Fiber Memory: Mechanisms & Applications

• Fiber memory is a family of phenomena where structured information is preserved during fiber-mediated transformations across optical, quantum, and mechanical systems.
• Research reveals deterministic correlations in multimode fibers and efficient quantum state storage in doped fibers with metrics like 1.4 µm focus precision and over 99% fidelity.
• Innovative implementations include recirculating delay-line memories for data centers and channel memory in coherent systems, enhancing energy efficiency and reducing replication.

“Fiber memory” is not a single concept but a family of research usages centered on persistence under fiber-mediated transformation. In multimode-fiber optics it denotes deterministic input–output correlations such as rotational, shift–shift, radial, and polarization memory effects; in quantum photonics it denotes fiber-compatible or fiber-integrated devices that store and retrieve photonic states; in coherent communications it denotes the temporal dependence induced by chromatic dispersion and Kerr nonlinearity; and in data-center architecture it denotes a recirculating optical delay-line memory for immutable data such as LLM weights (Amitonova et al., 2015, Zhang et al., 2023, Ming et al., 2021, Atmer et al., 9 Jul 2026).

1. Multimode-fiber memory effects

In multimode fibers, “memory” most often refers to partial invariances of a complex input–output map rather than to storage in time. Amitonova and co-workers identified the rotational memory effect in a circular-core step-index multimode fiber: rotating the incident wavefront around the fiber core axis rotates the output speckle by approximately the same angle while preserving substantial correlation, because in an ideal cylindrically symmetric fiber the propagation approximately commutes with rotations about the axis (Amitonova et al., 2015).

The experimentally demonstrated rotational case used a 12 cm long, 50 µm core, NA = 0.22 step-index multimode fiber at $\lambda = 633$ nm. The correlation $r(\alpha)$ never dropped to $0$, reached a minimum of about $0.3$ around $\pm 90^\circ$, showed revivals near $\pm 180^\circ$, and had $\Delta \alpha_{0.5} \approx 100^\circ$ when the rotation center coincided with the true fiber core axis. The same work showed that a single optimized phase mask could be rotated to scan a diffraction-limited focus around a full circle on the output facet, with focus FWHM $\approx 1.4\,\mu\text{m}$, standard deviation $0.08\,\mu\text{m}$, and power in the focus varying between $4\%$ and $r(\alpha)$0 during the rotation (Amitonova et al., 2015).

A different symmetry appears in square-core fibers. Caravaca-Aguirre and collaborators demonstrated a shift–shift memory effect in a $r(\alpha)$1, 30 mm square-core multimode fiber at 532 nm: translating the optimized input field causes the output focus to split into four foci moving in four symmetric directions set by the square symmetry. Their DMD-based optimization used 6400 input modes over an $r(\alpha)$2 area, achieved a focus enhancement factor exceeding 2000, and yielded memory-effect ranges $r(\alpha)$3 between $r(\alpha)$4 and $r(\alpha)$5, enabling fluorescence imaging over a $r(\alpha)$6 field of view without measuring a transmission matrix (Mezil et al., 2023).

A recent extension is the radial memory effect in step-index multimode fiber, defined by the observation that an input focused spot at radius $r(\alpha)$7 produces an output speckle with a ring of excess average intensity at the same radius. In the modal description, this arises because the input preferentially excites modes whose transverse intensity is concentrated near $r(\alpha)$8, and those modes remain radially concentrated after propagation even though their relative phases generate speckle. The effect was observed in fibers ranging from 20 cm to 20 m, was reported as robust against fiber perturbations such as bending, twisting, or shaking, and was proposed as a basis for spatial multiplexing; in ideal theory the ring width scales as $r(\alpha)$9, and the number of resolvable annular channels as $0$0 (Gokay et al., 15 Aug 2025).

Another orthogonal axis is polarization memory. Arora and Amitonova reported that, in a 50 cm step-index multimode fiber, the spatially resolved output polarization is random across the speckle field, yet the local Stokes vectors remain highly correlated with the input polarization state: rotating the input linear polarization by a half-wave plate causes the local polarization ellipse at each output point to rotate on the Poincaré sphere in a deterministic way, while the total speckle intensity distribution remains essentially unchanged. In their interpretation, the fiber preserves the relative phase between input circular components and transfers it into local output polarization structure, even in the presence of strong mode coupling and birefringence (Arora et al., 6 Sep 2025).

2. Fiber-compatible and fiber-integrated quantum memories

In quantum networking, “fiber memory” usually denotes a device that can reversibly map the quantum state of light into a storage medium while remaining spectrally or physically compatible with optical fiber infrastructure. Two representative realizations are an erbium-doped optical fiber memory and a fiber-pigtailed Er$0$1:LiNbO$0$2 waveguide memory, both operating directly in the telecom band near 1532 nm (Jin et al., 2015, Zhang et al., 2023).

Saglamyurek and colleagues demonstrated an atomic quantum memory in a 20 m erbium-doped silica optical fiber using the atomic frequency comb protocol. The memory operated on the $0$3 transition around 1532 nm, at 0.8 K and 600 G, with an 8 GHz AFC of tooth spacing $0$4 MHz and storage time $0$5 ns. The recall efficiency for 5 ns storage was about $0$6, storage times up to 50 ns were achieved by varying $0$7, the cross-correlation was $0$8 before storage and $0$9 after storage and recall, and full polarization-qubit tomography yielded fidelities at or above $0.3$0 for the tested states (Jin et al., 2015).

A distinct integrated implementation used a fiber-pigtailed Er$0.3$1:LiNbO$0.3$2 waveguide fabricated by femtosecond laser micromachining inside a 10×10×0.5 mm$0.3$3 erbium-doped lithium-niobate wafer. The device operated at 1532.05 nm in a dilution refrigerator at 13 mK, with a 4 GHz-wide AFC and a fixed storage time $0.3$4 ns set by $0.3$5 MHz, giving a time–bandwidth product of 800. It stored 330 temporal modes of heralded single photons, increased the coincidence detection rate by a factor of 167 relative to single-mode operation, and preserved non-classicality: $0.3$6 before storage, $0.3$7 after storage, and the heralded autocorrelation changed from $0.3$8 to $0.3$9 (Zhang et al., 2023).

These devices matter because they avoid wavelength conversion stages that would otherwise be needed to interface visible or near-IR memories with telecom fiber. A broader network implication is visible in the 420 km memory–memory entanglement experiment of 2025, where two cavity-enhanced $\pm 90^\circ$0Rb ensemble memories emitted 780 nm photons that were converted to telecom S-band at 1522 nm and sent through ultra-low-loss fiber to a middle station using a DLCZ protocol. That experiment demonstrated entanglement between two atomic ensemble quantum memories over 420 km and reported that the memory–memory entangling probability beat the repeaterless channel capacity for direct entanglement distribution (Luo et al., 8 Apr 2025).

3. Photonic fiber memories based on loops and cavities

A separate class of fiber memories stores photons as light in fiber, rather than as matter excitations. Here the memory medium is a loop or cavity formed directly in optical fiber, and storage is realized by trapping, recirculating, and later releasing guided pulses. This distinguishes photonic fiber memories from AFC, Raman, or spin-wave memories, even when both are described as “quantum memories” (Bonsma-Fisher et al., 2023, Cheng et al., 15 May 2025).

One route uses fiber Bragg grating cavities. In a spliced-fiber cavity formed by FBGs around a 5.08 m single-mode fiber, telecom pulses at 1585 nm were converted into a cavity wavelength near 1550 nm using Bragg-scattering four-wave mixing, stored for up to 11 round trips or $\pm 90^\circ$1, and retrieved with total memory efficiency $\pm 90^\circ$2 and SNR 12.8 after one round trip. A monolithic cavity with directly written chirped and anti-chirped FBGs extended the storage to 35 round trips or $\pm 90^\circ$3, with total memory efficiency $\pm 90^\circ$4 and SNR $\pm 90^\circ$5 after one round trip (Bonsma-Fisher et al., 2023).

A closely related cavity architecture used intracavity frequency translation in birefringent PM fiber with dichroic-coated end facets. In that system, a 902.5 nm pulse was switched into resonance at 925 nm by Bragg-scattering four-wave mixing driven by 790.1 nm and 807.4 nm control pulses, stored in a 1.285 m fiber cavity with round-trip time $\pm 90^\circ$6 ns, and released by the inverse translation. The device demonstrated storage of quantum-level THz-bandwidth coherent states for up to 16 cavity round trips or about 200 ns, a maximum overall efficiency of $\pm 90^\circ$7, write efficiency $\pm 90^\circ$8, read efficiency $\pm 90^\circ$9, and classical spectral fidelity above $\pm 180^\circ$0 (Bustard et al., 2022).

Telecom single-photon storage was realized in a related fiber-based cavity quantum memory using the FC-SWIFT protocol. Heralded O-band photons at 1260 nm with 81 GHz bandwidth were translated into a trapped wavelength at 1291.5 nm and later retrieved with a $\pm 180^\circ$1 lifetime of $\pm 180^\circ$2, corresponding to 32.8 cavity round trips. The internal memory efficiency after $\pm 180^\circ$3 of storage was $\pm 180^\circ$4, the total efficiency was $\pm 180^\circ$5, and non-classical cross-correlation remained above 2 for up to 70 round trips (Bonsma-Fisher et al., 2024).

The same general platform has also been used to generate photons inside the fiber cavity rather than only store externally generated ones. In that variant, spontaneous four-wave mixing inside a birefringent fiber cavity created heralded photons directly in the storage mode, and on-demand readout was performed by intracavity Bragg-scattering four-wave mixing. The reported figures were $\pm 180^\circ$6 in the first readout bin, readout frequency-translation efficiency of about $\pm 180^\circ$7, $\pm 180^\circ$8 memory lifetime of about 67 cavity cycles or $\pm 180^\circ$9, and $\Delta \alpha_{0.5} \approx 100^\circ$0 after one cavity cycle in an alternate low-noise cavity (Bustard et al., 2024).

A simpler photonic variant is the loop-and-switch architecture. A fiber-coupled broadband quantum memory for polarization-encoded photonic qubits used a fiber loop as the storage line and a fast polarization-preserving switch as the write/read element. It reported a pass-through efficiency of about $\Delta \alpha_{0.5} \approx 100^\circ$1, overall storage efficiency scaling as about $\Delta \alpha_{0.5} \approx 100^\circ$2 for $\Delta \alpha_{0.5} \approx 100^\circ$3 storage cycles, two storage-cycle times of about 40 ns and $\Delta \alpha_{0.5} \approx 100^\circ$4, and high-fidelity storage and retrieval of ultra-broadband single-photon polarization qubits in both cases (Cheng et al., 15 May 2025).

4. Channel memory in coherent optical communication

In coherent transmission research, “fiber memory” does not refer to a device but to the fact that the fiber channel is not memoryless. Chromatic dispersion spreads symbols over many symbol periods, Kerr nonlinearity makes the instantaneous phase depend on the surrounding waveform, and in WDM systems cross-phase modulation and four-wave mixing couple symbols across channels. The received symbol at time $\Delta \alpha_{0.5} \approx 100^\circ$5 therefore depends on many past and future symbols, a structure that is commonly modeled from the NLSE or Manakov equation and treated as a dispersive–nonlinear channel with long temporal memory (Deligiannidis et al., 2020, Ming et al., 2021).

This interpretation motivated the use of LSTMs as nonlinear equalizers. In the 2020 study of LSTM-based compensation, the sequence length $\Delta \alpha_{0.5} \approx 100^\circ$6 was taken as the effective channel-memory window. The paper connected the optimal word length directly to physical dispersion: at 1310 nm over 300 km, the delay spread was about 77 ps, or about 2 symbol durations at 25 Gbaud, and BER saturated for word lengths above 3 symbols; at 1550 nm over 500 km, the delay spread was about 3.4 ns, or about 85 symbols, and performance improved until the word length reached about 50 symbols. In WDM scenarios, the same work reported that LSTM post-processing could outperform single-channel DBP, especially because the recurrent model implicitly captured inter-channel nonlinear memory from the central channel waveform, and it noted that LSTM could be less complex than DBP in long distances above 1000 km (Deligiannidis et al., 2020).

A later refinement introduced the center-oriented LSTM (Co-LSTM) for a 10-channel, 64 Gbaud PDM-16QAM Nyquist-WDM system over up to 1600 km. The architecture used a bidirectional, center-oriented structure and a recycling-based simplified equalization mode whose per-symbol complexity was almost independent of transmission distance. At 1600 km and 1.0 dBm launch power, Co-LSTM achieved a $\Delta \alpha_{0.5} \approx 100^\circ$7 gain of $\Delta \alpha_{0.5} \approx 100^\circ$8 dB over CDC, slightly better $\Delta \alpha_{0.5} \approx 100^\circ$9 than Bi-LSTM, only $\approx 1.4\,\mu\text{m}$0 of Bi-LSTM complexity, and $\approx 1.4\,\mu\text{m}$1 of the complexity of DBP with 1 step per span (Ming et al., 2021).

5. Fiber memory in quantum repeater architectures

In quantum-repeater theory, “fiber memory” usually appears at the interface between lossy optical fiber and matter-based memory stations. The 2019 modular analysis of repeater cells modeled the fiber channel with attenuation length $\approx 1.4\,\mu\text{m}$2 km, corresponding to standard telecom loss $\approx 1.4\,\mu\text{m}$3, and characterized memory stations by three parameters: zero-length link efficiency $\approx 1.4\,\mu\text{m}$4, clock time $\approx 1.4\,\mu\text{m}$5, and coherence time $\approx 1.4\,\mu\text{m}$6. Within that framework, whether a repeater cell beats direct transmission depends on the interplay between fiber loss and the waiting-time budget set by the memory coherence (Loock et al., 2019).

The same analysis separated two protocol families. In node-sends-photons (NSP) schemes, a central memory node emits entangled photons through two fiber halves, so the relevant waiting time scales with the fiber propagation time $\approx 1.4\,\mu\text{m}$7; this makes long memory coherence essential. In node-receives-photons (NRP) schemes, incoming photons are written into memories locally, so the attempt time is set mainly by the local clock and fast sources can compensate shorter coherence. The comparison suggested that present or future rubidium memories are particularly attractive because of high $\approx 1.4\,\mu\text{m}$8 and long $\approx 1.4\,\mu\text{m}$9, while quantum dots benefit most from NRP-type protocols because of their high source clock rate (Loock et al., 2019).

The 420 km atomic-ensemble experiment provides a concrete realization of this fiber–memory interface at metropolitan to intercity scale. There, quantum frequency conversion from 780 nm to 1522 nm enabled the use of fiber with $0.08\,\mu\text{m}$0 dB/km loss, and fulltime far-off-resonant locking plus intermittent dual-band locking stabilized the phase between remote memories sufficiently for DLCZ entanglement generation. A plausible implication is that such demonstrations turn repeater-cell abstractions into experimentally parameterized modules, rather than purely asymptotic constructions (Luo et al., 8 Apr 2025).

6. Architectural and extended uses of “fiber memory”

A radically different meaning appears in data-center systems. The 2026 architecture titled Fiber Memory proposes using optical fiber itself as an active, recirculating delay-line memory for immutable data, especially LLM weights. In that design, a central weights server writes the model once into a 1000 km multi-core-fiber loop, the weights circulate continuously as an optical stream, and every accelerator taps a small fraction of the signal through passive optical interfaces and co-packaged optics instead of storing local copies in HBM. The case study used 14 cables of 19-core multi-core fiber, 256 active cores, 8 wavelengths per core at 100 Gb/s, aggregate bandwidth $0.08\,\mu\text{m}$1 TB/s, and in-flight capacity $0.08\,\mu\text{m}$2 GB (Atmer et al., 9 Jul 2026).

The architectural objective is to eliminate replicated static weight storage. For a Llama-3-70B INT8 deployment, the study argued that the ring can hold the 70 GB model plus slack and replicated Streamed Weight Packets, serve 10,000 AI accelerators in a data-parallel optical broadcast system, eliminate redundant weight storage across those 10,000 accelerators, and reduce weight-delivery energy by over 70% relative to traditional HBM3e. In the detailed power accounting, HBM-based weight delivery was estimated at 1024 kW and Fiber Memory at 284.8 kW, corresponding to a 72.1% reduction (Atmer et al., 9 Jul 2026).

A distinct, non-photonic extension of the phrase appears in mechanics of random fibrous matrices rather than optical fiber. There, “memory” refers to permanent, load-history-dependent remodeling after tension release, even in the absence of explicit plasticity mechanisms, inter-fiber cohesion, or fiber yielding. Numerical simulations showed that localized tension creates alignment gradients, and those gradients drive additional remodeling during unloading, so the matrix retains a mechanical memory of the applied deformation through persistent defects, altered density, and changes in effective stiffness (Sarkar et al., 23 Jan 2025).

Across these usages, “fiber memory” denotes persistence encoded in very different substrates: a symmetry of multimode propagation, a stored atomic or photonic state, a temporal dependency of a dispersive–nonlinear channel, a recirculating optical data stream, or a geometry-driven residual state in a fibrous matrix. This suggests that the common core of the term is not a single mechanism but the retention of structured information by a fiber system under transformation.