Memory Systems in Cognitive Neuroscience

• Memory systems in cognitive neuroscience are parallel, dissociable processes responsible for encoding, storing, and retrieving information across sensory, working, and long-term domains.
• Studies integrate lesion experiments, neuroimaging, and computational models to map dynamic interactions among hippocampal, cortical, and subcortical regions.
• Insights into these systems inform clinical interventions and bio-inspired models, advancing strategies to address cognitive impairments and optimize memory performance.

Memory systems in cognitive neuroscience denote an ensemble of parallel, partially dissociable processes that encode, store, and retrieve information over diverse timescales, content domains, and neural architectures. The classical dichotomy between declarative (explicit, hippocampus‐dependent) and non‐declarative (implicit, striatal/cerebellar/amygdalar/cortical) systems has evolved into a multi-component taxonomy encompassing sensory, working, episodic, semantic, procedural, priming, conditioning, and nonassociative memories. These systems interact dynamically, exhibit distinctive neurophysiological and computational signatures, and undergo distinct lifecycle phases: encoding, consolidation, retrieval, updating (reconsolidation), and forgetting. The following sections synthesize major theoretical constructs, anatomical mappings, computational models, emergent mechanisms, and current controversies in the field.

1. Taxonomy and Definitions of Memory Systems

Memory systems are classified according to duration, awareness, and content. Sensory memory refers to ultra-short (<500 ms) modality-specific buffers in primary cortical areas (iconic memory in V1, echoic in A1). Working memory comprises limited-capacity (≈4–9 chunks), time-limited (seconds–minutes) active maintenance and manipulation, primarily dorsolateral/ventrolateral PFC and parietal cortex, supporting online cognition, reasoning, and planning (Liang et al., 29 Dec 2025, Pastor, 2020, Fox et al., 2016).

Long-term memory divides into declarative and non-declarative:

• Declarative (explicit) memory: Episodic (autobiographical, spatiotemporal context, “mental time travel”) and semantic (facts, concepts, vocabulary, context-free), both depend initially on hippocampus and medial temporal lobe (MTL), maturing into distributed neocortical networks (Pastor, 2020, Fox et al., 2016, Liang et al., 29 Dec 2025).
• Non-declarative (implicit) memory: Procedural (motor skills, habits; striatum, cerebellum), priming (prior exposure; sensory/association cortex), classical conditioning (amygdala for emotional, cerebellum for motor), and nonassociative forms (habituation, sensitization; reflex/spinal circuits).

The table below summarizes primary memory system types, their key substrates, and major behavioral correlates:

System	Primary Neural Substrates	Behavioral Correlates
Sensory	V1/A1, modality-specific cortex	Iconic/echoic buffer
Working	DLPFC, parietal cortex, sensory cortices	Short-term holding, manipulation
Episodic	Hippocampus, MTL, neocortex	Event recall, spatiotemporal detail
Semantic	Neocortex, lateral temporal, IPL, vmPFC	Fact/concept knowledge
Procedural	Striatum, cerebellum, motor cortex	Skills, habits, S–R learning
Priming	Modality-specific neocortex	Facilitated recognition
Conditioning	Amygdala (emotional); cerebellum (motor)	CRs, emotional learning

2. Neuroanatomical Substrates and Dissociation Evidence

Distinct memory systems are underpinned by discrete anatomical circuits. Lesion and imaging studies reveal selective dependence:

• Working memory: Sustained delay-period activity in PFC; lesions in dorsolateral PFC impair manipulation and maintenance; parietal cortex holds spatial/feature content (Pastor, 2020, Fox et al., 2016).
• Episodic/Semantic: Hippocampal and MTL lesions (e.g. H.M.) cause anterograde amnesia for facts/events but spare procedural/priming; neocortical redistribution supports remote/semantic memory (Pastor, 2020).
• Procedural: Striatal lesions/Parkinson’s compromise habit learning but leave declarative recall unaffected (double dissociation) (Fox et al., 2016).
• Conditioning: Delay eyeblink (cerebellar interpositus), emotional CRs (amygdala); hippocampus not required (Pastor, 2020).

Functional MRI shows phase-dependent recruitment; early learning involves MTL, then progressively shifts toward striatum for habitual tasks (Pastor, 2020). Modulatory systems (cholinergic, noradrenergic, dopaminergic) further regulate encoding and consolidation (Fox et al., 2016, Liang et al., 29 Dec 2025).

3. Synaptic, Circuit, and Network Mechanisms

At the synaptic level, memory formation relies on Hebbian plasticity: $\Delta w_{ij} = \eta\,x_i\,x_j$ and spike-timing dependent plasticity (STDP): $$\Delta w = \begin{cases} A_+\,e^{-{\Delta t}/{\tau_+}} & \text{if } \Delta t > 0 \ -A_-\,e^{{\Delta t}/{\tau_-}} & \text{if } \Delta t < 0 \end{cases}$$ Memory traces undergo dual-phase dynamics: transient reverberating activity (short-term) and subsequent long-term synaptic modification [(Pastor, 2020); Hebb, 1949]. Single-cell and population timescales diverge—recent RNN models demonstrate optimal working-memory performance when fast and slow neurons (with τ_fast = 1, τ_slow = 10) are combined, with slow units causally essential for stable memory and fast units encoding transient inputs (Kurikawa, 9 Jun 2025).

Astrocytes provide a complementary, graded, and time-limited analog buffer, modulating local synaptic transmission through Ca²⁺-dependent gliotransmission (transients of ≈1–10 s), supporting activity-silent short-term retention and robust, noise-resistant recall (Tsybina et al., 2021).

Large-scale network models exploit bidirectional hippocampal–cortical replay for systems consolidation (gradual shift from HPC to neocortex) and reconsolidation (labile updating and re-indexing after retrieval), as formalized with AMPAR trafficking and depotentiation rules (Helfer et al., 2017).

4. Dynamic Organization, Computational Models, and Predictive Representations

Theoretical frameworks such as Complementary Learning Systems (CLS) posit fast, pattern-separated episodic encoding in the hippocampus and slow statistical extraction in neocortex. This division resolves the stability–plasticity dilemma, avoiding catastrophic interference of episodic facts with semantic regularities (Beton et al., 14 Jan 2026).

Memory spaces, as topological schemas, are constructed from hippocampal cell-assembly coactivity, formalized as finite Alexandrov topologies M = (X, τ) and nerve simplicial complexes N(X, τ), supporting persistent homology analysis (Betti numbers β_k reflect loops/holes in spatial and nonspatial domains) (Babichev et al., 2017).

Multiscale predictive representations—successor representations (SR) parameterizing discounted future occupancy: $$\mathrm{SR}(s,s') = \mathbb{E}_\pi\Bigl[\sum_{t=0}^\infty \gamma^t\,\mathbf{1}_{s_t=s'} \mid s_0=s\Bigr]$$ —are hierarchically encoded in the posterior–anterior hippocampus and PFC, facilitating episodic recall, abstraction, and efficient planning (Momennejad, 2024). Successor matrices at multiple γ scales enable both fine-grained episodic memory and high-level schemas.

Recurrent, gated network models (e.g. ORGaNICs) capture working memory as the emergent interaction of oscillatory, integrator, and gating motifs, supported by biophysically mapped circuit architecture (dendritic compartmentalization, thalamocortical loops) (Heeger et al., 2018).

5. Lifecycle Processes: Encoding, Consolidation, Retrieval, Reconsolidation, Forgetting

Lifecycle phases are orchestrated by both intra- and inter-regional mechanisms:

• Encoding: Selective synaptic potentiation in HPC/neocortex; neuromodulatory tagging for priority (Liang et al., 29 Dec 2025).
• Consolidation: Time-dependent systems-level transfer: rapid Hebbian growth in HPC, slow accumulation in ACC/neocortex by replay (model: AMPAR trafficking and depotentiation) (Helfer et al., 2017).

  $\frac{dM_c}{dt} = \alpha\,M_h(t) - \beta\,M_c(t)$

• Retrieval: Hippocampal pattern completion reactivates distributed neocortical traces; sharp-wave ripples coordinate reactivation (Liang et al., 29 Dec 2025).
• Reconsolidation: Post-retrieval memories become transiently labile (CI-AMPAR depletion), requiring HPC-mediated replay for restabilization; interference or lesion during this window causes memory loss (Helfer et al., 2017).
• Forgetting: Passive decay (exponential, $S(t) = S_0 \exp(-t/\tau)$), interference, or active neuromodulatory downregulation (e.g. endocannabinoids) (Liang et al., 29 Dec 2025, Fox et al., 2016).

6. Experimental Paradigms, Measurement, and Clinical Implications

Memory systems are interrogated via specific paradigms:

• Sensory memory: Partial-report tasks (iconic/echoic span).
• Working memory: Digit span, n-back, delayed match-to-sample; metrics: span, accuracy, reaction time (Liang et al., 29 Dec 2025).
• Long-term memory: Free/cued recall, recognition, paired-associates, DNMS; metrics: capacity, fidelity, decay, signal-detection index d′.

Lesion studies, tES/tDCS, optogenetics, two-photon imaging, and pharmacological manipulations parse system dissociation, consolidation routes, and storage substrates (Fox et al., 2016, Pastor, 2020, Tsybina et al., 2021).

Experimental predictions include enhanced delay-period impairment after selective optogenetic suppression of slow WM units, or astrocytic Ca²⁺ transients time-locked to memory item loading, and schema formation via topological quotient-space reduction (Kurikawa, 9 Jun 2025, Babichev et al., 2017, Tsybina et al., 2021).

Clinical interventions target specific substrates (cholinesterase inhibitors for semantic/episodic; L-DOPA for procedural; tDCS for WM) with differing safety/fairness/authenticity trade-offs by system (Fox et al., 2016).

7. System Interactions, Controversies, and Translational Relevance

Systems can cooperate (e.g. episodic+semantic integration) or compete (stress-induced hippocampal→striatal shift). Independence and interaction mechanisms remain debated—e.g., sequence learning’s dependence on the hippocampus vs. striatum (Pastor, 2020). Multiple-trace theory contests the standard consolidation model, positing permanent HPC involvement in episodic detail.

CLS theory is foundational for bridging neuroscience and neuro-inspired AI: knowledge object architectures paired with neural weights (mirroring hippocampal fast and neocortical slow learning) mitigate interference and stability gaps seen in artificial agents (Beton et al., 14 Jan 2026).

Advances in predictive representation theory, topological modeling, and astrocytic mechanisms extend the memory-systems view into planning, abstraction, and analog working-memory storage. Adaptive agent architectures extracting principles from neurobiology—multi-timescale storage, content-addressable retrieval, prioritized replay, controlled forgetting, and dynamic updating—approach human levels of flexibility and robustness (Liang et al., 29 Dec 2025, Momennejad, 2024).

Current research focuses on further elucidating cooperation/competition dynamics, refining computational and topological frameworks for memory integration, and leveraging system-specific interventions for enhancement and remediation. Biological and artificial memory systems increasingly share design principles shaped by the neurobiological mosaic of human cognition.