Long-Term Memory in Biology and AI

Updated 1 July 2025

Long-Term Memory is the capacity to encode, store, and retrieve information over extended timescales in both biological neural circuits and artificial systems.
In neuroscience, memory persists through synaptic modifications like LTP and LTD, forming robust engram networks that support enduring behavioral change.
In artificial intelligence, hybrid parametric and non-parametric architectures enable scalable retrieval and continual learning for adaptive problem-solving.

Long-term memory is the capacity of systems—biological or artificial—to encode, store, and retrieve information over extended timescales, supporting knowledge accumulation, enduring behavioral change, cross-domain reasoning, and adaptive responses. In neuroscience, it underpins the persistence of synaptic modifications, behavioral competencies, and engram formation over months to years. In artificial intelligence, long-term memory refers to architectures, algorithms, and hardware that enable scalable, robust storage and retrieval of knowledge beyond transient computational windows, enabling continual or lifelong learning, compositional reasoning, and personalized adaptation.

1. Synaptic and Computational Foundations of Long-Term Memory

Early theoretical and empirical work established that memory traces are physically realized as lasting changes in neural circuits, notably through synaptic plasticity. Long-Term Potentiation (LTP) and Long-Term Depression (LTD) are central molecular mechanisms by which patterns of activity alter synaptic efficacy and thus store information within neural connectivity. Traditional models posited that bistability at individual synapses—whereby synapses transition between "weak" and "strong" persistent states—was necessary. However, empirical data indicates that both synaptic strength and correlated dendritic spine volume exhibit unimodal, typically log-normal, distributions, and bistability in vivo has not been demonstrated (1504.01792).

Contemporary computational frameworks, such as the cascade and bidirectional cascade models, have shown that complex, multi-timescale and metastable synaptic processes can vastly expand memory capacity and preserve memories over timescales approaching an organism's lifetime (1706.04946). Metaplasticity—synaptic rules that are sensitive to the history of activity—further enhances longevity by dynamically modulating stability and plasticity in response to experience.

From a network-theoretic perspective, the physical substrate of long-term memory is proposed to be distributed connected subgraphs ("engram networks") within the brain's immense directed connectivity graph (2411.01164). These connected subgraphs are structurally robust, high-capacity, and distributed across the cortex, linking the disparate sensory and associative components of a memory. Mathematical models grounded in random graph theory and empirical connectomics demonstrate that in a cortex-sized graph, the number of possible richly interconnected, weakly overlapping subgraphs is astronomically large, explaining the cortex's massive storage potential and robustness.

2. Long-Term Memory Mechanisms in Artificial Intelligence

Modern AI systems realize long-term memory through both parametric and non-parametric mechanisms (2411.00489):

Parametric memory: Knowledge encoded directly in model parameters (weights), as in LLMs. Here, learning results in distributed traces across potentially billions of parameters. This supports rapid retrieval but is bounded by model size, training set diversity, and subject to catastrophic forgetting if continually updated.
Non-parametric (external) memory: Dedicated memory modules (e.g., vector databases, key-value stores, content-addressable memory) store and index information independently from primary model parameters. Examples include retrieval-augmented generation (RAG), lifelong learning systems with content-addressable vector memories (1610.06402), or latent-space/episodic memory networks (1812.04227).

Table: Mapping Human and AI Memory Types (2411.00489)

Human Memory	AI Parametric Memory	AI Non-Parametric Memory
Episodic	Model weights updated with history	Experience replay, event logs
Semantic	Generalization in model weights	Knowledge bases, vector stores
Procedural	Action policies in RL networks	Explicit rules, codebases

Architectures such as memory-augmented neural networks, Long-Term Memory Networks (LTMN) (1707.01961), and SuMem/M+ models (2502.00592) combine these elements, dynamically accessing both fixed-parameter and external memory banks. Critically, models such as LongMem (2306.07174) and M+ integrate scalable long-term memory modules and co-trained retrievers, enabling retention and recall of information across hundreds of thousands of tokens—a scale previously unattainable.

3. Mechanisms of Storage, Retrieval, and Forgetting

Long-term memory systems in both brains and machines require mechanisms for encoding/storage, retrieval, and controlled forgetting.

Storage: Biological systems encode information in distributed patterns of synaptic change, often involving clustering of potentiated synapses within dendritic branches (1504.01792). In AI, external memory is updated via write operations indexed by content or temporal context, enabling non-destructive, extensible storage (1610.06402).

Retrieval: Memory is accessed via cues (context, content, partial observations) that drive associative or key-based addressing; in AI, this is realized via similarity-based vector lookup or trainable retrievers (2502.00592). In continual learning systems, retrieval strategies may exploit structure (e.g., time-aware queries, fact expansion, round-level granularity (2410.10813)) and adaptive querying.

Forgetting: Biological forgetting arises through synaptic turnover, interference, consolidation, and active suppression. In AI, forgetting can be managed via active mechanisms (memory compression, deduplication, rehearsal) or is an unavoidable consequence of bounded memory or overwriting in parametric systems (2411.00489). Advanced models use learned retention policies to avoid discarding information crucial for long-term tasks (1812.04227).

4. Empirical and Benchmark Evaluations

Recent research has initiated rigorous empirical evaluation of long-term memory capabilities in both cognitive and artificial systems. Benchmarks such as LongMemEval (2410.10813) assess core competencies including:

Information extraction from extended histories,
Multi-session reasoning (combining information across epochs),
Temporal reasoning (navigating past events and their time order),
Knowledge updates (tracking changes to stored information),
Abstention (knowing when information is missing or outdated).

Commercial LLM-based chat assistants and long-context LLMs exhibit substantial performance degradation in these settings, with accuracy drops of 30% or more as session history length increases (2410.10813). Memory system designs with effective value decomposition, key expansion, and time-aware query-processing yield significant improvements in recall and QA accuracy.

5. Long-Term Memory, Lifelong Learning, and AI Self-Evolution

Long-term memory is foundational to lifelong learning and self-evolving AI. It enables the accumulation and dynamic refinement of knowledge, adaptation to new experiences, and the formation of personalized or agent-specific competencies. Systems such as OMNE (2410.15665), which integrate multi-agent architectures with robust, flexible memory principles (inspired by cortical columnar organization), demonstrate that LTM supports continual adaptation, reasoning, and collaborative problem-solving—achieving state-of-the-art performance on complex benchmarks (GAIA).

Long-term memory also underpins the transition from fixed-foundation models to self-evolving agents: models that can internally represent, retain, and update individualized knowledge in the face of changing environments and novel tasks. Critical infrastructure includes distributed, scalable memory stores, adaptive retrieval and summarization, privacy/security frameworks, and architectures that enable memory-driven prompt/service orchestration.

6. Challenges, Theoretical Insights, and Future Directions

Despite rapid progress, scaling, evaluating, and engineering long-term memory in AI presents formidable challenges:

Catastrophic forgetting persists in parametric models, and hybrid memory systems must balance efficient usage of parametric and non-parametric resources.
Retrieval scalability remains a technical bottleneck, particularly as external memory stores expand into millions of entries (1610.06402).
Memory consolidation, interference, and adaptivity need further modeling, drawing inspiration from biological mechanisms such as metaplasticity, multi-timescale cascades, and active forgetting (1706.04946).
Personalization and privacy, particularly in medical or multi-agent contexts, demand system-level solutions.

Theoretical frameworks—such as the Cognitive Architecture of Self-Adaptive Long-term Memory (SALM) (2411.00489)—provide a unifying model encompassing storage, retrieval, and forgetting adapters across six memory subtypes (episodic, semantic, procedural, each as parametric and non-parametric), integrating adaptive, feedback-driven processes analogous to those in the brain.

Promising research avenues include advanced memory-guided planning, cross-modal and agent-to-agent memory transfer, neuroadaptive brain-computer interfaces informed by long-term memory prediction (2012.03510), and graph-theoretic principles for building robust memory networks (2411.01164).

7. Summary Table: Core Concepts in Long-Term Memory

Concept	Biological Realization	AI Implementation	Impact/Significance
Storage substrate	Synaptic clusters, subgraphs	Vector DBs, memory pools, model weights	Defines capacity, robustness
Retrieval mechanism	Cue-based, associative	KNN, co-trained retrievers, attention fusion	Enables recall, reasoning
Forgetting/adaptation	Synaptic turnover, consolidation	Memory scheduling, compression, rehearsal	Balances flexibility/stability
Capacity/principles	Combinatorial subgraphs	Scalable latent memory, graph-theoretic design	Lifelong and robust learning
Evaluation	Behavioral, physiological	Long-context QA, reasoning, benchmarking	Advances practical deployment

Long-term memory, as formalized in neuroscience and realized in advanced AI systems, is essential for scalable, adaptive, and context-rich intelligence. Ongoing research at the intersection of biological theory, computational neuroscience, and machine learning continues to uncover the mechanisms and architectures by which lasting, flexible memory supports both natural and artificial intelligence.