Quantum memories

What gets measured

A memory's quality is not a single number. Four metrics matter for network use, and a platform that wins on one usually loses on another. The platform table below quotes a representative storage time and a TRL for each row; the rest of this section is the vocabulary needed to read it ^{Ezratty 2025} .

• Storage time — how long the stored state retains its quantum character. Two underlying constants drive it. $T_1$ is the energy-relaxation time (how long the excited state stays excited); $T_2$ is the dephasing time (how long the relative phase between $|0⟩$ and $|1⟩$ survives). Coherent storage of a qubit — not just a classical bit — is bounded by $T_2$, the smaller of the two, and in practice $T_2\le 2T_1$. The published "storage time" is typically the time at which retrieved-state fidelity has dropped to a quoted threshold.
• Retrieval efficiency — the probability that a stored excitation comes back out as a photon when the readout pulse fires. A memory that holds for a long time but only retrieves 5 % of the time is a different engineering proposition from one that holds briefly but retrieves with 80 % efficiency. The Sangouard RMP review treats efficiency and time together because they trade off through the storage protocol's optical depth and bandwidth ^{Sangouard et al. 2011} .
• Retrieval fidelity — how close the retrieved state is to the one that was written in. Storage that returns a maximally mixed state on cue is useless; the memory has to preserve the qubit's relative phase, not just the populations.
• Multimode capacity — how many qubits the memory can hold in parallel. Multimode storage is what makes the entanglement-distribution rate of a long chain workable: a single-mode memory has to sit idle between attempts, while an N-mode memory can absorb N parallel attempts and let the classical heralding signal pick out which one succeeded ^{Sangouard et al. 2011} .

Per-platform landscape

Seven platforms cover essentially all current memory work. Numbers in the table are the best published demonstrations cited in the surveys below, not Pareto- optimal joint operating points: the same memory does not simultaneously hold for an hour, retrieve at 80 % efficiency, store thousands of modes, and operate at telecom wavelengths. Reading the table requires holding all four metrics from the previous section in mind together. Storage-time and TRL figures track Meddeb 2025 Table 5 ^{Meddeb 2025} ; per-platform context follows Sangouard RMP 2011 for atomic ensembles and Ezratty UQT Part 2 for the surrounding hardware survey ^{Ezratty 2025} .

• Atomic ensembles. Few ms in DLCZ-type demos; up to 1 h in atomic-frequency-comb storage.
• Solid-state defects. Communication electron spin holds for ~100 µs; nearby ¹³C nuclear-spin register has been demonstrated past 1 s.
• Trapped ions. Hyperfine-encoded qubits in a clock state. Hour-scale single-ion coherence has been reported at the qubit level; network-relevant memory experiments sit nearer the 10-s mark.
• Neutral atoms. Network-memory experiments in this platform are newer than the survey cycle; covered alongside atomic ensembles where the DLCZ-style read/write applies.
• Semiconductor quantum dots. The 4-day figure is a write-and-erase charge memory; network-relevant spin-photon memories in dots sit at ms–s.
• Photonic delay-line. Storage at telecom wavelengths in single-mode fibre, no transduction needed. The cited number is short, and the platform inherits the link's own loss-per-km — every µs of storage is ~200 m of fibre.
• Superconducting. Random-Access Quantum Memory: a superconducting qubit is parked in a long-lived microwave resonator. Couples natively to superconducting processors but needs a microwave→optical transducer to sit on a fibre network.

Two demonstrations anchor the solid-state-defect row to network deployment. Pompili et al. (2021) ran a three-node NV-centre network using on-chip nuclear- spin memories to hold the first link's Bell pair while the second link was generated ^{Pompili et al. 2021} . Knaut et al. (2024) ran a two-node SiV-centre memory experiment over 35 km of deployed Boston-area telecom fibre, with a quantum-frequency-conversion stage carrying the SiV's 737 nm photons to telecom and back ^{Knaut et al. 2024} . Both experiments treat the memory as the network's scheduling primitive — what makes asynchronous heralding actually compose.

Where memories sit on the spectrum

Each memory speaks a fixed wavelength of light — set by the atomic transition or defect level it stores into. Most memories speak something other than the telecom C-band: rare-earth-doped crystals at 580–795 nm, NV at 637 nm, SiV at 737 nm, ion-trap UV–visible lines, quantum dots in the 900–1100 nm range. To travel along deployed fibre, the photon those memories absorb or emit has to first be converted to the C-band by a quantum frequency converter (QFC); for superconducting memories the conversion is microwave-to-optical and is the field's hardest open transducer. The spectrum page places all of these on a common log-frequency axis with the C-band marked as the convergence target; the transduction page covers the conversion devices themselves.

TRL benchmark

Meddeb's 2025 Computer Networks survey rates each memory technology on the standard TRL 1–9 scale, alongside its current and decade-ago storage time. The summary, condensed from Table 5 ^{Meddeb 2025} :

Two patterns stand out. Atomic ensembles and trapped ions are the most mature memory platforms on the survey, with TRLs in the 6–8 range and storage times long enough to absorb realistic link latency. Solid-state, quantum-dot, and photonic memories sit in the 5–6 range — viable but pre-deployment. The superconducting row is a special case: 34 ms is enough to hold a microwave qubit for a long-distance fibre link's worth of latency in principle, but the missing microwave-to-optical transducer means it cannot today be the memory in a fibre quantum network.

The maturity page sets these per-memory TRLs against the per-capability QTRL framework Purohit et al. propose, so a single platform's TRL for "memory" can differ from its TRL for "compute" or "source".

Network role

The memory is what makes asynchronous protocols composable. Three patterns from the surrounding subjects use it directly:

• Memory-based repeaters chain Bell-pair generation with entanglement swapping. Each segment heralds independently; the memory at each node holds the segment-A half of a Bell pair while the segment-B attempt cycles. Without memory, both segments would have to herald in the same optical-flight window — a vanishingly small probability over realistic distances. Covered in the repeaters subject.
• Purification needs at least two Bell pairs co-located at each end. The first pair sits in memory while the second is being delivered; local CNOTs and measurement then trade the two pairs for one of higher fidelity. Covered in the purification subject.
• Quantum-frequency conversion ties memory wavelength to the telecom C-band. A memory at 737 nm cannot itself sit on a deployed fibre link; a QFC stage between the memory and the link translates incoming and outgoing photons to and from the C-band. Covered in the transduction subject.

The all-photonic-repeater family is the exception that proves the rule: by encoding the entanglement into a graph state of photons up front, it removes the memory requirement at the price of an order-of-magnitude harder photon source ^{Kumar et al. 2025} . Covered in the all-photonic QR subject.