ResonanceDB: Wave-Based Semantic Memory

• ResonanceDB is a wave-based semantic memory system that represents each item as a complex waveform combining amplitude and phase to capture intensity and modulation.
• It uses resonance-based retrieval where similarity is measured through constructive interference, effectively distinguishing negation and operator effects unlike cosine similarity.
• The system implements exact scan-based comparison on commodity CPUs for interactive retrieval, balancing practical efficiency with the challenges of linear scaling and memory-mapped I/O.

ResonanceDB is a wave-based semantic memory system in which each stored item is represented as a complex-valued waveform rather than as a purely real embedding. In the formulation reported in "Wave-Based Semantic Memory with Resonance-Based Retrieval: A Phase-Aware Alternative to Vector Embedding Stores" (Listopad, 21 Aug 2025), its purpose is to address a stated weakness of conventional vector stores: real-valued similarity measures such as cosine or inner product are phase-insensitive and therefore poorly suited to polarity, negation, contextual shifts, and other operator-like semantic transformations.

1. Conceptual definition and problem setting

ResonanceDB is presented as an alternative to conventional vector embedding stores. The underlying claim is not that vector methods are ineffective for approximate semantic proximity, but that they ignore sign- or phase-like structure and can therefore conflate opposites or structured variants, including examples such as “happy” versus “not happy”. In that framing, semantic memory is modeled not as a point in a real vector space, but as a wave pattern whose alignment depends jointly on magnitude and phase (Listopad, 21 Aug 2025).

This design targets semantic phenomena that the source describes as especially difficult for cosine-style retrieval: negation, inversion, modality, discourse or context shifts, and intensity changes. The central distinction is that ResonanceDB attempts to preserve not only how strongly a feature is expressed, but also how that feature is modulated. The system is therefore phase-aware by construction.

A common misconception is to interpret the name as referring to a generic database of physical resonances or to resonance-counting work in random-matrix or localization settings. That interpretation is not supported by the source associated with (Kutlin et al., 2024), which the provided record explicitly identifies as a SciPost LaTeX submission template containing no substantive scientific content on resonance counting and no database called ResonanceDB (Kutlin et al., 2024).

2. Wave representation and semantic encoding

The core representational object is the complex waveform

$\psi(x)=A(x)e^{i\phi(x)},$

where $x$ indexes vector dimensions, $A(x)\ge 0$ is the amplitude, and $\phi(x)\in[-\pi,\pi)$ is the phase (Listopad, 21 Aug 2025).

Within this representation, amplitude encodes semantic intensity or salience, while phase encodes contextual or structural modulation. The source associates phase specifically with contextual modulation, semantic polarity, operator effects, and transformations such as negation or shift. This is the principal departure from standard embedding storage: instead of retaining only a real-valued vector, ResonanceDB stores a polar representation with both magnitude and phase.

The paper also states that existing real-valued vectors can be mapped into the wave format without retraining. For a real vector $v$, the sign-phase mapping is

$$A(x)=|v(x)|,\qquad \phi(x)= \begin{cases} 0, & v(x)\ge 0\ \pi, & v(x)<0 \end{cases}$$

and for nonnegative vectors one may use $\phi(x)=0$. This mapping preserves sign information while yielding a valid polar complex representation.

This representational choice supports operator-sensitive semantics. Negation is modeled as an anti-phase relation; controlled phase shifts correspond to semantic transformation of the same underlying concept under different contextual or logical status; amplitude changes encode intensity modulation. The paper’s explicit operator examples are NEG, SHIFT+, INT_UP, and INT_DOWN.

3. Resonance-based retrieval and its formal properties

Retrieval in ResonanceDB is based on a resonance score defined through constructive interference between two wave patterns. In prose, the score is highest when patterns have similar energy, amplitudes align, and phases align; it penalizes imbalance so that a high-energy pattern does not overwhelm a low-energy one (Listopad, 21 Aug 2025).

Several mathematical properties are stated in the appendix. The score is bounded,

$0\le S(\psi_1,\psi_2)\le 1,$

symmetric,

$S(\psi_1,\psi_2)=S(\psi_2,\psi_1),$

and maximal at self-match,

$S(\psi,\psi)=1 \quad (\psi\neq 0).$

It is also globally phase-invariant,

$x$0

and attains its anti-phase minimum when

$x$1

The anti-phase minimum is central to the system’s treatment of negation. If two patterns are anti-phased, they are driven to the minimum similarity despite comparable magnitudes. This is the formal mechanism by which negated or opposed patterns can be separated even when cosine similarity would keep them close.

The same appendix states that for real vectors of equal norm, the resonance score reduces to

$x$2

This implies that resonance retrieval is positioned as a generalization of cosine similarity rather than as an unrelated replacement. The added degrees of freedom arise from complex phase and energy balancing.

4. System architecture, storage model, and computational organization

ResonanceDB is implemented as a practical storage and retrieval system for fixed-length amplitude-phase patterns. The reported design stores amplitude-phase pairs $x$3, converts to complex form on demand, uses memory-mapped binary segments, and performs retrieval through a deterministic comparison kernel. The paper characterizes it as closer to content-addressable memory than to a neural memory system, and it explicitly states that there is no training and no learnable controller (Listopad, 21 Aug 2025).

The implementation exposes two backends. JavaKernel is the scalar CPU baseline and is the backend used for the reported benchmarks. SimdKernel is experimental, uses the Java 22 Panama Vector API, and is excluded from the reported performance numbers.

The indexing strategy is deliberately simple. The reported implementation uses exact scan-based retrieval, not approximate nearest-neighbor indexing. Its stated complexity is

$x$4

where $x$5 is the number of patterns and $x$6 is the pattern length. This is presented as a conscious engineering tradeoff: exactness and simplicity are retained, but search is not sublinear.

This design also constrains the system’s scaling claims. The abstract states scalability to millions of patterns with millisecond latency, whereas the detailed benchmark configuration in the provided record reports measurements up to $x$7 patterns under a specified CPU-only setup. The paper’s more conservative position in the detailed description is that exact full-waveform comparison remains interactive on commodity CPUs at moderate scale, while latency grows linearly with dataset size.

5. Empirical evaluation and reported retrieval behavior

The reported experiments were run on a single machine with Windows 11, Java HotSpot 22.0.1, a 12-core Intel CPU at 3.60 GHz, 32 GB RAM, no GPU, pattern lengths $x$8, dataset sizes from $x$9 to $A(x)\ge 0$0, top-$A(x)\ge 0$1 retrieval with $A(x)\ge 0$2, cold-cache runs, a 12-thread Java executor, and the scalar backend only (Listopad, 21 Aug 2025).

The evaluation criteria are listed as latency and scalability, sensitivity to amplitude and phase variation, and retrieval precision under structured semantic perturbations. The most direct comparative result concerns operator retrieval.

Operator	ResonanceDB P@1	Cosine P@1
NEG	1.0	$A(x)\ge 0$3
SHIFT+	1.0	$A(x)\ge 0$4
INT_UP	1.0	$A(x)\ge 0$5
INT_DOWN	1.0	$A(x)\ge 0$6

The paper further reports that cosine distances collapse near $A(x)\ge 0$7 for all operators, whereas resonance separates operators into distinct bands. The approximate resonance distance means are reported as NEG $A(x)\ge 0$8, SHIFT+ $A(x)\ge 0$9, INT_UP $\phi(x)\in[-\pi,\pi)$0, and INT_DOWN $\phi(x)\in[-\pi,\pi)$1. Heatmaps are described as dense and unstructured under cosine, but block-diagonal under resonance, with NEG aligning with anti-phased opposites, SHIFT+ forming stable offset clusters, and INT_UP and INT_DOWN separating into amplitude-driven regions.

Latency measurements are also given for the largest reported configuration.

Setting	Average latency	p95 latency
top-1, $\phi(x)\in[-\pi,\pi)$2, 500K patterns	$\phi(x)\in[-\pi,\pi)$3 ms	$\phi(x)\in[-\pi,\pi)$4 ms
top-10, $\phi(x)\in[-\pi,\pi)$5, 500K patterns	$\phi(x)\in[-\pi,\pi)$6 ms	$\phi(x)\in[-\pi,\pi)$7 ms
top-100, $\phi(x)\in[-\pi,\pi)$8, 500K patterns	$\phi(x)\in[-\pi,\pi)$9 ms	$v$0 ms

The interpretation supplied by the source is that these are interactive CPU-only latencies even without indexing or approximation. This suggests that the system’s empirical argument is not centered on asymptotic search acceleration, but on retaining exact resonance-based comparison while keeping moderate-scale retrieval operationally practical.

6. Limitations, applications, and positioning

The paper is explicit about several limitations. ResonanceDB has no training or adaptation; phases must be supplied by an encoder or by a procedure, and there is no learnable phase modulator. The system is sensitive to phase noise, especially high-frequency perturbations. Retrieval remains exact scan-based with complexity $v$1. Memory-mapped I/O can dominate at scale, amplitude calibration across corpora remains a concern, and the SIMD backend is experimental, so the reported performance claims rest on the scalar backend (Listopad, 21 Aug 2025).

These constraints shape the system’s scope. It is not presented as a drop-in ANN replacement optimized for sublinear search, nor as a learned memory architecture. It is instead a deterministic, phase-aware retrieval substrate intended to preserve operator semantics that real-valued similarity collapses.

The paper positions this substrate as relevant to RAG systems with richer query semantics, knowledge representation with polarity and modality, reasoning over negation and contextual change, structured semantic retrieval, and cognitively inspired memory substrates for AGI-oriented architectures. The stated rationale is that memory should represent not just which concepts are near each other, but also how concepts are transformed, qualified, or opposed.

A plausible implication is that ResonanceDB is best understood as a semantic-memory formalism and systems design that elevates phase to a first-class representational variable. Within the source’s own terms, its defining claim is that operator-aware retrieval can be realized through complex-valued storage, constructive interference scoring, exact scan-based comparison, and a storage layout designed for deterministic large-scale lookup rather than for learned adaptation.