Long-Term and Short-Term Memory

Abstract

This paper systematically examines the biological foundations, functional characteristics, and interrelationships of human short-term and long-term memory, and on this basis proposes a unifying hypothesis framework: the core function of the memory system is not merely to store information, but to generate a weighted priority table for the decision system. Short-term memory is a temporary workspace with extremely limited capacity (approximately 3–5 items) and extremely brief duration (under typical experimental conditions without rehearsal, information decays rapidly within tens of seconds), maintained by sustained neuronal firing in the prefrontal cortex. Long-term memory is persistent storage formed through emotional tagging, sleep triage, and synaptic consolidation; estimates based on synaptic information capacity suggest its potential storage scale may reach the terabyte-to-petabyte range. A 2024 study using inhibitory avoidance tasks in mice suggests that, at least in certain learning paradigms, long-term memory can form through a pathway independent of CaMKII-mediated short-term memory, indicating that the linear STM→LTM model is not universally applicable. This paper proposes a tripartite distinction among functional weights — encoding weight (the priority tag assigned when information enters memory), consolidation weight (resource allocation during sleep/replay phases), and decision weight (the bias strength exerted on behavioral choices during retrieval) — and argues that these three weight types constitute a complete chain from memory formation to behavioral output. Highly Superior Autobiographical Memory (HSAM), as a boundary case, suggests that exceptionally strong autobiographical memory may be accompanied by emotional burden and difficulty with selective forgetting, though the evidence is insufficient to prove that “lossless global memory necessarily leads to decision paralysis.” The validity of this framework is strictly delimited to the observable memory-decision channel and neither covers nor excludes alternative information pathway hypotheses explored by the author in other papers that remain at the theoretical exploration stage.

Short-Term Memory: A Capacity-Limited Temporary Workbench

1.1 Definition and Core Characteristics

Short-term memory (STM) is a cognitive system for briefly maintaining small quantities of information. Its core characteristics can be summarized by three parameters: extremely limited capacity — Miller’s (1956) classic “7±2” items has been revised to approximately 3–5 items; extremely brief duration — unrehearsed information decays within 15–30 seconds; and extreme susceptibility to interference — incoming information can immediately overwrite existing contents.

Baddeley and Hitch (1974) expanded short-term memory into the multi-component model of working memory, comprising the central executive, the phonological loop, the visuospatial sketchpad, and the episodic buffer. Working memory is not merely passive storage — it is an active “workbench” for real-time manipulation of information, supporting reasoning, comprehension, and learning.

1.2 Biological Foundations

The neural basis of short-term memory is fundamentally different from that of long-term memory. Short-term memory depends on sustained firing of prefrontal cortex neurons — neurons maintain their active state to “hold” information, and once firing ceases, the information is immediately lost. The energy cost of this maintenance mechanism is far lower than the structural synaptic changes required for long-term storage, but its capacity is limited by the number of neuronal ensembles that can simultaneously sustain firing.

At the molecular level, CaMKII (calcium/calmodulin-dependent protein kinase II) is involved in the maintenance of short-term memory. A 2024 study at the Max Planck Florida Institute using inhibitory avoidance tasks in mice found that optogenetic or genetic inhibition of CaMKII impairs short-term memory at the 1-hour time point but does not affect long-term memory formation at 1 day — suggesting that the two types of memory may rely on different molecular pathways and that long-term memory does not necessarily require prior passage through short-term memory. [Evidence level: strong, but limited to a specific task paradigm in mice]

Long-Term Memory: Selectively Consolidated Permanent Storage

2.1 Definition and Types

Long-term memory (LTM) is persistent storage formed after information undergoes a consolidation process, lasting from days to a lifetime. Unlike the unitary form of short-term memory, long-term memory comprises multiple types:

Episodic memory — memory of personal experiences that occurred at a specific time and place. “I remember what happened at my birthday party last year.” This is the memory type most closely linked to the personal weight table.

Semantic memory — knowledge of facts and concepts. “Paris is the capital of France.” It carries no temporal or contextual information.

Procedural memory — memory of skills and operational procedures. “How to ride a bicycle.” Once acquired, it is almost never forgotten.

2.2 Biological Foundations: From Synapses to Systems

The formation of long-term memory requires structural changes at the synaptic level — this is the most fundamental distinction from short-term memory. Short-term memory is merely a transient active state of neurons, whereas long-term memory involves the physical remodeling of synaptic connections:

Long-term potentiation (LTP) is the core molecular mechanism of long-term memory formation. When two neurons repeatedly fire synchronously, the synaptic connection between them is strengthened — the postsynaptic density enlarges, receptor numbers increase, and new synapses can grow. This process depends on NMDA receptor activation and protein synthesis.

Systems consolidation — memories are initially encoded in the hippocampus, then gradually transferred to the cortex for permanent storage over days to years. The hippocampus is the “temporary warehouse,” and the cortex is the “permanent archive.” This transfer occurs primarily during sleep, driven by sharp-wave ripple (SWR)–mediated replay.

The Relationship Between STM and LTM: Serial or Parallel?

3.1 The Traditional Model: Serial Conversion

Atkinson and Shiffrin’s (1968) multi-store model proposed the classic serial chain: sensory memory → short-term memory → long-term memory. In this model, information must first enter short-term memory and, through rehearsal and consolidation, can then be transferred to long-term storage. Short-term memory is the mandatory gateway to long-term memory.

3.2 The 2024 Discovery: Evidence for Parallel Pathways

A 2024 study at the Max Planck Florida Institute using inhibitory avoidance tasks in mice found that optogenetic or genetic inhibition of CaMKII impairs short-term memory at the 1-hour time point but does not impair long-term memory formation at 1 day. This suggests that, at least in certain emotional/avoidance learning paradigms, long-term memory can form through a pathway independent of CaMKII-mediated short-term memory, supporting the view that the linear STM→LTM model is not universally applicable. [Evidence level: strong, but limited to a specific task paradigm in mice]

3.3 The Core Question: What Determines the Fate of a Piece of Information?

Whether through serial or parallel pathways, the volume of information humans receive through their sensory systems each day is enormous — yet the vast majority disappears from short-term memory within seconds to minutes, and only a tiny fraction is consolidated into long-term memory. What determines whether a piece of information will decay rapidly in short-term memory or be selected for long-term storage?

The traditional answer is “rehearsal” — thinking about something repeatedly enables you to remember it. But this cannot explain why the memory of a single car accident can last a lifetime without any rehearsal, while a repeatedly memorized phone number is quickly forgotten. The answer lies in the memory weight-tagging system — the brain does not decide what to retain based on “number of repetitions” but rather allocates weights based on “emotional arousal intensity.” This is the central theme of the subsequent sections of this paper.

Memory Is Not an Archive, but a Decision Engine

Traditional cognitive science treats memory as an information storage system — a passive “archive” whose function is to preserve past experiences for future retrieval. This metaphor has profoundly shaped more than half a century of memory research, directing researchers’ attention to encoding fidelity, storage capacity, and retrieval efficiency — dimensions of “archival management.”

However, viewed from an evolutionary perspective, a pure “archive” has no survival value — remembering everything but being unable to make rapid, accurate decisions based on those memories is tantamount to stopping to consult an encyclopedia while fleeing a predator. Natural selection preserved not the most complete memory systems, but the memory systems most capable of driving adaptive behavior.

This paper advances a core proposition: In the default operating mode without metacognitive intervention, the primary function of the human memory system is not only to store information but also to generate a weighted priority table for the decision system. Encoding, consolidation, forgetting, and retrieval — the four processes traditionally regarded as “memory management” — are fundamentally processes for constructing and maintaining the decision weight table. Of course, human metacognitive abilities can review and override this default output — but metacognition itself also relies on reference data from the memory store to operate.

Formation of High-Weight Memories: From Catecholamines to Synaptic Tags

5.1 The Stress Cascade: System-Wide Reconfiguration in Milliseconds

When humans face a life-threatening situation (such as a traffic accident or predator encounter), the brain initiates a precisely orchestrated two-wave hormonal cascade within milliseconds. The first wave is driven by the sympathetic-adrenomedullary (SAM) system: the amygdala detects the threat → the hypothalamus activates the sympathetic nervous system → the adrenal medulla releases epinephrine and norepinephrine. This process is initiated before conscious awareness has completed its threat assessment.

The second wave is driven by the HPA axis: the hypothalamus releases CRH → the pituitary releases ACTH → the adrenal cortex releases cortisol. Cortisol amplifies the effects of the first-wave catecholamines, switching mechanisms across the entire brain to amplify the encoding of emotional memories.

5.2 Neural Gain: Channel Amplification and Bandwidth Expansion

The action of norepinephrine in the brain is not simply to “turn up the volume” but rather to alter the input-output transfer function of neurons — a mechanism known as neural gain. The gain parameter reshapes the input-output curve to improve the signal-to-noise ratio within each brain region: strong signals are further amplified while weak signals (noise) are further suppressed.

The GANE model (Glutamate Amplifies Noradrenergic Effects) further reveals the selectivity of this amplification: when norepinephrine release surges, it enhances the activity of neurons transmitting high-priority mental representations while suppressing the activity of neurons transmitting low-priority mental representations. [Evidence level: strong, validated by multiple experiments] This paper further translates this selective amplification into information-theoretic language — characterizing it as “enhanced signal-to-noise ratio of the channel” — this is the author’s model-based interpretation, not the formulation used in the original GANE literature. [Evidence level: author’s theoretical extrapolation]

5.3 An Information-Theoretic Account of Time Perception

Traffic accident survivors commonly report the experience of “time slowing down” — two studies by Noyes and Kletti (1976, 1977) showed that approximately 72–75% of accident survivors perceived time as slowing. David Eagleman’s classic free-fall experiment demonstrated that subjects in a state of fear could not read rapidly flashing numbers (temporal resolution was not enhanced), yet they retrospectively estimated their fall duration as approximately 36% longer than that of others.

This paper proposes a unifying explanation: enhanced neural gain → increased channel capacity → actual increase in information processed per unit time → accelerated internal clock sampling rate → perceptual experience of time “slowing down” (Arstila is correct); simultaneously, higher information throughput → genuinely denser encoding → dense memories replayed at normal speed during retrospective recall → producing the retrospective experience that “it felt very slow at the time” (Eagleman is also correct).

5.4 Synaptic Tags and Weight Inscription

The weight tags assigned during the encoding phase are not unidimensional but constitute a multidimensional tagging system:

5.5 Integration Mechanism of Multidimensional Weights

The five weight dimensions do not operate independently — they converge through an integration mechanism into a single consolidation priority score. Based on existing neurobiological evidence, we propose the following conceptual integration model:

Memory Consolidation Priority W(m) = f(θ, NE, PE, S, G)

Where:
  θ  = Theta oscillation power at encoding (baseline weight, continuous value)
  NE = Amygdala–locus coeruleus norepinephrine release level (emotional weight)
  PE = |Actual outcome − Expected outcome| (prediction error, surprise weight)
  S  = Survival-relevance detection activation level (evolutionary weight, threshold-triggered)
  G  = Prefrontal goal-relevance signal (goal weight, volitionally modulable)

Integration rules (speculative, based on existing evidence):
  · NE and θ interact multiplicatively — neither θ nor NE
    alone is sufficient to produce strong consolidation; both must co-occur
  · PE gates the effect of NE via dopaminergic pathways —
    high prediction error amplifies NE's memory-enhancing effect
  · S acts as a threshold multiplier — once survival threat is activated,
    overall weight is nonlinearly amplified (analogous to the GANE hotspot mechanism)
  · G's influence is constrained by NE level —
    under extremely high emotional arousal, goal weight can be overridden by emotional weight

It should be noted that the model above is currently conceptual — the precise functional form and parameters require determination through computational neuroscience modeling and experimental validation. However, existing evidence clearly supports the core feature that multidimensional weights interact nonlinearly (rather than summing linearly). The hotspot interaction between norepinephrine and local glutamate levels in the GANE model is one experimentally validated case of nonlinear integration.

5.6 Tripartite Distinction of the Weight Concept

This paper uses the term “weight” to encompass priority signals at three different stages of the memory processing chain. To avoid conceptual confusion, the distinctions are made explicit here:

Sleep Triage: Weight-Based Sorting and Selective Consolidation

Sleep is not passive “rest” but the memory system’s active triage station. During NREM sleep, the brain performs three core operations:

First, selective consolidation of high-weight memories. Sharp-wave ripples (SWRs) in the hippocampus selectively replay memories bearing high-weight tags. Research suggests that the power of theta oscillations during encoding may predict the density of subsequent slow oscillation–sleep spindle coupling during sleep — however, as a complete causal chain, “theta tag → sleep coupling density → consolidation resources,” the independent evidence for each node is strong, while the direct causal links between nodes still require further experimental validation. [Evidence level: strong at individual nodes, moderate for the causal chain]

Second, active forgetting of low-weight memories. The synaptic homeostasis hypothesis holds that brain-wide synaptic downscaling during sleep is necessary — it “resets” synapses that were excessively potentiated during the day to sustainable levels, at the cost of losing low-weight memories.

Third, gist extraction. Sleep triage is not a simple “retain/delete” binary but also includes extracting abstract regularities from specific events — stripping away detailed noise while preserving patterns and rules.

Information Input
  ↓
Encoding Phase — Real-time weight tag inscription
  ├─ Theta oscillation power → Baseline weight
  ├─ Emotional arousal level → Emotional weight
  ├─ Reward prediction error → Surprise weight
  ├─ Survival relevance → Evolutionary weight
  └─ Top-down commands → Goal weight
  ↓
Sleep Triage — Read tags, execute sorting
  ├─ High weight → Priority SWR replay → Consolidated into long-term memory
  ├─ Medium weight → Partial consolidation, gist extraction, detail discarded
  └─ Low weight → Synaptic downscaling / Neurogenesis-driven overwriting → Active forgetting
  ↓
Long-Term Memory Store — Reference database for the decision system

Forgetting: Not a Bug, but the Core Algorithm of Intelligence

7.1 The Evolutionary Significance of the Ebbinghaus Curve

The forgetting curve documented by Ebbinghaus (1885) — 50% lost within 20 minutes, 70% within 24 hours — was long regarded as a “defect” of the memory system. But research over the past decade has completely reversed this understanding: forgetting is an active function of the brain, not passive decay.

Dopaminergic circuits participate simultaneously in both memory formation and clearance — the same neurons bidirectionally regulate “retention” and “deletion.” Adult neurogenesis in the hippocampus promotes active forgetting — the integration of newborn neurons into memory engrams may facilitate the fading of old memories by modifying those engrams.

7.2 HSAM: A Suggestive Boundary Case

Highly Superior Autobiographical Memory (HSAM) is an extremely rare condition characterized by exceptionally detailed and rapid autobiographical recall. It provides a valuable boundary case for this framework — but requires cautious interpretation to avoid overextension. [Evidence level: suggestive, extremely small sample]

It must be emphasized that HSAM represents superior autobiographical episodic memory, not lossless whole-brain, all-modality memory. How HSAM individuals avoid being continuously overwhelmed by past memories, and their inhibitory control mechanisms, remain poorly understood. Therefore, HSAM cannot simply be equated with “lossless global memory leads to cognitive paralysis” — it is a suggestive boundary case hinting that forgetting mechanisms may play an important role in maintaining cognitive efficiency, but the evidential strength of this inference is limited.

Memory Weights → Decision Weights: The Causal Mapping

8.1 Reinterpreting Damasio’s Somatic Marker Hypothesis

Antonio Damasio’s somatic marker hypothesis proposes that emotional/bodily signals bias decision-making in complex, uncertain situations. Within this framework, we reinterpret somatic markers as the physiological expression that occurs when the emotional tags of high-weight memories are reactivated in a decision context. When facing a choice, the brain retrieves high-weight memories relevant to the current situation; the emotional tags carried by these memories produce somatic responses (accelerated heartbeat, stomach contractions, muscle tension), which bias choice preferences. [Evidence level: theoretical reinterpretation, a within-framework re-reading of an existing hypothesis]

It should be noted that the somatic marker hypothesis itself is not without controversy — Dunn et al. (2006) critically evaluated its core experimental paradigm and interpretation. Therefore, the invocation of somatic markers in this paper should be understood as “this framework can accommodate and reinterpret the hypothesis,” rather than “Damasio has proven that decision-making is the reading of memory weight tags.” Competing explanations exist, including alternative models based on attentional prioritization and reinforcement learning value functions.

Patients with vmPFC lesions provide crucial evidence: their intelligence is intact, working memory is normal, and language ability is normal — yet they cannot make decisions. What they have lost is not “memory” or “rationality” but the ability to “read emotional weight tags.” The weight table exists but cannot be accessed — and the decision system is consequently paralyzed.

8.2 The Underlying Implementation of Kahneman’s System 1

Daniel Kahneman’s distinction between System 1 (fast, automatic, intuitive) and System 2 (slow, effortful, rational) receives a neurobiological-level explanation within this framework. System 1’s “intuition” is not a mysterious cognitive shortcut but the rapid retrieval output of the memory weight table — high-weight memories are preferentially retrieved, and the emotional tags they carry directly bias choice preferences. System 2 is the prefrontal cortex’s secondary evaluation and potential override of System 1’s output.

8.3 Three-Layer Decision Architecture

Three-Layer Decision Architecture (Dynamic Priority Spectrum)

Perceptual Input (current situational information)
  ↓
Layer 1: Physiological Homeostasis Assessment (continuously running)
  Hypothalamus/brainstem monitors homeostatic parameters
  → Hunger / thirst / fatigue / temperature / pain signals
  → Upon reaching threshold, automatically drives behavior (eating / drinking / sleeping)
  ↓
Layer 2: Threat Assessment (continuously monitoring)
  Amygdala assesses threat level
  → Low threat: continue normal processing
  → High threat: strongly suppresses prefrontal cortex → stress reflexes take over
  ↓
Layer 3: Deliberative Behavioral Decision (operates when permitted by the first two layers)
  Hippocampus retrieves relevant high-weight memories
  → Encoding weight + consolidation weight → decision weight ranking
  → Somatic marker activation → System 1 intuitive output
  → Prefrontal System 2 review → accept / modify / override
  → Final decision output
  → Outcome feedback → updates memory weight table

A hierarchical priority relationship exists among the physiological layer, the stress layer, and the deliberative layer. In extreme cases, the amygdala’s threat detection can powerfully suppress prefrontal activity — in the face of acute, life-threatening danger, this suppression approaches total shutdown. However, in most everyday stress situations (social conflict, work anxiety, mild fear), the relationship between the amygdala and the prefrontal cortex more closely approximates a gradual spectrum of dynamic interplay — the emotional substrate continuously biases and interferes with cognitive control, while the prefrontal cortex can reciprocally regulate the amygdala’s response intensity. Only under the most extreme stress states does this interplay collapse into unilateral “takeover.”

Bio-Economics of the Stress System: The Gecko-Tail Survival Trade-off

9.1 The Cost of “Overclocking”

The stress system is a biological apparatus optimized for “single-use, short-duration, high-intensity” scenarios. Its design assumption is that activation frequency is low, each episode is brief, and intervals between activations allow sufficient recovery time. Much like a gecko shedding its tail — sacrificing the tail (energy reserves, balance) in exchange for the opportunity to escape.

The potential costs of a single acute stress episode include: in animal pharmacological catecholamine-overload models, a single dose of isoproterenol can cause approximately 10% cardiomyocyte damage with signs of recovery — suggesting that extreme catecholamine exposure may carry a cardiac cost, but this dosage far exceeds natural stress levels and cannot be directly equated with a single episode of acute psychological stress in ordinary humans [Evidence level: animal pharmacological model]. Additionally, high-energy-consuming components of the immune system are transiently suppressed, and oxidative stress increases. When any of these three design assumptions is violated — frequency too high, duration too long, or recovery period insufficient — the system transitions from “life-saving tool” to “pathogenic source.”

9.2 Recovery Capacity Is the True Rate-Limiting Variable

The bottleneck determining how frequently and how sustainably a person can utilize high-performance states is not “how far the accelerator can be pressed” but “how quickly the brakes can restore the system.” Heart rate variability (HRV) is a precise tool for measuring this recovery capacity — it measures the efficiency with which the nervous system switches from “high-performance mode” back to “repair mode.”

The formula for sustainable peak performance: alternating cycles of high-arousal windows × complete recovery periods. Recovery capacity depends on four pillars: deep sleep quality, nutritional supply quality, oxygen delivery efficiency, and parasympathetic activation depth (deep meditation).

Energy Budget of Human Intelligence: From Biological Maintenance to Cognitive Dominance

The human brain constitutes only 2% of body weight yet consumes approximately 20% of the body’s resting metabolic rate — more than 10 times the value predicted by body weight alone. During childhood, the brain’s metabolic demands peak — consuming glucose equivalent to 66% of the body’s resting metabolic rate and 43% of daily energy requirements.

Even more noteworthy is that the signaling pathways in the most recently expanded brain regions (those responsible for higher cognition) cost 67% more energy than those in sensorimotor areas. This precisely corresponds to a critical transition: the intelligent system has supplanted the biological maintenance system as humanity’s primary energy consumer.

This reallocation of the energy budget was not the result of the brain “actively choosing” — the causal direction is precisely the opposite. According to the LEECHO Evolutionary Causal Chain Hypothesis Model (detailed in the author’s evolutionary paper), changes began at the most fundamental level of energy input and reshaped the entire species layer by layer upward. The independent data for each node in the following narrative are relatively robust, but the entire chain as a coherent causal narrative should be regarded as a synthetic evolutionary hypothesis, not a conclusion with a completed empirical loop: [Evidence level: individual nodes strong, overall causal chain is a hypothesis model]

Energy foundation: Approximately 2 million years ago, natural wildfires killed large animals, providing human ancestors with far greater nutritional density and digestive efficiency than raw food — a zero-cognitive-threshold natural event. Initially, this supply of cooked meat was intermittent (dependent on wildfire frequency), so only a slow accumulation of energy advantage occurred in individual members who habitually encountered cooked meat, not yet sufficient to alter species-level traits.

Metabolic layer: When human ancestors learned to collect and preserve fire, cooked meat transitioned from an occasional food source to a dietary staple. Sustained high-energy input shifted the selection pressures on the digestive system — the long digestive tract previously needed to process large amounts of coarse fiber gradually simplified and shortened, releasing energy budget previously allocated to the digestive system.

Structural layer: Reduced digestive burden + energy abundance → bipedal locomotion gained a selective advantage. Fire brought back to camp → nighttime safety → ground sleeping replaced arboreal sleeping.

Neural development layer: Ground-based deep sleep → substantial increase in REM duration → an unprecedented time window for the development, pruning, and consolidation of neural circuits → brain volume and complexity began to increase — at which point the energy budget released by the digestive system could sustain this increasingly costly organ.

The initial direction of the causal chain is from the bottom layer upward — the change in energy supply is the initiating condition for the entire chain. However, once brain development reached a certain level of complexity, bidirectional feedback emerged: a larger brain could invent more efficient fire-use and foraging strategies, thereby further increasing energy intake, accelerating the simplification of the digestive system and the continued growth of the brain. Therefore, the complete picture is: a bottom-up energy revolution unidirectionally initiated the entire causal chain, but once the chain was in motion, a positive feedback loop of mutual acceleration formed between upper and lower layers, driving the exponential acceleration of human evolution.

Core Proposition: What You Store Determines Who You Become

Synthesizing the analysis across all the foregoing dimensions, the core proposition of this paper can be stated as follows:

The contents of a person’s memory store determine the weight table. The weight table determines the ranking. The ranking determines which option is prioritized for presentation to consciousness. The option prioritized for presentation has a high probability of becoming the final decision.

This means that a person’s “personality,” “preferences,” and “intuitions” — at the level observable by neuroscience — are highly correlated with the statistical distribution characteristics of high-weight memories in their long-term memory store. When two people make different decisions when confronted with an identical set of options, one important reason is that their memory stores contain different contents, and therefore the ranking results they produce differ.

However, it must be emphasized that memory weight distribution is the dominant factor in behavioral decision-making, not the sole factor. Human self-awareness possesses the capacity for dynamic construction — individuals can monitor, evaluate, and even actively override the automatic output of the memory weight system through metacognition (“reflecting on one’s own thinking”). This ability to “audit the weight table itself” is a higher-order emergent property of human cognition and cannot be simply reduced to the statistical characteristics of the weight table itself. Moreover, although values and belief systems are highly correlated with memory contents, once formed they can function as independent cognitive frameworks that retroactively shape the encoding of new experiences — this is a top-down modulation that assigns different weight tags to the same event within different value frameworks.

Regarding the relationship between values and the memory weight table, this paper’s position is: values can be understood as “higher-order meta-rules for the assignment of weight tags.” They both originate from the historical accumulation of the memory weight table (your past high-weight experiences shaped your value judgments) and can retroactively modulate the encoding weights of new experiences (your values determine how high a weight tag the same event receives in your system). This constitutes a recursive structure — the weight table generates values, values modulate the weight table — the two do not stand in a simple containment relationship but are a co-evolving coupled system.

Within the framework of the memory weight system, there are two primary pathways for changing behavioral patterns:

Discussion: Architectural Implications for Personalized AI Alignment

This framework poses a fundamental redefinition of the question for LLM personalized alignment research. Current AI personalization research asks “how to learn the user’s surface preferences” — style, format, tone. But the analysis in this paper suggests that genuine alignment requires understanding the user’s implicit memory weight distribution — that is, what resides in this person’s high-weight long-term memory store.

The fundamental problem with current RLHF is that training with the “average human weight table” amounts to installing the same memory store in every user. System prompts attempt to use “working memory” to override “long-term memory” — but with insufficient influence weight. A worthwhile engineering direction to explore is writing personal preferences into the model’s parameter layer through lightweight adapters such as LoRA, rather than relying solely on the context window.

But the core challenge remains: human weight tables are implicit — users themselves typically do not know what their weight tables contain. Therefore, genuine personalized alignment requires not asking users “what do you want” but reverse-engineering their implicit weight distribution from long-term behavioral patterns.

In AI interaction scenarios, complete life-event data is unavailable — this is an environment of extreme information sparsity. But this framework suggests three feasible engineering pathways for addressing this sparsity:

Falsifiable Predictions

A good hypothesis framework should explicitly enumerate its falsification conditions. Below are the evidence levels and verifiable predictions for the core propositions of this framework:

Conclusion

By integrating existing evidence from stress neuroscience, sleep biology, information theory, and evolutionary psychology, this paper proposes a unifying hypothesis framework: the human memory system is the core infrastructure of the decision system, operating according to the logic of “emotional weighting → sleep triage → selective consolidation/forgetting → weight-based ranking → decision output.”

The framework’s main contributions include: (1) systematically defining the biological foundations and functional differences between short-term and long-term memory, and introducing the 2024 parallel pathway discovery (limited to a specific mouse paradigm) to suggest that the relationship between the two may be more complex than the traditional serial model allows; (2) proposing a tripartite distinction of the weight concept (encoding weight / consolidation weight / decision weight) and a conceptual nonlinear integration model of five-dimensional weight tags; (3) using information-theoretic language to propose a hypothetical explanation that unifies the Eagleman and Arstila camps in the time perception debate; (4) offering a compatibility-based reinterpretation within this framework of the Damasio somatic marker hypothesis (acknowledging competing explanations), the Kahneman dual-system theory, and the Shohamy episodic memory–decision model; (5) using HSAM boundary cases and vmPFC lesion evidence to suggest that forgetting mechanisms may serve an important adaptive function in maintaining cognitive efficiency; and (6) proposing a “three-layer decision architecture” — a dynamic priority spectrum relationship among the physiological homeostasis layer, the stress layer, and the deliberative behavioral decision layer.

What memories a person stores largely determines their future behavioral patterns and decision outputs. Memory is not an archive of the past but a blueprint for future behavior.

Key References

Damasio, A. (1994). Descartes’ Error: Emotion, Reason, and the Human Brain. Putnam.

Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.

Kandel, E. (2006). In Search of Memory: The Emergence of a New Science of Mind. W.W. Norton.

Schacter, D. (2001). The Seven Sins of Memory. Houghton Mifflin.

Cowan, N. (2001). The magical number 4 in short-term memory: A reconsideration of mental storage capacity. Behavioral and Brain Sciences, 24(1), 87-114.

Dunn, B.D. et al. (2006). The somatic marker hypothesis: A critical evaluation. Neuroscience & Biobehavioral Reviews, 30(2), 239-271.

Baddeley, A. & Hitch, G. (1974). Working memory. Psychology of Learning and Motivation, 8, 47-89.

Atkinson, R.C. & Shiffrin, R.M. (1968). Human memory: A proposed system and its control processes. Psychology of Learning and Motivation, 2, 89-195.

Small, S. (2021). Forgetting: The Benefits of Not Remembering. Crown.

Hermans, E.J. et al. (2011). Stress-Related Noradrenergic Activity Prompts Large-Scale Neural Network Reconfiguration. Science, 334(6059), 1151-1153.

Eagleman, D.M. et al. (2007). Does time really slow down during a frightening event? PLoS ONE, 2(12), e1295.

Sederberg, P.B. et al. (2003). Theta and Gamma Oscillations during Encoding Predict Subsequent Recall. J. Neuroscience, 23(34), 10809-10814.

McGaugh, J.L. (2004). The amygdala modulates the consolidation of memories of emotionally arousing experiences. Annual Review of Neuroscience, 27, 1-28.

Rouhani, N. et al. (2023). Multiple routes to enhanced memory for emotionally relevant events. Trends in Cognitive Sciences.

Dunsmoor, J.E. et al. (2022). Tag and Capture: How Salient Experiences Target and Rescue Nearby Events in Memory. Trends in Cognitive Sciences.

Stickgold, R. & Walker, M. (2013). Sleep-dependent memory triage. Nature Neuroscience, 16, 139-145.

Nairne, J.S. et al. (2007). Adaptive memory: Survival processing enhances retention. J. Experimental Psychology: Learning, Memory, and Cognition.

Shohamy, D. & Daw, N.D. (2015). Integrating memories to guide decisions. Current Opinion in Behavioral Sciences, 5, 85-90.

Aston-Jones, G. & Cohen, J.D. (2005). An integrative theory of locus coeruleus-norepinephrine function. Annual Review of Neuroscience, 28, 403-450.

Shin, M.E., Parra-Bueno, P. & Yasuda, R. (2024). Formation of long-term memory without short-term memory revealed by CaMKII inhibition. Nature Neuroscience, 28(1), 35-39. doi: 10.1038/s41593-024-01831-z

Bartol, T.M. et al. (2015). Nanoconnectomic upper bound on the variability of synaptic plasticity. eLife, 4, e10778. doi: 10.7554/eLife.10778

Noyes, R. & Kletti, R. (1976). Depersonalization in the face of life-threatening danger: A description. Psychiatry, 39(1), 19-27.

Noyes, R. & Kletti, R. (1977). Depersonalization in response to life-threatening danger. Comprehensive Psychiatry, 18(4), 375-384.

LePort, A.K. et al. (2012). Behavioral and neuroanatomical investigation of Highly Superior Autobiographical Memory (HSAM). Neurobiology of Learning and Memory, 98(1), 78-92.

Kuzawa, C.W. et al. (2014). Metabolic costs and evolutionary implications of human brain development. PNAS, 111(36), 13010-13015.

Mather, M. et al. (2016). Norepinephrine ignites local hotspots of neuronal excitation. J. Neuroscience, 36(11), 3095-3107.

Willis, M.A. et al. (2016). Acute catecholamine exposure causes reversible myocyte injury without cardiac regeneration. Circulation Research, 119(7), 865-879.

Davis, R.L. & Bhatt, D. (2024). Active forgetting and neuropsychiatric diseases. Molecular Psychiatry, 29, 3207-3218.

Arstila, V. (2012). Time slows down during accidents. Frontiers in Psychology, 3, 196.