Holographic Declarative Memory

• Holographic Declarative Memory (HDM) is a computational paradigm that encodes declarative items as distributed, overlapping traces retrievable via cue-driven reconstruction.
• The framework leverages vector-symbolic architectures like Holographic Reduced Representations and circular convolution to enable robust partial matching and reconstructive recall.
• HDM spans diverse implementations—from ACT-R adaptations to transformer-based systems, magnonic devices, and quantum codes—addressing challenges in scalability and interference.

Holographic Declarative Memory (HDM) is used in the cited literature for memory schemes in which declarative content is encoded in distributed form and retrieved by cue-driven reconstruction rather than by direct location-based lookup. In the most explicit usage, HDM is a vector-symbolic alternative to ACT-R’s Declarative Memory; in adjacent work, the same holographic-declarative idea is developed through latent-space addressing in transformers, external knowledge manifolds, magnonic interference devices, and holographic tensor-network codes (Ray et al., 21 Aug 2025).

1. Conceptual scope and relation to declarative memory

Declarative memory, in the cognitive sense, is explicit memory that can be “declared” in words or languages. It comprises episodic memory, which concerns events together with their temporal and spatial context, and semantic memory, which concerns factual or conceptual knowledge independent of specific episodes. Zhang’s review emphasizes that these two forms are dissociated yet closely related, with episodic learning associated primarily with hippocampal mechanisms and semantic learning with neocortical abstraction and consolidation (Zhang, 6 Feb 2026).

Within that broader cognitive frame, HDM names a particular computational strategy: declarative items are not stored as isolated symbolic records but as overlapping, distributed traces that can be bound, superposed, and later reconstructed from partial cues. In the ACT-R line, this strategy is explicitly realized as a vector-symbolic replacement for standard chunk storage, with the stated advantages of scalability, architecturally defined similarity, and chunk recall without storing the actual chunk as a discrete structure (Ray et al., 21 Aug 2025).

The literature does not present a single canonical HDM formalism. Instead, it uses the term explicitly in some settings and as an interpretive bridge in others. This produces a family resemblance across otherwise different implementations: distributed encoding, content-addressable retrieval, reconstructive recall, and tolerance for partial or noisy cues.

2. Representational primitives: superposition, binding, and matched filtering

The most direct mathematical basis for HDM is the Holographic Reduced Representation framework used in the ACT-R adaptation. Each basic symbol—slot name, value, or word—is assigned a high-dimensional vector, typically sampled from a zero-mean normal distribution. A chunk

$c = \{s_1{:}v_1, s_2{:}v_2, \dots, s_n{:}v_n\}$

is encoded as

$C = \sum_{i=1}^n s_i \circledast v_i,$

where $\circledast$ denotes circular convolution. Superposition is ordinary vector addition, unbinding uses correlation with an approximate inverse, and retrieval relies on vector similarity, typically dot product or cosine similarity (Ray et al., 21 Aug 2025).

This construction makes declarative storage inherently distributed. Slot:value bindings share the same fixed-width vector space, and multiple associations can occupy that space simultaneously. Retrieval is therefore reconstructive: a cue vector is unbound from a memory trace, and the resulting noisy approximation is cleaned up by nearest-neighbor comparison against the token vocabulary. Partial matching is not an added exception rule; it emerges from geometry in the vector space.

A separate but related idiom appears in HDRAM for tokenized LLMs. There, hypertokens are symbolic identifiers derived from linear block codes and mapped into latent space by a phase-preserving projection

$\Phi: \mathbb{F}_2^n \rightarrow \mathbb{R}^d.$

The basic memory operation is

$$\text{Memory}(h_j) = \arg\max_i \langle \Phi(c_i), h_j \rangle,$$

so retrieval is a matched-filter or despreading operation in a latent space treated as a spread-spectrum channel. ECC grammar, bifix-free coding, symbolic RIP, compressed sensing, and Grover-style search are used to stabilize K:V and V:K retrieval without changing transformer architecture (Augeri, 2 Jun 2025).

Taken together, these two strands define the core HDM logic. Declarative items are encoded as distributed combinations of address-like and content-like components; recall proceeds by cue projection, correlation, or interference; and similarity is architectural rather than post hoc.

3. HDM in ACT-R and the problem of whole-chunk recall

The ACT-R adaptation is the clearest full-system instantiation of HDM. It adapts HDM to Lisp ACT-R so that extant models designed for standard Declarative Memory can run with HDM without major changes. Goal chunks remain in standard DM, while non-goal declarative knowledge can be stored in HDM through ACT-R-compatible requests. Lisp-side commands call Python HDM routines through the ACT-R 7 Python connection, and familiar interfaces such as (sgp) and (dm) are preserved in vector-based form (Ray et al., 21 Aug 2025).

Its most technically distinctive contribution is a mechanism for retrieving an entire chunk from only vector representations of tokens. Each memory vector $m$ is paired with a time-memory vector $mt$. Time is discretized by chunk index and encoded by a 320-dimensional oscillator-based vector $\mathbf{T}(t)$ built from 15 oscillator functions. Because these oscillator vectors are not HRR-compatible by default, the system converts them into an HRR-compatible time representation by fractional binding:

$$\widetilde{\mathbf{T}(t)} = \mathcal{F}^{-1} \left\{ \sum_{l=1}^{320} \mathcal{F}(e_{t_l})^{\mathbf{T}(t)_l} \right\}.$$

When a chunk is stored, the time HRR is bound to the chunk’s summed slot:value bindings and added into the time-memory vectors. For a partial request

$Q = \sum_{c=1}^{m} s'_c \circledast v'_c,$

the system computes

$C = \sum_{i=1}^n s_i \circledast v_i,$0

matches reconstructed time HRRs against actual time vectors, and then uses multiple-unknown request chaining to reconstruct missing slot values and thus the full chunk (Ray et al., 21 Aug 2025).

This mechanism changes the role of ACT-R declarative retrieval. Standard DM returns a single symbolic chunk or none, with optional mismatch penalties. HDM instead permits graded retrieval over distributed traces, with chunk structure reconstructed from the interaction of slot:value bindings, time context, and similarity clean-up. The adaptation also adds a text-processing pipeline: (preprocess-text) removes stopwords and segments raw text into sentences, (read-corpus-hdm) loads those sentences into HDM, and (dm) prints unique values together with a 2D visualization of the vector space. The paper presents this as especially relevant to instance-based learning, where large numbers of situation–action–utility instances must be compared by similarity rather than exact symbolic identity (Ray et al., 21 Aug 2025).

4. Transformer-centered and external-memory variants

A separate strand of research treats declarative memory in LLMs as an explicitly writable layer above frozen model weights. In In-Memory Learning, the model parameters are fixed and learning occurs through natural-language notes $C = \sum_{i=1}^n s_i \circledast v_i,$1 inserted into context. Action selection is written as

$C = \sum_{i=1}^n s_i \circledast v_i,$2

and learning is decomposed into inference, induction, and revision. On a 4-class benchmark built from 10 adjective-based dimensions, the system iteratively summarizes trajectories into new rules and merges them into existing notes, producing accuracy curves over learning steps without gradient updates or human-labeled data (Wang et al., 2024).

HDRAM takes a more explicitly holographic route. It treats transformer latent space as a spread-spectrum communication channel and uses hypertokens as phase-coherent memory addresses for key–value storage, reverse lookup, and Grover-style search. The proposal combines classical ECC structure, holographic superposition in a “holobasis,” and quantum-inspired amplitude steering. Reported empirical effects include “precision recall extended by ~2x or more” and a “False activation rate decreased by 65%,” together with support for exact recall windows, associative retrieval, and in-context algorithmic operations (Augeri, 2 Jun 2025).

The Holographic Knowledge Manifold externalizes knowledge even more strongly. It defines a manifold

$C = \sum_{i=1}^n s_i \circledast v_i,$3

with concept nodes, probabilistic edges, and fractal levels. Its four-phase pipeline consists of probabilistic entanglement, fractal quantization, holographic sampling and training integration, and dynamic diffraction chipping. Knowledge is updated through Fourier-space interference,

$C = \sum_{i=1}^n s_i \circledast v_i,$4

followed by RL-pruned EWC. The paper reports 0% catastrophic forgetting, 3× compression, 67% storage savings, 100% integration, 1% memory growth per update, support for over 1,020 updates, and 53% training time reduction on its evaluation setup (Arndt, 3 Sep 2025).

Across these systems, declarative memory appears in three different guises: editable natural-language rules, ECC-structured latent addresses, and an external holographic manifold. Their commonality lies less in surface representation than in the use of distributed codes and reconstructive access.

5. Physical and quantum realizations

The magnonic holographic memory device provides a concrete hardware example of how an HDM-like system could be physically instantiated. The prototype is an eight-terminal device based on a $C = \sum_{i=1}^n s_i \circledast v_i,$5 YIG waveguide matrix with four junction magnets and edge micro-antennas that generate and detect spin waves. Input information is encoded in spin-wave phases, while output is the amplitude of the inductive voltage produced by interfering waves. Different magnet configurations produce different interference landscapes and therefore different mappings from input phase patterns to output voltage. Recognition is implemented by comparing the output voltage against a reference threshold $C = \sum_{i=1}^n s_i \circledast v_i,$6; changing the threshold adjusts the tolerated Hamming distance for similar patterns. The paper also notes an estimated data storage density up to $C = \sum_{i=1}^n s_i \circledast v_i,$7 and stresses CMOS compatibility, but it likewise discusses fabrication constraints, damping, coherence, and scaling limits (Kozhevnikov et al., 2014).

The quantum “H-code” offers a different template. Built from an absolutely maximally entangled four-qutrit state, it generates a tensor-network code on a triangular lattice in which the boundary determines the bulk. The defining local rule is the neutralization condition

$C = \sum_{i=1}^n s_i \circledast v_i,$8

and the full code is an even superposition over boundary configurations whose corresponding bulk product states are orthogonal. On tori of size $C = \sum_{i=1}^n s_i \circledast v_i,$9, any two codewords differ in at least $\circledast$0 positions, the code decomposes into three sectors labeled by a global charge $\circledast$1, and the topological entanglement entropy is $\circledast$2. Latorre and Sierra present this as a natural quantum memory; in HDM terms, it is a boundary-indexed, holographically organized store with very large codeword separation (Latorre et al., 2015).

These realizations are not identical to the ACT-R or LLM variants, but they sharpen the meaning of “holographic.” In the magnonic case, holography is literal interference of propagating waves. In the H-code, it is an exact boundary-to-bulk mapping with strong entanglement structure. Both provide technical models for distributed declarative storage that is global rather than pointwise.

6. Limits, misconceptions, and open problems

The literature discourages three common simplifications. First, “holographic” is not confined to optical storage: the cited work spans spin-wave hardware, HRR-based vector-symbolic memory, transformer latent-space overlays, tensor-network quantum codes, and external manifolds (Kozhevnikov et al., 2014, Ray et al., 21 Aug 2025, Augeri, 2 Jun 2025, Latorre et al., 2015). Second, declarative access does not require literal symbolic chunk objects: ACT-R HDM reconstructs chunks from traces, In-Memory Learning stores editable notes, and HKM keeps knowledge in an external manifold rather than only in model weights (Wang et al., 2024, Arndt, 3 Sep 2025). Third, associative retrieval is not equivalent to exact or noise-free recall: these systems explicitly discuss crosstalk, thresholds, approximate inverses, local minima, entropic limits, phase coherence, and growth-control tradeoffs (Augeri, 2 Jun 2025, Kozhevnikov et al., 2014).

Open problems are correspondingly diverse. The ACT-R adaptation remains preliminary and identifies better time-context representations as a major direction for improving chunk reconstruction, while the LLM note-revision framework reports instability under weak momentum and a tendency to become trapped in copy-biased local minima (Ray et al., 21 Aug 2025, Wang et al., 2024). HDRAM points to interference, holobasis optimization, anisotropic embeddings, and “entropic limits” in latent-space coherence; HKM notes that its current implementation is limited to 2,997 nodes and assumes that DDPM-like noise, RL-pruned EWC, and manifold-aware attention will scale to much larger systems (Augeri, 2 Jun 2025, Arndt, 3 Sep 2025).

At the systems level, Zhang’s episodic–semantic review highlights a further unresolved question: how fast index-like episodic storage, slow semantic abstraction, and replay-based consolidation should be coupled in a biologically grounded declarative architecture. The review presents hippocampal indexing, neocortical abstraction, and impairment-sensitive learning dynamics as central, but also notes that its basic unit is not neuron-like, that demonstrated semantic content is largely numerical, and that the system processes sequences only (Zhang, 6 Feb 2026).

HDM is therefore best understood not as a single settled architecture but as a research program. Its recurring thesis is that declarative memory can be made scalable, content-addressable, and similarity-sensitive when symbols, facts, or episodes are encoded in distributed holographic form and recalled by reconstruction rather than direct lookup.