Memory Strength Research Articles

Phenotypic memory in quorum sensing.

Quorum sensing (QS) is a regulatory mechanism used by bacteria to coordinate group behavior in response to high cell densities. During QS, cells monitor the concentration of external signals, known as autoinducers, as a proxy for cell density. QS often involves positive feedback loops, leading to the upregulation of genes associated with QS signal production and detection. This results in distinct steady-state concentrations of QS-related molecules in QS-ON and QS-OFF states. Due to the slow decay rates of biomolecules such as proteins, even after removal of the initial stimuli, cells can retain elevated levels of QS-associated biomolecules for extended periods of time. This persistence of biomolecules after the removal of the initial stimuli has the potential to impact the response to future stimuli, indicating a memory of past exposure. This phenomenon, which is a consequence of the carry-over of biomolecules rather than genetic inheritance, is known as "phenotypic" memory. This theoretical study aims to investigate the presence of phenotypic memory in QS and the conditions that influence this memory. Numerical simulations based on ordinary differential equations and analytical modeling were used to study gene expression in response to sudden changes in cell density and extracellular signal concentrations. The model examined the effect of various cellular parameters on the strength of QS memory and the impact on gene regulatory dynamics. The findings revealed that QS memory has a transient effect on the expression of QS-responsive genes. These consequences of QS memory depend strongly on how cell density was perturbed, as well as various cellular parameters, including the Fold Change in the expression of QS-regulated genes, the autoinducer synthesis rate, the autoinducer threshold required for activation, and the cell growth rate.

Effect of levels-of-processing on rates of forgetting.

The levels-of-processing (LOP) framework, proposing that deep processing yields superior retention, has provided an important paradigm for memory research and a practical means of improving learning. However, the available levels-of-processing literature focuses on immediate memory performance. It is assumed within the LOP framework that deep processing will lead to slower forgetting than will shallow processing. However, it is unclear whether, or how, the initial level of processing affects the forgetting slopes over longer retention intervals. The present three experiments were designed to explore whether items encoded at qualitatively different LOP are forgotten at different rates. In the first two experiments, depth of processing was manipulated within-participants at encoding under deep and shallow conditions (semantic vs. rhyme judgement in Experiment 1; semantic vs. consonant-vowel pattern decision in Experiment 2). Recognition accuracy (d prime) was measured between-participants immediately after learning and at 30-min, 2-h, and 24-h delays. The third experiment employed a between-participants design, contrasting the rates of forgetting following semantic and phonological (rhyme) processing at immediate, 30-min, 2-h, and 6-h delays. Results from the three experiments consistently demonstrated a large effect size of levels of processing on immediate performance and a medium-to-large level effect size on delayed recognition, but crucially no LOP × delay group interaction. Analysis of the retention curves revealed no significant differences between the slopes of forgetting for deep and shallow processing. These results suggest that the rates of forgetting are independent of the qualitatively distinct encoding operations manipulated by levels of processing.

Want to learn? think again!

One of the best ways to improve new learning and increase memory strength is by reprocessing the recently acquired information, for example, by thinking of it again. Synaptic plasticity, the process by which neurons change the strength of their connections with each other, is fundamental for learning and memory formation. Yet, at present, it is unclear how reprocessing information drives synaptic plasticity to support memory improvement. A new study suggests that reprocessing enhances memory formation by recruiting more synapses to represent the new memory, thus increasing its strength.

Evaluating the machine learning models based on natural language processing tasks

<span lang="EN-US">In the realm of natural language processing (NLP), a diverse array of language models has emerged, catering to a wide spectrum of tasks, ranging from speaker recognition and auto-correction to sentiment analysis and stock prediction. The significance of language models in enabling the execution of these NLP tasks cannot be overstated. This study proposes an approach to enhance accuracy by leveraging a hybrid language model, combining the strengths of long short-term memory (LSTM) and gated recurrent unit (GRU). LSTM excels in preserving long-term dependencies in data, while GRU's simpler gating mechanism expedites the training process. The research endeavors to evaluate four variations of this hybrid model: LSTM, GRU, <a name="_Hlk158992630"></a>bidirectional long short-term memory (Bi-LSTM), and a combination of LSTM with GRU. These models are subjected to rigorous testing on two distinct datasets: one focused on IBM stock price prediction, and the other on Jigsaw toxic comment classification (sentiment analysis). This work represents a significant stride towards democratizing NLP capabilities, ensuring that even in resource-constrained settings, NLP models can exhibit improved performance. The anticipated implications of these findings span a wide spectrum of real-world applications and hold the potential to stimulate further research in the field of NLP. </span>

A Stock Index Futures Price Prediction Approach Based on the MULTI-GARCH-LSTM Mixed Model

As a type of financial derivative, the price fluctuation of futures is influenced by a multitude of factors, including macroeconomic conditions, policy changes, and market sentiment. The interaction of these factors makes the future trend become complex and difficult to predict. However, for investors, the ability to accurately predict the future trend of stock index futures price is directly related to the correctness of investment decisions and investment returns. Therefore, predicting the stock index futures market remains a leading and critical issue in the field of finance. To improve the accuracy of predicting stock index futures price, this paper introduces an innovative forecasting method by combining the strengths of Long Short-Term Memory (LSTM) networks and various Generalized Autoregressive Conditional Heteroskedasticity (GARCH)-family models namely, MULTI-GARCH-LSTM. This integrated approach is specifically designed to tackle the challenges posed by the nonstationary and nonlinear characteristics of stock index futures price series. This synergy not only enhances the model’s ability to capture a wide range of market behaviors but also significantly improves the precision of future price predictions, catering to the intricate nature of financial time series data. Initially, we extract insights into the volatility characteristics, such as the aggregation of volatility in futures closing prices, by formulating a model from the GARCH family. Subsequently, the LSTM model decodes the complex nonlinear relationships inherent in the futures price series and incorporates assimilated volatility characteristics to predict future prices. The efficacy of this model is validated by applying it to an authentic dataset of gold futures. The empirical findings demonstrate that the performance of our proposed MULTI-GARCH-LSTM hybrid model consistently surpasses that of the individual models, thereby confirming the model’s effectiveness and superior predictive capability.

Dorsal raphe to basolateral amygdala corticotropin-releasing factor circuit regulates cocaine-memory reconsolidation.

Environmental stimuli elicit drug craving and relapse in cocaine users by triggering the retrieval of strong cocaine-related contextual memories. Retrieval can also destabilize drug memories, requiring reconsolidation, a protein synthesis-dependent storage process, to maintain memory strength. Corticotropin-releasing factor (CRF) signaling in the basolateral amygdala (BLA) is necessary for cocaine-memory reconsolidation. We have hypothesized that a critical source of CRF in the BLA is the dorsal raphe nucleus (DR) based on its neurochemistry, anatomical connectivity, and requisite involvement in cocaine-memory reconsolidation. To test this hypothesis, male and female Sprague-Dawley rats received adeno-associated viruses to express Gi-coupled designer receptors exclusively activated by designer drugs (DREADDs) selectively in CRF neurons of the DR and injection cannulae directed at the BLA. The rats were trained to self-administer cocaine in a distinct environmental context then received extinction training in a different context. Next, they were briefly re-exposed to the cocaine-predictive context to destabilize (reactivate) cocaine memories. Intra-BLA infusions of the DREADD agonist deschloroclozapine (DCZ; 0.1 mM, 0.5 µL/hemisphere) immediately after memory reactivation attenuated cocaine-memory strength, relative to vehicle infusion. This was indicated by a selective, DCZ-induced and memory reactivation-dependent decrease in drug-seeking behavior in the cocaine-predictive context in DREADD-expressing males and females at test compared to respective controls. Notably, BLA-projecting DR CRF neurons that exhibited increased c-Fos expression during memory reconsolidation co-expressed the glutamatergic neuronal marker, vesicular glutamate transporter 3. Together, these findings suggest that the DRCRF → BLA circuit is engaged to maintain cocaine-memory strength after memory destabilization, and this phenomenon may be mediated by DR CRF and/or glutamate release in the BLA.

Identification of structural defects and their impact on the magnetic memory, static and cyclic strength of VNS9-SH thin sheet trip-steel

Identification of microstructure defects, which are stress concentrators (SC) during the operation of mechanical engineering products, is an important scientific and practical task relevant for manufacturing enterprises. This problem becomes especially urgent and difficult for critical helicopter parts made of sheet TRIP steel VNS9-Sh (23Kh15N5AM3-Sh) and operating under cyclic loads due to the complex microstructure of the steel and small thickness of strips and sheets. To assess the impact of structural defects on the cyclic strength of products made of the steel under study, the specimens were preliminary sorted proceeding from the results of their testing using the method of metal magnetic memory (MMM) and metallographic studies. The MMM method is a structure-sensitive procedure which provides information about the presence of structural defects that arise during the manufacture. Magnetic anomalies in the form of sharp local changes in the intrinsic stray magnetic field (SSMF) (H) and its gradient |ΔH| along the length of the controlled section Δx, were identified on the surface of sheets cut from five different batches. A conventional classification of the identified anomalies was made according to the magnitude of the magnetic field gradient. The specimens of two types were cut in zones of magnetic anomalies and outside them: type 1 — for cyclic tests and type 2 — for metallographic studies. The geometric parameters and field gradient values of magnetic anomalies on specimens of type 1 and type 2 were the same. Metallographic studies in zones of maximum magnetic field gradient on type 2 specimens revealed defects in the form of a strip at the boundary of different structures, which is a structural stress concentrator (SSC) and a source of the inhomogeneity and changes in the magnetic properties. Type 1 specimens with similar magnetic anomalies and Type 1 specimens cut from sheets outside zones of magnetic anomalies were then selected for cyclic testing. Comparative tests for cyclic strength of the specimens with and without specified SSC were carried out. It is shown that the presence of SSC zones in the specimens reduces the number of cycles to failure during cyclic tests by 1 – 2 orders of magnitude compared to the specimens free of SSC. Based on cyclic tensile tests of specimens, a limiting value of the magnetic field gradient was determined that corresponds to the acceptable level of stress concentration on structural defects. This value is recommended for use as a rejection criterion when examining a new tape by the MMM method.

Adaptive forgetting speed in working memory.

Working memory is known to be capacity-limited and is therefore selective not only for what it encodes but also what it forgets. Explicit forgetting cues can be used effectively to free up capacity, but it is not clear how working memory adaptively forgets in the absence of explicit cues. An important implicit cue that may tune forgetting in working memory is the passage of time. When information becomes irrelevant more quickly, working memory should also forget information more quickly. In three delayed-estimation experiments, we systematically manipulated how probing probability changed as time passed on after encoding an item (i.e., the "probing hazard"). In some blocks, probing hazard decreased after encoding an item, requiring participants to only briefly retain the memory item. In other blocks, the probing hazard increased or stayed flat, as the retention interval was lengthened. In line with our hypothesis, we found that participants adapted their forgetting rate to the probing dynamics of the working memory task. When the memory item quickly became irrelevant ("decreasing" probing hazard), forgetting rate was higher than in blocks where probing hazard increased or stayed flat. The time course of these adaptations in forgetting implies a fast and flexible mechanism. Interestingly, participants could not explicitly report the order of conditions, suggesting forgetting is implicitly sped up. These findings suggest that implicit adaptations to the temporal structure of our environment tune forgetting speed in working memory, possibly contributing to the flexible allocation of limited working memory resources.

The digital economy and advertising diffusion models: Critical mass and the Stalling equilibrium

In the digital economy it is frequently observed that products become more valuable the larger is the number of people that use it. To account for such network effects, we introduce a new diffusion equation in a dynamic model of the firm with the aim to obtain the advertising policy that maximizes firm profits. Also an advertising budget is introduced.First, we find that introduction of the network effect leads to a critical mass of the sales level. Only if the sales are greater or equal to this level, is it optimal for the firm to have a substantial sales level in the long run. Second, introduction of an advertising budget could result in the existence of a Stalling equilibrium. This new type of equilibrium, at which the advertising amount is at its upper bound, serves as the critical mass threshold separating growth and decline. Third, the advertising policy is continuous in the sales level either when the discount rate is large and the rate of forgetfulness is low, when the Stalling equilibrium exists, or when the rate of forgetfulness is that high that the firm is always in decline. Furthermore, existence of the Stalling equilibrium involves an important message to the management of the firm in that a slight increase of the advertising budget would open up an avenue of growth in sales.

Paradoxical boosting of weak and strong spatial memories by hippocampal dopamine uncaging.

The ability to remember changes in the surroundings is fundamental for daily life. It has been proposed that novel events producing dopamine release in the hippocampal CA1 region could modulate spatial memory formation. However, the role of hippocampal dopamine increase on weak or strong spatial memories remains unclear. We show that male mice exploring two objects located in a familiar environment for 5 minutes created a short-term memory (weak) that cannot be retrieved one day later, whereas 10 minutes exploration created a long-term memory (strong) that can be retrieved one day later. Remarkably, hippocampal dopamine elevation during the encoding of weak object location memories (OLMs) allowed their retrieval one day later but dopamine elevation during the encoding of strong OLMs promoted the preference for a familiar object location over a novel object location after 24 hours. Moreover, dopamine uncaging after the encoding of OLMs did not have effect on weak memories whereas on strong memories diminished the exploration of the novel object location. Additionally, hippocampal dopamine elevation during the retrieval of OLMs did not allow the recovery of weak memories and did not affect the retrieval of strong memory traces. Finally, dopamine elevation increased hippocampal theta oscillations, indicating that dopamine promotes the recurrent activation of specific groups of neurons. Our experiments demonstrate that hippocampal dopaminergic modulation during the encoding of OLMs depends on memory strength indicating that hyperdopaminergic levels that enhance weak experiences could compromise the normal storage of strong memories.Significance statement Increased levels of dopamine have been related to cognitive enhancement. Hippocampal dopamine elevation caused by novelty exposure has been proposed as a strategy to enhance memory based on the observation that surprising events create flashbulb memories that are remembered for long time. However, hyperdopaminergic levels could also underlie maladaptive memories, such as non-desired preservation of traumatic experiences. Our experiments show that dopamine elevation in the dorsal hippocampus during the encoding of spatial memories has paradoxical effects, while the enhancement of weak memories allows their retrieval, the dopaminergic modulation of strong memories limits the ability to modify preexisting spatial memories by changes in the environment. We conclude that cognitive enhancement through dopamine boosting must consider diverse aspects of memory formation.

A lifespan study of the confidence-accuracy relation in working memory and episodic long-term memory.

The relation between an individual's memory accuracy and reported confidence in their memories can indicate self-awareness of memory strengths and weaknesses. We provide a lifespan perspective on this confidence-accuracy relation, based on two previously published experiments with 320 participants, including children aged 6-13, young adults aged 18-27, and older adults aged 65-77, across tests of working memory (WM) and long-term memory (LTM). Participants studied visual items in arrays of varying set sizes and completed item recognition tests featuring 6-point confidence ratings either immediately after studying each array (WM tests) or following a long period of study events (LTM tests). Confidence-accuracy characteristic analyses showed that accuracy improved with increasing confidence for all age groups and in both WM and LTM tests. These findings reflect a universal ability across the lifespan to use awareness of the strengths and limitations of one's memories to adjust reported confidence. Despite this age invariance in the confidence-accuracy relation, however, young children were more prone to high-confidence memory errors than other groups in tests of WM, whereas older adults were more susceptible to high-confidence false alarms in tests of LTM. Thus, although participants of all ages can assess when their memories are weaker or stronger, individuals with generally weaker memories are less adept at this confidence-accuracy calibration. Findings also speak to potential different sources of high-confidence memory errors for young children and older adults, relative to young adults. (PsycInfo Database Record (c) 2024 APA, all rights reserved).

A Study of Optimization Algorithms for English Vocabulary Memory and Review Plans: Integrating the Principles of Forgetting Curve and Memory Reinforcement

Mastering English vocabulary is a crucial steppingstone towards achieving fluency in the language. Building a strong English vocabulary foundation is essential for spoken and written fluency. This inefficiency can lead to wasted time reviewing known words or forgetting recently learned ones. This work proposes a novel deep learning-based system for personalized English vocabulary learning: the Contextual Awareness Controlled Generative Adversarial Network with Forget Curve & Memory Reinforcement (CACGAN-FCMR-PEVL) model. CACGAN-FCMR-PEVL leverages deep learning to create a personalized optimization algorithm, analyzing user data and incorporating forgetting curve principles. It utilizes a self-attention mechanism and a conditional generative adversarial network (CGAN) to generate personalized vocabulary maps and review schedules, optimizing memory retention through memory reinforcement techniques. Hyperparameter tuning with the Binary Waterwheel Plant Optimization Algorithm (BWpOA) further enhances the model's effectiveness. When compared to other existing methods, the proposed CACGAN-BWpOA-FCMR-PEVL model shows 53.55%, 31.703% and 32.403% higher Vocabulary Recall and 49.46%, 58.06% and 30.98% higher Word Similarity, indicating superior learning and retention. It also attains 47.42%, 56.701% and 73.21% higher Review Interval Prediction Accuracy (RIPA), which suggests effective personalization based on forgetting curves. These findings suggest that CACGAN-BWpOA-FCMR-PEVL effectively personalizes vocabulary learning, leading to improved memory retention.

U T’aan Nukuch Máak (‘Words of the Elders’): Defining a Yucatec Maya ‘Speech’ Genre

Yucatec Maya orality has been a popular topic of study of scholars from a wide array of disciplines. These studies often rely on academically generated categories of speech that have often been stylized in forms that descend from Western thought. The generated speech categories may overlook more performance-based forms that are more common of Indigenous knowledge systems. Most of the collected and analyzed Maya oral literature appears to be recounted by men, leaving women’s orality and their unique ways of interpreting the world largely under-documented. In this paper, I expand our understanding of Yucatec Maya women’s oral literature by providing a systematic documentation and description of u t’aan nukuch máak, words of the elders, a ‘speech’ genre that relies largely on performance. The performance of u t’aan nukuch máak are an embodiment of my female collaborators’ culture, as they occur in their daily routines. U t’aan nukuch máak are performed (or uttered) in the context of certain bodies, objects, times, and spaces that index concepts that reflect the strength of Maya cultural memory.

Memory for rewards guides retrieval

Rewards paid out for successful retrieval motivate the formation of long-term memory. However, it has been argued that the Motivated Learning Task does not measure reward effects on memory strength but decision-making during retrieval. We report three large-scale online experiments in healthy participants (N = 200, N = 205, N = 187) that inform this debate. In experiment 1, we found that explicit stimulus-reward associations formed during encoding influence response strategies at retrieval. In experiment 2, reward affected memory strength and decision-making strategies. In experiment 3, reward affected decision-making strategies only. These data support a theoretical framework that assumes that promised rewards not only increase memory strength, but additionally lead to the formation of stimulus-reward associations that influence decisions at retrieval.

Credit Risk Modelling Using RNN-LSTM Hybrid Model for Digital Financial Institutions

In response to the rapidly evolving financial market and the escalating concern surrounding credit risk in digital financial institutions, this project addresses the urgency for accurate credit risk prediction models. Traditional methods such as Neural network models, kernel-based virtual machines, Z-score, and Logit (logistic regression model) have all been used, but their results have proven less than satisfactory. The project focuses on developing a credit scoring model specifically tailored for digital financial institutions, by leveraging a hybrid model that combines long short-term memory (LSTM) networks with recurrent neural networks (RNN). This innovative approach capitalizes on the strengths of the Long-Short Term Memory (LSTM) for long-term predictions and Recurrent Neural Network (RNN) for its recurrent neural network capabilities. A key component of the approach is feature selection, which entails extracting a subset of pertinent features from the credit risk data using RNN in order to help classify loan applications. The researcher chose to use data from Kaggle to study and compare the efficacy of different models. The findings reveal that the RNN-LSTM hybrid model outperforms other RNNs, LSTMs, and traditional models. Specifically, the hybrid model demonstrated distinct advantages, showcasing higher accuracy and a superior Area Under the Curve (AUC) compared to individual RNN and LSTM models. While RNN and LSTM models exhibited slightly lower accuracy individually, their combination in the hybrid model proved to be the optimal choice. In summary, the RNN-LSTM hybrid model developed stands out as the most effective solution for predicting credit risk in digital financial institutions, surpassing the performance of standalone RNN and LSTM models as well as traditional methodologies. This research contributes valuable insights for banks, regulators, and investors seeking robust credit risk assessment tools in the dynamic landscape of digital finance.

Dynamic Recommendation Based on Graph Diffusion and Ebbinghaus Curve

Nowadays, many dynamic recommendations still suffer from the insufficiency of finding user online interest evolving patterns because of those complicated interactions. In general, each interaction is usually impacted by multiple underlying reasons, which needs us to open the “box” of each interaction instance instead of simply treating them as a pair-wise link. Besides, different users usually perform differently for their long-term and short-term tastes, leaving traditional sequential models far from personalized. In this article, we propose a novel recommendation model based on Graph Diffusion and Ebbinghaus Curve. Specifically, to explore the underline reasons for different interactions, we explore an underlying sub-graph for each interaction and find important reasoning paths within the sub-graph via a well-designed graph diffusion method. To capture users’ personalized strategies on long-term and short-term tastes, we are inspired by the Ebbinghaus Curve, which can naturally describe users’ memory patterns, and design an effective neural network to process users’ evolving behaviors. We conduct extensive experiments on four real-world datasets and the results further confirm the superiority of our model compared with existing state-of-the-art baselines.

Accessibility of working memory representations in the focus of attention: Heightened or reduced?

Many researchers agree that information residing in the focus of attention in working memory benefits from a boost in memory strength and activation, as well as heightened accessibility. However, recent studies have questioned this heightened accessibility. More specifically, these recent studies found reduced accessibility for an item in the focus of attention compared to another item in working memory, which was referred to as an "inhibition-of-return-like" effect. Our study aimed to provide a detailed examination of the accessibility of information in the focus of attention. Across a series of experiments, varying task characteristics related to the time course of the effect (Experiments 2-3) and the potential role of response inhibition (Experiments 4a-4b), we repeatedly failed to find evidence for an inhibition-of-return-like effect. Instead, we mostly found heightened accessibility for an item in the focus of attention. Given that an inhibition-of-return-like effect seems to be limited to a very specific task condition, reduced accessibility of information in the focus of attention appears to be far from a general phenomenon. Therefore, based on our findings, we propose that information in the focus of attention generally benefits from heightened accessibility, although there may be factors, most likely unrelated to the functioning of the focus of attention, that could sometimes mask this and even result in data patterns that are consistent with reduced accessibility. The theoretical implications for the focus of attention and working memory are discussed. (PsycInfo Database Record (c) 2024 APA, all rights reserved).

Thermal-mechanical enhancement investigation of shape memory polymer nanocomposite for anti-/deicing

The oriented arrangement of graphene oxide (GO) and magnetic graphene oxide (MGO) in shape memory epoxy resin (EP) by employing the coupling effect of electric and magnetic fields were investigated. The heat transfer experiments showed that the thermal conductivity of the aligned shape memory MGO/GO/EP composites was 90.29% higher than that of the random doped composites. Through the two-dimensional aligned of MGO, the tensile strength of the shape memory MGO/GO/EP composite was significantly improved. The tensile strength reaches 24.52 and 22.67 MPa along the X and Y direction respectively, which increased of 53.92% and 42.31%. The shape recovery experiments showed that the shape recovery and shape fixity of the two-dimensional aligned composites were 95.55% and 91.49%, which increased by 4.44% and 7.82%. The results demonstrated that the two-dimensional directional arrangements of MGO layers significantly enhance the heat transfer and mechanical properties of shape memory polymer. The improvement of thermal conductivity and mechanical properties directly affected the anti-deicing performance of the composites. Hence, the deicing time of the shape memory MGO/GO/EP composite was 324 s, which represented an 18.2% improvement of the deicing efficiency.

I remember it now, so I'll remember it later: Working memory strength guides predictions for long-term memory performance.

Judgments of learning (JOLs) are assumed to be made inferentially, based on cues. This cue-utilization approach substituted the theory that memory strength guides JOLs. The rejection of this theory ignores the existence of two memory systems: working memory (WM), which holds representations immediately accessible, and long-term memory (LTM), which is a permanent store. By manipulating and measuring WM strength, we tested a revised version of the memory-strength theory in which JOLs are guided by WM representations. In Experiment 1, participants memorized sequences of two or four colored objects, then they provided JOLs for an LTM test of these objects, and performed a WM test on the objects' colors. After learning 200 objects, the LTM test followed. Sequence-length affected WM, but not LTM performance. JOLs, however, were higher for sequences of two than for four objects and correlated higher with WM than LTM performance. We replicated these results with a simultaneous presentation of the objects (Experiment 2), in the absence of a WM test (Experiment 3), and in a word-pair task (Experiment 4). Overall, our findings are consistent with the revised memory-strength theory. WM strength should therefore be considered when examining the factors guiding JOLs.

MemoryBank: Enhancing Large Language Models with Long-Term Memory

Large Language Models (LLMs) have drastically reshaped our interactions with artificial intelligence (AI) systems, showcasing impressive performance across an extensive array of tasks. Despite this, a notable hindrance remains—the deficiency of a long-term memory mechanism within these models. This shortfall becomes increasingly evident in situations demanding sustained interaction, such as personal companion systems, psychological counseling, and secretarial assistance. Recognizing the necessity for long-term memory, we propose MemoryBank, a novel memory mechanism tailored for LLMs. MemoryBank enables the models to summon relevant memories, continually evolve through continuous memory updates, comprehend, and adapt to a user's personality over time by synthesizing information from previous interactions. To mimic anthropomorphic behaviors and selectively preserve memory, MemoryBank incorporates a memory updating mechanism, inspired by the Ebbinghaus Forgetting Curve theory. This mechanism permits the AI to forget and reinforce memory based on time elapsed and the relative significance of the memory, thereby offering a more human-like memory mechanism and enriched user experience. MemoryBank is versatile in accommodating both closed-source models like ChatGPT and open-source models such as ChatGLM. To validate MemoryBank's effectiveness, we exemplify its application through the creation of an LLM-based chatbot named SiliconFriend in a long-term AI Companion scenario. Further tuned with psychological dialog data, SiliconFriend displays heightened empathy and discernment in its interactions. Experiment involves both qualitative analysis with real-world user dialogs and quantitative analysis with simulated dialogs. In the latter, ChatGPT acts as multiple users with diverse characteristics and generates long-term dialog contexts covering a wide array of topics. The results of our analysis reveal that SiliconFriend, equipped with MemoryBank, exhibits a strong capability for long-term companionship as it can provide emphatic response, recall relevant memories and understand user personality.