Multiple Memory Systems (MMS)

• Multiple Memory Systems (MMS) is defined as a framework where memory is composed of specialized, interacting systems that differ in how they encode, store, and retrieve information.
• Empirical evidence such as double dissociations in clinical studies and specific behavioral tasks supports the separation of working, episodic, semantic, and procedural memory.
• Computational models and AI architectures inspired by MMS demonstrate improved performance through distinct processing pathways, with measurable gains in metrics like F1 scores and BLEU ratios.

Searching arXiv for the cited MMS-related papers to ground the article and confirm details. arXiv search: (Ren et al., 2023) "Rethinking Memory Profiling and Migration for Multi-Tiered Large Memory Systems" arXiv search: (Pastor, 2020) "Memory systems of the brain" arXiv search: (Fox et al., 2016, Zhang et al., 21 Aug 2025, Zhang, 6 Feb 2026, Mao et al., 13 Mar 2026, You et al., 19 Mar 2026, Li et al., 6 Jul 2026) Multiple Memory Systems (MMS) denotes the view that memory is not a unitary faculty but a set of specialized, interacting systems with distinct operating principles, representational content, and neural substrates. In the neurocognitive literature, MMS is typically organized around declarative and non-declarative divisions and further differentiated into systems such as working, episodic, semantic, and procedural memory; in recent AI research, the same idea has been translated into architectures that separate short-term context, episodic traces, semantic commitments, retrieval mechanisms, and control processes rather than treating memory as undifferentiated conversation history (Pastor, 2020).

1. Definition, scope, and historical emergence

MMS is defined by converging behavioral, computational, and neurobiological criteria. A memory “system” is identified jointly by behavioral signatures, operating principles, neural substrates, and representational content. On this view, memory systems differ in how they acquire information, how retrieval is expressed, what is stored, and which circuits are necessary for the relevant behavior. The rationale is empirical rather than purely taxonomic: selective impairments, double dissociations, and distinct learning phenomena that survive hippocampal amnesia all indicate that mnemonic function does not reduce to a single mechanism (Pastor, 2020).

The modern MMS framework emerged from mid- to late-twentieth-century neuropsychology and behavioral neuroscience. H.M.’s bilateral medial temporal lobe resection provided the canonical dissociation: severe anterograde declarative amnesia with spared immediate memory and preserved skill learning. Tulving’s episodic/semantic distinction and Squire’s declarative/nondeclarative framework then supplied the dominant conceptual vocabulary. The resulting model presents memory as a consortium of systems that can compete for behavioral control, cooperate during encoding and consolidation, and fail selectively under lesion or disease (Pastor, 2020).

A central consequence of this framework is that the phrase “memory enhancement” or “memory impairment” is too coarse unless the target system is specified. The human enhancement literature explicitly argues that at least four major systems—working, procedural, episodic/autobiographical, and semantic memory—have largely dissociable neural substrates and therefore differ in feasible interventions, risks, and ethical significance (Fox et al., 2016).

2. Taxonomy of major and minor memory systems

The core taxonomy distinguishes declarative memory, which supports conscious recollection of episodes and facts, from non-declarative memory, which is expressed through performance rather than recollection. Declarative memory includes episodic and semantic memory and depends critically on the hippocampus and surrounding medial temporal lobe structures for encoding. Non-declarative memory includes procedural skills and habits, priming, classical conditioning, and nonassociative learning, with canonical substrates in the striatum, cerebellum, amygdala, neocortex, and reflex pathways (Pastor, 2020).

System	Characteristic content / timescale	Canonical substrates
Working memory	Short-term manipulation; “the mind’s blackboard”; ~1 minute	Prefrontal cortex
Procedural or skill memory	Skills and habits; gradual acquisition over practice	Striatum/basal ganglia and cerebellum
Episodic or autobiographical memory	Specific events in time and place; remote consolidation over years	Hippocampus and adjacent MTL for formation; remote memories in neocortex
Semantic memory	Facts and concepts independent of acquisition context	Hippocampus and adjacent MTL for formation; remote knowledge in neocortex

This four-system summary is not exhaustive. Minor systems identified within the MMS framework include priming, perceptual representation systems, classical conditioning, habituation, and sensitization. Classical conditioning itself is subdivided by substrate: emotional conditioning is linked to the amygdala, whereas motor/reflex conditioning is linked to the cerebellum. Priming is described as modality-specific facilitation supported by active neocortical circuits, and habituation or sensitization is associated with reflex pathways (Fox et al., 2016).

The taxonomy is functional as well as anatomical. Working memory supports span tasks, n-back, verbal and visuospatial manipulation, language comprehension, and reasoning. Procedural memory is indexed by mirror tracing, probabilistic classification, motor sequence learning, and automaticity measures. Episodic memory is assessed by autobiographical recall, recognition, vividness, and recent-versus-remote tests. Semantic memory is assessed by factual recall, recognition, vocabulary, and category knowledge (Fox et al., 2016).

3. Dissociation, interaction, and systems-level dynamics

The evidential core of MMS is dissociation. Bilateral medial temporal lobe damage abolishes new episodic and semantic learning while leaving working memory, priming, conditioning, and procedural skill learning largely intact. Conversely, Parkinson’s disease, which compromises striatal function, can impair probabilistic habit learning while preserving declarative recall of the training episode. The literature therefore uses double dissociations, not merely single deficits, to justify distinct systems (Fox et al., 2016).

A standard example is the contrast between medial temporal lobe amnesic patients and Parkinson’s patients in probabilistic habit classification. Medial temporal lobe amnesics can learn the habit without recalling the training episodes, whereas Parkinson’s patients recall the session but fail to acquire the habit. Comparable dissociations appear in eyeblink conditioning: delay conditioned responses can be acquired without forebrain involvement, whereas trace procedures require forebrain and hippocampal participation. These patterns support the claim that different tasks recruit different memory architectures rather than a common store with variable difficulty (Pastor, 2020).

Systems are not isolated. They operate in parallel and may compete or cooperate. Training duration can shift control from medial temporal lobe-dependent strategies to striatal strategies; stress can bias behavior toward striatal learning; semantic scaffolding can facilitate episodic encoding; and amygdala-mediated arousal can modulate consolidation across systems. MMS therefore combines dissociation with interaction: the systems are specialized, but behavior emerges from coordinated and sometimes competitive deployment (Pastor, 2020).

Several formal motifs recur across this literature. Hebbian plasticity is used as a compact encoding principle,

$\Delta w_{ij} = \eta x_i y_j,$

while reinforcement-based procedural learning is often formalized by temporal-difference error,

$\delta_t = r_t + \gamma V(s_{t+1}) - V(s_t).$

Complementary Learning Systems (CLS) further posits fast episodic encoding in hippocampus and slow interleaved learning in neocortex, reconciling rapid acquisition with stable generalization (Pastor, 2020).

4. Computational models of declarative MMS

Computational work on episodic and semantic memory has largely concentrated on hippocampus–neocortex interaction. The reviewed models share a common asymmetry: fast hippocampal binding or indexing supports episodic acquisition, while slower cortical learning supports semantic abstraction and long-term stabilization. This division is central to the Binding model, TraceLink, and CLS, though the architectural details differ (Zhang, 6 Feb 2026).

The Binding model of Alvarez and Squire represents two reciprocally connected cortical areas and a small medial temporal lobe module. The hippocampal-like component rapidly stores indices that bind cortical patterns; repeated reactivation slowly strengthens direct cortex–cortex associations until retrieval becomes independent of the index. TraceLink introduces a trace system, a link system, and a modulatory system, with fast trace–link learning and slow trace–trace consolidation. CLS implements the same asymmetry with rapid hippocampal storage and slow interleaved cortical training to avoid catastrophic interference (Zhang, 6 Feb 2026).

These models are used to simulate classic impairments. Retrograde amnesia is reproduced by lesioning the hippocampal or link component after variable consolidation intervals, yielding a Ribot gradient in which older memories are better preserved. Anterograde amnesia follows from disabling fast binding before learning. Developmental amnesia is modeled as impaired sequential association encoding with preserved recognition of individual elements and relatively preserved semantic knowledge. Dense amnesia is modeled as a state in which new semantic learning is possible only after very long repetition schedules, matching the observation that patients with severe hippocampal damage can acquire some semantic knowledge over years but not under ordinary laboratory exposure (Zhang, 6 Feb 2026).

The chapter also reviews symbolic and hybrid cognitive architectures. ACT-R includes a symbolic declarative memory but does not explicitly model episodic memory. CLARION uses dual-level explicit and implicit representations, with decision policy given by

$$p(a \mid x) = \frac{\exp(Q(x,a)/\tau)}{\sum_{a'} \exp(Q(x,a')/\tau)}.$$

Zhang’s multi-leveled system attempts a more explicit episodic–semantic decomposition by combining a hippocampal-like episodic storage module, semantic symbol and representation subsystems, and a “dreaming” mechanism that randomly replays stored segments to drive semantic abstraction (Zhang, 6 Feb 2026).

A notable implication of this line of work is that declarative MMS is not only a lesion-based taxonomy but also a design principle for learning systems: rapid binding, replay, consolidation, and abstraction are treated as distinct algorithmic operations rather than as side effects of a single learner. This suggests that the biological MMS framework has direct computational content.

5. MMS in long-lived AI agents and personalized language-model systems

Recent AI research has explicitly adopted MMS terminology to address the inadequacy of storing long interaction histories as raw transcripts or simple vector memories. One line of work constructs multiple long-term memory fragments from short-term interactions. In “Multiple Memory Systems for Enhancing the Long-term Memory of Agent,” each dialogue round is transformed into keywords, cognitive perspectives, episodic memory, and semantic memory; retrieval memory units use $(M_{key}, M_{short}, M_{cog}, M_{epi})$, contextual memory units use $(M_{key}, M_{short}, M_{cog}, M_{sem})$, and the two are linked one-to-one so that retrieval is query-aligned while generation is knowledge-aligned. On LoCoMo, MMS improves over NaiveRAG, MemoryBank, and A-MEM; for GPT-4o the reported average rises from A-MEM’s F1 $=23.37$ and BLEU-1 $=19.65$ to MMS F1 $=30.54$ and BLEU-1 $=25.74$, with average memory-content tokens per query of 744 and average latency of 1.309 seconds (Zhang et al., 21 Aug 2025).

A second line of work emphasizes control and revision rather than fragment design alone. MRMS organizes memory along a representational axis—structured records, vector embeddings, and graph relations—and a temporal axis—short-term traces, medium-term abstractions, and long-term semantic commitments. Memory objects are explicitly typed as

$$m_i = (\ell_i, q_i, v_i, r_i, a_i, \sigma_i, c_i, h_i),$$

and pre-generation selection combines semantic similarity, recency, task importance, contradiction penalties, and graph adjudication. The resulting selection score is

$\delta_t = r_t + \gamma V(s_{t+1}) - V(s_t).$0

with inclusion gated by support/contradiction signals and active superseders. In controlled long-lived scenarios comprising 800 synthetic tasks, Full MRMS reaches 98.8% overall pre-generation accuracy with 95% CI 98.0–99.4, and evidence attribution is explicitly tested at 95% overall (Li et al., 6 Jul 2026).

D-Mem reformulates MMS as a dual-process architecture for LLM agents. Its fast path, Mem0$\delta_t = r_t + \gamma V(s_{t+1}) - V(s_t).$1, uses compressed semantic retrieval from a vector database; its slow path, Full Deliberation, performs query-guided reading over the raw conversation history. The escalation decision is made by a Multi-dimensional Quality Gating policy:

$\delta_t = r_t + \gamma V(s_{t+1}) - V(s_t).$2

where relevance, faithfulness and consistency, and completeness must all pass for the fast path to be accepted. On LoCoMo with GPT-4o-mini, Quality Gating achieves F1 $\delta_t = r_t + \gamma V(s_{t+1}) - V(s_t).$3, compared with Mem0$\delta_t = r_t + \gamma V(s_{t+1}) - V(s_t).$4 at 51.2 and Full Deliberation at 55.3, thereby recovering 96.7% of Full Deliberation’s performance while using 12,681 tokens and 8.03 seconds, versus 35,435 tokens and 23.73 seconds for Full Deliberation (You et al., 19 Mar 2026).

CoMAM extends MMS to a collaborative multi-agent setting for personalized memory systems. It decomposes the pipeline into an Extraction Agent for fine-grained memory, a Profile Agent for coarse-grained user profile memory, and a Retrieval Agent for memory selection and response generation. The agents are optimized jointly as a sequential MDP with local rewards for coverage, abstraction rationality, and retrieval precision, plus a global reward for exact-match answer accuracy. Adaptive credit assignment is based on ranking consistency measured by NDCG, and the integrated reward is

$\delta_t = r_t + \gamma V(s_{t+1}) - V(s_t).$5

On PersonaMem, the reported accuracies for the Qwen family are 0.64, 0.70, and 0.66 at 32K, 128K, and 1M contexts, outperforming leading baselines; joint training also reduces convergence steps relative to independent optimization (Mao et al., 13 Mar 2026).

Across these systems, a common design principle is explicit separation among memory types, access routes, and control processes. This suggests that contemporary AI MMS is less a direct copy of the human taxonomy than an engineering reinterpretation of the same problem: rapid capture, selective retrieval, revision under contradiction, and durable commitments require heterogeneous stores and heterogeneous control logic rather than a single memory buffer.

6. Practical implications, controversies, and terminological ambiguity

In human neuroscience and bioethics, MMS has immediate practical consequences because intervention strategies differ by target system. Working memory is associated with prefrontal stimulation and arousal-modulating drugs; procedural memory raises special concern because dopaminergic modulation intersects reward circuitry; episodic/autobiographical memory is linked to deep-brain stimulation and reconsolidation-related interventions; semantic memory is often targeted indirectly through cholinergic therapies in neurodegenerative disease. The ethical profile also varies by system: fairness and coercion are especially salient for semantic, working, and procedural enhancement, whereas authenticity and identity concerns are concentrated in autobiographical memory and therapeutic forgetting (Fox et al., 2016).

The framework remains well supported but not uncontroversial. Some critiques argue for semi-dissociable processes rather than sharply bounded systems, and some taxonomy disputes concern how much to “split” or “lump” phenomena such as implicit memory, short-term memory, and long-term memory. Even within the mainstream view, semantic and episodic memory are partly coupled during encoding, and frontal control processes permeate declarative retrieval. MMS is therefore presented not as an absolute partition but as a robust model of interacting specialized systems (Pastor, 2020).

A parallel controversy in AI is whether longer prompt windows or retrieval-only pipelines can substitute for a genuine memory system. Recent work argues that prompt extension cannot enforce boundaries, provenance, or supersession, and that retrieval-only systems surface stale or out-of-scope items when similarity remains high. This suggests that the defining issue in AI MMS is not merely storage capacity but controlled influence: memory must be authorized, scoped, revised, and selectively exposed (Li et al., 6 Jul 2026).

The acronym “MMS” also has a separate meaning in computer architecture. In “Rethinking Memory Profiling and Migration for Multi-Tiered Large Memory Systems,” MMS denotes multi-tiered large memory systems: multi-terabyte systems comprising more than two memory tiers with differing latency and bandwidth characteristics. That paper introduces MTM, an application-transparent page management system built on overhead-aware profiling, a universal page migration policy, and huge page awareness, and reports performance gains of up to 42% and 17% on average over seven state-of-the-art solutions on big-data workloads with working sets from hundreds of GB to 1 TB (Ren et al., 2023). The terminological overlap is significant because it separates two unrelated research programs: one on multiple memory systems in cognition and AI, the other on heterogeneous hardware memory hierarchies.

Taken together, these literatures define MMS as a framework for heterogeneity. In neuroscience, heterogeneity concerns substrates, representations, and behavioral expression. In AI, heterogeneity concerns storage formats, temporal layers, retrieval routes, revision policies, and agent roles. The shared thesis is that memory becomes tractable only when its multiplicity is modeled explicitly rather than absorbed into a single monolithic store.