Ghost Memory in Multi-Domain Research

• Ghost Memory is a concept where indirect carriers, such as zero-photon time bins and proxy signals, enable reconstruction of information without direct storage.
• It spans diverse applications from ghost imaging in optics and quantum memory readouts to efficient software proxies and privacy leakage detection.
• Recent studies quantify substantial resource savings and improved performance, such as a 1.5 Mbit memory reduction in FPGA-based imaging and tenfold FLOP reductions in neural networks.

Searching arXiv for relevant papers on “ghost memory” and closely related usages across optics, quantum memory, systems, and LLM privacy. arXiv search query: ghost memory ghost imaging memory “Ghost memory” is not a single standardized technical term in the arXiv literature. Rather, it appears across several research areas as a recurring description for situations in which information is retained, reconstructed, or made operational through indirect carriers: zero-photon time bins in ghost imaging, streaming differential accumulators that replace full measurement histories, spin-wave storage that delays one arm of a nonlocal image, non-Markovian variables that steer optical relaxation into a saddle-node bottleneck, persistent DRAM side effects from transient execution, and privacy-relevant traces resurfacing from LLM memory systems (Chen et al., 9 Apr 2026, Yang et al., 2019, Mazelanik et al., 2021, Boer et al., 13 May 2026, Zhang et al., 2020, Zhang et al., 2024). This suggests a unifying motif: usable structure survives even when the direct signal, direct storage model, or direct causal path has been discarded, hidden, or rendered nonlocal.

1. Semantic range and core motif

Across the cited literature, “ghost memory” functions less as a formal theory than as a family resemblance term. In some papers, it denotes an explicit memory mechanism; in others, it is a metaphor for distributed encoding, hidden-state retention, or side effects that outlive the nominal event. A common misconception is to treat all such uses as references to ordinary digital storage or to biological memory. Several of the relevant works are explicit that they address neither cognitive memory nor conventional storage semantics, but instead imaging correlations, indefinite-metric sectors, or proxy-object footprint (Ceddia et al., 2018, Jatkar et al., 2016, Peck et al., 2013).

Area	Ghost-memory interpretation	Representative paper
Ghost imaging with zero photons	Image reconstruction from zero-photon time bins	(Chen et al., 9 Apr 2026)
Instant ghost imaging	One-frame-plus-accumulator memory model	(Yang et al., 2019)
Quantum-memory ghost imaging	Spin-wave storage and on-demand idler release	(Mazelanik et al., 2021)
Ghost state of light	Memory steers dynamics into a saddle-node ghost	(Boer et al., 13 May 2026)
Software and systems	Low-footprint proxies or persistent side effects in memory	(Peck et al., 2013, Zhang et al., 2020)
LLM privacy	Past inputs and RAG memories resurface or enable inference leakage	(Zhang et al., 2024)

The significance of this semantic spread is methodological. In optics, the “ghost” often refers to nonlocal or correlation-based recovery. In systems work, it can refer either to compact memory use or to architectural effects that survive beyond their authorized execution path. In privacy research, it denotes latent information that remains accessible or inferable after the original interaction has receded from immediate view.

2. Distributed reconstruction and information from absence

In ghost imaging, the canonical mechanism is already memory-like in a distributed sense: the object is not recorded pixel-by-pixel at the detecting arm, but reconstructed from correlations between a bucket detector and a spatially resolved reference. “Ghost imaging with zero photons” extends this logic to the statistics of absences. The reported experiment uses pseudothermal light from a continuous-wave laser passing through a rotating ground glass, split by a 1:1 non-polarizing beam splitter into two arms. One arm passes through the object to a bucket detector $D_2$; the other is scanned spatially by a fiber coupler and detected by $D_1$. Reconstruction is performed by conditioning on zero-photon time bins of width $T$, so that the relevant information is carried by photon-number vacancies rather than detected arrivals (Chen et al., 9 Apr 2026).

The formalism is written in terms of a joint probability generating function,

$$M(x,y)= \frac{1}{1-\bar n(x-1)-\bar n(y-1)+(1-\mu)\bar n^2(x-1)(y-1)},$$

with $\bar n=\langle I\rangle T$ and $\mu = |g^{(1)}(\vec r_1,t_1;\vec r_2,t_2)|^2$. The work further gives

$$P_{m0}= \frac{\bigl[\bar n+(1-\mu)\bar n^2\bigr]^m} {\bigl[1+2\bar n+(1-\mu)\bar n^2\bigr]^{m+1}}.$$

Within this framework, a negative ghost image is reconstructed via $g^{(2)}_{10}(0)$, while a positive ghost image is obtained via $g^{(2)}_{00}(0)$. The paper emphasizes that all photons interacting with the object can be discarded, and that in the $g^{(2)}_{00}(0)$ case both arms are conditioned on zero photons, yet the image remains recoverable (Chen et al., 9 Apr 2026).

A second strand interprets ghost imaging as a random-basis reconstruction problem rather than as a direct memory process. In “On Random-Matrix Bases, Ghost Imaging and X-ray Phase Contrast Computational Ghost Imaging,” an image $D_1$0 is synthesized from random matrices $D_1$1 with weights $D_1$2, yielding

$D_1$3

The resulting global scaling,

$D_1$4

makes the encoding explicitly distributed over many random illuminations rather than concentrated in a single direct measurement (Ceddia et al., 2018). This is one reason the “memory” language is plausible but metaphorical: the object is effectively recalled from a set of correlation coefficients and known patterns.

A related extension appears in reciprocity-assisted ghost imaging through dynamic random media. There, optical reciprocity,

$D_1$5

is combined with the memory effect,

$D_1$6

to produce correlations between counter-propagating beams across a finite spatial patch. The object and camera are placed at opposite ends of the random medium; the object arm is measured by a bucket detector, the reference arm by a camera, and the object must be smaller than the memory-effect range. The importance of this result is that a dynamic random medium need not destroy ghost imaging if reciprocity and the memory effect remain usable (Tananyan et al., 2024).

3. Streaming memory reduction and quantum-memory-assisted ghost imaging

“Instant Ghost Imaging: Algorithm and On-chip Implementation” turns ghost memory into an explicit storage problem. Conventional background-subtraction ghost imaging computes

$D_1$7

and therefore requires retaining all bucket signals $D_1$8, all reference images $D_1$9, and often the running sums. For an image of size $T$0 and $T$1 measurements, the reported storage for the measured $T$2 data is about $T$3, whereas the FPGA used in the work has only $T$4 of storage (Yang et al., 2019).

The instant ghost imaging algorithm replaces this batch model with a differential streaming rule,

$T$5

Operationally, the system stores $T$6 in register $T$7, $T$8 in register $T$9, and the running sum $$M(x,y)= \frac{1}{1-\bar n(x-1)-\bar n(y-1)+(1-\mu)\bar n^2(x-1)(y-1)},$$0 in register $$M(x,y)= \frac{1}{1-\bar n(x-1)-\bar n(y-1)+(1-\mu)\bar n^2(x-1)(y-1)},$$1. When $$M(x,y)= \frac{1}{1-\bar n(x-1)-\bar n(y-1)+(1-\mu)\bar n^2(x-1)(y-1)},$$2 and $$M(x,y)= \frac{1}{1-\bar n(x-1)-\bar n(y-1)+(1-\mu)\bar n^2(x-1)(y-1)},$$3 arrive, it computes the two differences, multiplies them, adds the result to the accumulator, and overwrites the old values. The paper states that IGI needs only $$M(x,y)= \frac{1}{1-\bar n(x-1)-\bar n(y-1)+(1-\mu)\bar n^2(x-1)(y-1)},$$4 to store one $$M(x,y)= \frac{1}{1-\bar n(x-1)-\bar n(y-1)+(1-\mu)\bar n^2(x-1)(y-1)},$$5 frame, reports an on-chip system with two CMOS detectors, an FPGA, and a monitor on a PCB with no external memory, and gives a throughput of $$M(x,y)= \frac{1}{1-\bar n(x-1)-\bar n(y-1)+(1-\mu)\bar n^2(x-1)(y-1)},$$6 measurements per second. After $$M(x,y)= \frac{1}{1-\bar n(x-1)-\bar n(y-1)+(1-\mu)\bar n^2(x-1)(y-1)},$$7 measurements, the accumulated value for one pixel reaches about $$M(x,y)= \frac{1}{1-\bar n(x-1)-\bar n(y-1)+(1-\mu)\bar n^2(x-1)(y-1)},$$8 in conventional GI, needing about $$M(x,y)= \frac{1}{1-\bar n(x-1)-\bar n(y-1)+(1-\mu)\bar n^2(x-1)(y-1)},$$9 bits, versus about $\bar n=\langle I\rangle T$0 in IGI, needing about $\bar n=\langle I\rangle T$1 bits, for a stated saving of about $\bar n=\langle I\rangle T$2 over the full $\bar n=\langle I\rangle T$3 image (Yang et al., 2019).

A distinct but complementary meaning appears in real-time ghost imaging of Bell-nonlocal entanglement between a photon and a quantum memory. Here the memory is not a storage bottleneck but an active physical element in the imaging protocol. The experiment uses a pencil-shaped cold cloud of $\bar n=\langle I\rangle T$4 atoms in a 3D magneto-optical trap operating at about $\bar n=\langle I\rangle T$5, with a $\bar n=\langle I\rangle T$6 write pulse detuned by $\bar n=\langle I\rangle T$7 and a $\bar n=\langle I\rangle T$8 read pulse. A detected signal photon at a bucket detector triggers, via FPGA feedback, the readout of the stored spin wave and the gating of an intensified CMOS camera for the idler. Because the quantum memory is spatially multimode and wavevector-multiplexed, the image is acquired in full field rather than by sequential scanning (Mazelanik et al., 2021).

The work reports a Bell parameter of

$\bar n=\langle I\rangle T$9

obtained from a single far-field image, with local visibilities around $\mu = |g^{(1)}(\vec r_1,t_1;\vec r_2,t_2)|^2$0. The memory-assisted architecture replaces the image-preserving optical delay lines used in earlier ghost-imaging experiments with programmable storage whose delay can extend to tens of microseconds, exceeding the nanosecond-scale delays of optical fiber lines. In this usage, “ghost memory” is literal in both senses: the image is nonlocal, and the quantum memory itself stores and releases one arm of the correlated pair (Mazelanik et al., 2021).

4. Non-Markovian optical ghosts and long-lived transient states

In “Ghost State of Light,” memory is the mechanism that realizes a ghost rather than merely compressing or transporting one. The system is a plano-concave Fabry–Pérot microcavity filled with olive oil, whose thermo-optical response introduces a delayed nonlinearity. The cavity is driven just outside the bistable region, near a saddle-node bifurcation (SNB). Although the stable and unstable fixed points have annihilated, the phase space retains a bottleneck in which trajectories evolve very slowly. The paper identifies a nonlinear response with memory as the ingredient that steers the system into this bottleneck and thereby realizes a long-lived non-stationary ghost state (Boer et al., 13 May 2026).

The model is

$\mu = |g^{(1)}(\vec r_1,t_1;\vec r_2,t_2)|^2$1

with $\mu = |g^{(1)}(\vec r_1,t_1;\vec r_2,t_2)|^2$2, $\mu = |g^{(1)}(\vec r_1,t_1;\vec r_2,t_2)|^2$3, and hidden variable

$\mu = |g^{(1)}(\vec r_1,t_1;\vec r_2,t_2)|^2$4

The visible signature is a plateau in the transmitted intensity: near the SNB, transmission rises quickly, remains on a plateau for milliseconds, and only later reaches the final steady state. Farther from the SNB, the plateau disappears (Boer et al., 13 May 2026).

The quantitative claims are unusually strong. The reported plateau lasts milliseconds, exceeds the picosecond cavity photon lifetime by more than $\mu = |g^{(1)}(\vec r_1,t_1;\vec r_2,t_2)|^2$5 orders of magnitude, and exceeds the oil’s microsecond thermal relaxation time by about a factor of $\mu = |g^{(1)}(\vec r_1,t_1;\vec r_2,t_2)|^2$6. For sufficiently large $\mu = |g^{(1)}(\vec r_1,t_1;\vec r_2,t_2)|^2$7, specifically $\mu = |g^{(1)}(\vec r_1,t_1;\vec r_2,t_2)|^2$8, the plateau duration follows

$\mu = |g^{(1)}(\vec r_1,t_1;\vec r_2,t_2)|^2$9

while in the regime $$P_{m0}= \frac{\bigl[\bar n+(1-\mu)\bar n^2\bigr]^m} {\bigl[1+2\bar n+(1-\mu)\bar n^2\bigr]^{m+1}}.$$0 it scales approximately linearly with $$P_{m0}= \frac{\bigl[\bar n+(1-\mu)\bar n^2\bigr]^m} {\bigl[1+2\bar n+(1-\mu)\bar n^2\bigr]^{m+1}}.$$1. At fixed driving conditions, the lifetime distribution approximately follows

$$P_{m0}= \frac{\bigl[\bar n+(1-\mu)\bar n^2\bigr]^m} {\bigl[1+2\bar n+(1-\mu)\bar n^2\bigr]^{m+1}}.$$2

over a limited range, with a fat tail attributed to randomness in the initial detuning $$P_{m0}= \frac{\bigl[\bar n+(1-\mu)\bar n^2\bigr]^m} {\bigl[1+2\bar n+(1-\mu)\bar n^2\bigr]^{m+1}}.$$3 (Boer et al., 13 May 2026). The broader interpretation is that memory is neither a passive record nor an ancillary feature; it reshapes the relaxation pathway so that a transient remnant of the bifurcation becomes experimentally accessible.

5. Memory-efficient ghost mechanisms in software and neural architectures

In software infrastructure, “Ghost” designates a proxy framework whose stated goals include low memory consumption. “Ghost: A Uniform and General-Purpose Proxy Implementation” separates interception from handling through AbstractProxy, AbstractProxyHandler, and Interception objects. AbstractProxy defines no structure, and handlers can be shared across many proxies. On a $$P_{m0}= \frac{\bigl[\bar n+(1-\mu)\bar n^2\bigr]^m} {\bigl[1+2\bar n+(1-\mu)\bar n^2\bigr]^{m+1}}.$$4-bit system, a stateless shared handler saves $$P_{m0}= \frac{\bigl[\bar n+(1-\mu)\bar n^2\bigr]^m} {\bigl[1+2\bar n+(1-\mu)\bar n^2\bigr]^{m+1}}.$$5 bytes by avoiding a per-proxy handler field and $$P_{m0}= \frac{\bigl[\bar n+(1-\mu)\bar n^2\bigr]^m} {\bigl[1+2\bar n+(1-\mu)\bar n^2\bigr]^{m+1}}.$$6 bytes by avoiding the handler object itself, for a total of $$P_{m0}= \frac{\bigl[\bar n+(1-\mu)\bar n^2\bigr]^m} {\bigl[1+2\bar n+(1-\mu)\bar n^2\bigr]^{m+1}}.$$7 bytes saved per proxy. The compact classes TargetBasedProxy and TargetBasedClassProxy use a $$P_{m0}= \frac{\bigl[\bar n+(1-\mu)\bar n^2\bigr]^m} {\bigl[1+2\bar n+(1-\mu)\bar n^2\bigr]^{m+1}}.$$8-byte header instead of $$P_{m0}= \frac{\bigl[\bar n+(1-\mu)\bar n^2\bigr]^m} {\bigl[1+2\bar n+(1-\mu)\bar n^2\bigr]^{m+1}}.$$9 bytes, or $g^{(2)}_{10}(0)$0 instead of $g^{(2)}_{10}(0)$1 bytes for larger bodies, by encoding the class reference in $g^{(2)}_{10}(0)$2 bits and allowing up to $g^{(2)}_{10}(0)$3 compact classes (Peck et al., 2013).

The memory logic is architectural rather than metaphorical. Proxies for regular objects, classes, and methods are designed so that class-specific layout does not inflate every proxy type, and handler sharing avoids duplicating behavior state. In the Marea case study, these choices contribute to reported application memory footprint reductions of $g^{(2)}_{10}(0)$4 to $g^{(2)}_{10}(0)$5 (Peck et al., 2013).

In neural super-resolution, the relevant mechanism is feature-memory economy. “GRAN: Ghost Residual Attention Network for Single Image Super Resolution” argues that many intermediate feature maps generated by standard convolutions are redundant and therefore expensive in both computation and memory. Its Ghost Module computes a smaller set of intrinsic features and generates additional ghost features through cheap linear operations:

$g^{(2)}_{10}(0)$6

GRAN combines this with a Channel and Spatial Attention Module (CSAM) to attend to “where and what” is being extracted (Niu et al., 2023).

The reported efficiency numbers are concrete. Table 2 gives RCAN as $g^{(2)}_{10}(0)$7M parameters and $g^{(2)}_{10}(0)$8G FLOPs, MSRN as $g^{(2)}_{10}(0)$9M and $g^{(2)}_{00}(0)$0G, DBPN as $g^{(2)}_{00}(0)$1M and $g^{(2)}_{00}(0)$2G, HRAN as $g^{(2)}_{00}(0)$3M and $g^{(2)}_{00}(0)$4G, and GRAN as $g^{(2)}_{00}(0)$5M parameters and $g^{(2)}_{00}(0)$6G FLOPs. The paper states that parameters and FLOPs decrease by more than ten times. In the ablation study, “Ghost Module + Channel Attention” gives $g^{(2)}_{00}(0)$7M parameters, $g^{(2)}_{00}(0)$8G FLOPs, and $g^{(2)}_{00}(0)$9 PSNR, while “Ghost Module + Channel + Spatial Attention” gives $g^{(2)}_{00}(0)$0M, $g^{(2)}_{00}(0)$1G, and $g^{(2)}_{00}(0)$2 PSNR (Niu et al., 2023). Here the “ghost” label marks a resource-saving generative shortcut: feature richness is synthesized from a compact base representation.

6. Persistent traces, integrity attacks, and privacy leakage

“GhostKnight: Breaching Data Integrity via Speculative Execution” introduces a markedly different sense of ghost memory. The paper argues that speculative execution can induce side effects within DRAM, not only traces in microarchitectural buffers. An attacker mistrains a Spectre-V1 bounds check, flushes the relevant data to delay resolution, and on the mispredicted path triggers an out-of-bounds speculative load that can reach DRAM. If repeated often enough, the resulting activations produce permanent bit flips. The critical distinction from spectre-type attacks is therefore confidentiality versus integrity: transient execution is used not as a readout oracle but as a fault-injection mechanism that survives rollback (Zhang et al., 2020).

The attack’s timing envelope is explicit. The work reports that if the time cost per hammer exceeds about $g^{(2)}_{00}(0)$3 cycles, no bit flip is observed within $g^{(2)}_{00}(0)$4 hours; with GhostKnight, the first bit flip appears within about $g^{(2)}_{00}(0)$5 minutes. All measured speculative-hammering costs were below the $g^{(2)}_{00}(0)$6-cycle threshold, with $g^{(2)}_{00}(0)$7 in the range $g^{(2)}_{00}(0)$8 and $g^{(2)}_{00}(0)$9 below $D_1$00 cycles. The described targets include user/kernel separation, inter-process isolation, intra-process sandboxing, MMU-virtualization-based TEEs such as Intel EPT, AMD NPT, and ARM second-stage translation, with a future-work example involving a $D_1$01-bit RSA exponentiation implementation inside a TEE (Zhang et al., 2020). In this usage, “ghost memory” names an architecturally forbidden access whose physical effect becomes permanent.

A privacy-oriented analogue appears in “Ghost of the past: identifying and resolving privacy leakage from LLM’s memory through proactive user interaction.” The paper distinguishes short-term context-window memory from long-term RAG-based memory and argues that both can leak private information either directly or by inference over aggregated past inputs. The formative study reports that only $D_1$02 participants understood long-term memory mechanisms, $D_1$03 were unaware of notification when memory or context were added, $D_1$04 were totally unaware of memory usage, and $D_1$05 had never clicked the memory management panel (Zhang et al., 2024).

MemoAnalyzer addresses this by inferring sensitive information from past inputs and memories, displaying it with background color temperature mapped to sensitivity and transparency mapped to confidence, and exposing the original sources and keywords for editing. The RGBA mapping is

$D_1$06

In a $D_1$07-day evaluation with $D_1$08, the paper reports significant reductions in inferred private information relative to the GPT and Manual baselines, with over $D_1$09 and $D_1$10 reduction in certain categories when evaluated with GPT-4o. Total interaction times are given as $D_1$11s for MemoAnalyzer, $D_1$12s for GPT-4o, and $D_1$13s for Manual, with no significant differences in total time across techniques overall except on Day 1 (Zhang et al., 2024). The underlying concern is that memory can resurface or combine dispersed disclosures long after the original utterance has disappeared from immediate attention.

7. Hidden sectors, negative norms, and metaphorical memory

At the most abstract end of the spectrum, “Entangled spins and ghost-spins” treats ghost memory as an effect of tracing over hidden indefinite-metric degrees of freedom. A single ghost-spin is a two-state system with

$D_1$14

or, in the diagonal basis,

$D_1$15

When a full state $D_1$16 of spins plus ghost-spins is reduced by tracing out the ghost sector,

$D_1$17

the sign structure of the ghost metric survives in the reduced density matrix (Jatkar et al., 2016).

The paper’s central result is structural. If the spin and ghost-spin sectors are disentangled, positive-norm states yield positive entanglement entropy, whereas negative-norm states yield entanglement entropy with a negative real part and a constant imaginary part, with $D_1$18 under the stated branch choice. When spins are entangled with ghost-spins, positive norm no longer guarantees positive entropy. With an even number of ghost-spins, subsectors exist in which positive-norm states always lead to positive entanglement entropy after tracing out the ghost sector. With an odd number of ghost-spins, positive-norm states with negative real part for entanglement entropy always exist (Jatkar et al., 2016).

This usage clarifies an important boundary condition for the broader term. In some fields, ghost memory is literal storage; in others it is a proxy for hidden-state inheritance. The ghost sector is not a memory device, but tracing over it leaves a persistent signature—sign inversions, negative eigenvalues, and branch-dependent imaginary contributions—in the observable subsystem. That outcome parallels the wider literature surveyed here: what disappears from direct access may still continue to structure what can later be reconstructed, inferred, or corrupted.