter.
Quantum information travels on photons. Call these flying qubits: fast, good at covering distance, impossible to hold still. It gets stored in matter – atoms, ions, spins in a crystal lattice. Call these stationary qubits: they stay put, they can be manipulated with lasers and microwaves, and they go nowhere on their own.
A quantum memory is a device that converts one into the other, in both directions, without disturbing the state. The photon has to be absorbed in a way that maps its quantum state onto an excitation in the material, held there, and then released as a photon carrying the same state. Every step leaks.
Wavelength makes it worse. Telecom fibre is engineered around 1550 nm, where attenuation is lowest, and the global installed base of fibre is optimised for that window. Most good memory materials do not absorb at 1550 nm. They absorb in the visible or near-infrared, where fibre loss is several times higher. So a practical node needs frequency conversion – a nonlinear optical process that shifts a photon’s wavelength while preserving its quantum state. Conversion works, and it costs efficiency at both ends.
That’s the constraint shaping platform selection: not which material has the longest coherence time, but which combination of material, cavity, and conversion stage delivers acceptable numbers on all four properties simultaneously.
What the platforms actually offer
There’s no winner. There is a set of trades, and each family sits somewhere different on them.
Trapped ions are individual atoms, stripped of an electron and held in a vacuum by electromagnetic fields. Their coherence is exceptional, running to minute-scale storage in well-isolated ion qubits, achieved through extreme isolation and pulse sequences that cancel slow noise. They also come with mature control techniques borrowed from ion-trap quantum computing, so gate fidelities are high. The difficulty is photons: ions emit at inconvenient wavelengths and getting a single photon reliably into or out of a single ion generally requires an optical cavity built around it. Systems are large, slow, and difficult to replicate.
Atomic ensembles store a photon not in one atom but as a collective excitation shared across a cloud of them – typically rubidium or caesium, either laser-cooled in a trap or as a warm vapour in a glass cell. Because the interaction is collective, coupling light in and out is comparatively easy, and multiplexing comes fairly naturally. Storage times run from microseconds in warm vapour up to the second range in cold, well-controlled setups. Warm vapour cells are the strongest current candidate for something that works outside a specialist laboratory, since they need no cryogenics.
Colour centres in diamond are point defects in the crystal lattice – a nitrogen atom next to a vacancy, or a silicon atom in a similar arrangement. The defect’s electron spin can be initialised, manipulated with microwaves, and read out optically, and a nearby nuclear spin can serve as a much longer-lived storage register alongside it. This split is powerful: the electron spin talks to photons quickly, and once entanglement is established the state is transferred to the nuclear spin, which holds it for far longer than the electron could. Because these are solid-state devices, they can be fabricated with photonic structures around them, which makes them the most credible route to something manufacturable. Nanophotonic diamond memory modules have been operated over fibre links, including spooled and campus-scale runs rather than benchtop optics alone.
Rare-earth doped crystals dope a host crystal with ions such as europium, praseodymium or erbium, whose optical and spin transitions are unusually narrow and long-lived. This family holds the storage-time records – hours of spin coherence has been demonstrated in europium-doped material under carefully chosen magnetic conditions – and it multiplexes well. Erbium is particularly interesting because it absorbs near 1550 nm, sidestepping frequency conversion. The cost is cryogenics: these devices generally run at a few kelvin or below.
Quantum dots are engineered semiconductor structures that behave like artificial atoms. They emit single photons very fast and integrate directly into semiconductor fabrication, which makes them attractive as sources. As memories they are weaker, with spin coherence typically in the microsecond range because of interactions with nuclear spins in the surrounding lattice.
Long coherence tends to come with hard photon coupling. Easy photon coupling tends to come with short coherence. Convenient wavelengths tend to come with cryogenics.
Which is why the most interesting designs stop trying to find one material that does everything. A dual-species ion node uses one ion species chosen for its photon-emission properties to establish entanglement, then transfers the state to a second species chosen for its coherence, sitting in the same trap. The diamond electron-plus-nuclear-spin arrangement is the same idea inside one defect. The architectural pattern is consistent enough to name: a communication qubit optimised for talking to the network, coupled to a memory qubit optimised for holding still. Expect to see it in almost every serious node design.
What this changes for a security architect
Quantum key distribution, or QKD, is the application most architects encounter first: two parties generate shared random keys using quantum signals, with the physics guaranteeing that eavesdropping disturbs the signal detectably.
Deployed QKD today is limited to roughly a hundred kilometres of fibre per link. To go further, operators use trusted nodes – intermediate stations that terminate one quantum link, decrypt the key material, and re-encrypt it onto the next. It works, and it moves the security boundary: every trusted node is a point where key material exists in classical form and must be physically and operationally protected. For a bank linking two data centres in one metropolitan area, that may be acceptable. For a link crossing jurisdictions, it usually isn’t.
Quantum repeaters built on memories remove that compromise. Entanglement swapping never exposes the key, because the intermediate node’s measurement destroys its own qubits and reveals nothing about the correlation it establishes between the endpoints. A repeater node cannot be compromised for key material in the way a trusted node can. It can be attacked for availability – an adversary who disrupts a node prevents entanglement distribution – but not for confidentiality.
That’s the real security argument for quantum memory, and it is narrower than the one usually made. Memories do not make QKD longer-range as a matter of convenience. They change what has to be trusted.
Two corrections are worth making explicitly, because the source literature blurs both.
First, none of this affects your post-quantum cryptography migration. QKD distributes symmetric keys between two endpoints over a purpose-built quantum link. It does not authenticate parties, does not produce digital signatures, does not protect data at rest, and does not work over the general internet. The algorithms standardised by NIST – ML-KEM for key encapsulation, ML-DSA and SLH-DSA for signatures – address a different problem, and no quantum network deployment substitutes for that work.
Second, quantum repeaters are not deployed infrastructure. Multi-node entanglement-based networks have been demonstrated across a handful of laboratory nodes, and memory modules have been operated over fibre links rather than benchtop optics alone. Neither is a product. Timelines for operational quantum repeater networks are genuinely uncertain and any vendor offering you one now is selling something else.
Four questions worth asking
If quantum networking appears in a procurement document or a technology roadmap on your desk, four questions separate a serious proposal from an aspirational one.
What is the coherence time, and what does the topology require? The second half is the one that gets skipped. A memory holding state for ten milliseconds is impressive hardware and useless in a chain of six 80 km segments. Ask the proposer to show the arithmetic connecting the device to the link.
What is the retrieval efficiency, at the wavelength you will actually use? Efficiency quoted at the memory’s native wavelength, on a device that will need frequency conversion in deployment, is a laboratory number. Ask for the end-to-end figure.
How many modes does it store? Multiplexing depth sets the link rate at least as much as coherence time does, and it’s the property most often omitted.
What is the operating environment? Millikelvin, a few kelvin, or room temperature is the difference between a physics result, a specialist installation, and something that could sit in a telecom exchange. Cryogenic requirements dictate the cost, footprint and maintenance profile of every node in the network.
None of these questions requires a physics background to ask, and the pattern of answers tells you quickly whether you are looking at engineering or at a research programme with a sales team attached.
The shape of the problem
Quantum networking gets described as a distance problem, and that framing is close enough to be misleading. Distance is the symptom. The cause is that the one operation classical networks rely on to beat distance – read the signal, then send a fresh copy – is physically prohibited for quantum information.
Everything downstream follows from that prohibition. Entanglement swapping exists because copying does not. Quantum memory exists because swapping needs two probabilistic events to coincide and they won’t coincide on their own. The four-way trade between coherence, efficiency, multiplexing and operating temperature exists because holding a quantum state still and coupling it to a travelling photon are opposing engineering goals.
An architect who holds that chain of reasoning can read almost any quantum networking claim and place it correctly, without following the atomic physics underneath. The question to bring to any such claim is always the same one: what is being stored, for how long, at what cost in loss, and does that match the network being proposed?
Quantum memory sits at the centre of quantum network architecture, and the reasoning above is the entry point rather than the full picture. Repeater protocols, entanglement purification, network routing under probabilistic link generation, and the topology decisions that follow from real hardware parameters are covered in depth in the quantum networking programs at Quantum Academy, which are built for architects and engineers who need to evaluate these systems rather than build them.
For the cryptographic migration work that runs in parallel with, and independently of, any quantum networking deployment, the methodology at pqcframework.org is the better starting point. Deeper technical treatment of quantum networking hardware is available at PostQuantum.com.