Short-term memory is a primary cognitive function of the brain. The transitions between the spontaneous and persistent states could lead to the emergence and disappearance of short-term memory in a bistable system (Amit and Brunel, 1997). Continuous neural activity without external inputs was deemed an expression of short-term memory (Wang, 2001). A phenomenological model of spatial working memory was developed to examine the dynamical interactions of multiple feedback mechanisms (Carter and Wang, 2007). A growing body of evidence suggests memories may be kept through the mutual effect of persistent neural activity and activity-silent dynamics (Stokes, 2015; Barbosa et al., 2020). Gaussian noise was treated as an essential factor in neuronal activity and its toggle switch (memory maintenance) (Zheng et al., 2020b). Memory maintenance through persistent neural activity and a synaptic mechanism was compared in mice and two types of artificial neural networks to show their differences (Hu et al., 2021). Then, computational modeling was constructed to prove how the circuits and networks affect working memory, which provides a novel theory for memory maintenance (Ghazizadeh and Ching, 2020). In addition, the encoding style of the input information of the short-term memory was investigated to illustrate the dynamical mechanisms of short-term memory (Ichikawa and Kaneko, 2021; Jones and Ching, 2022; Zhou et al., 2023). Hamilton energy, representing the utilization of energy (actual energy in the generation of short-term memory), should be considered to illustrate the dynamic mechanism of the generation of short-term memory.

Hindmarsh–Rose (HR) model (1) (x is the membrane potential, y is the recovery variable of the fast current of K ⁺ or Na ⁺, and z is the adaptation variable of the slow current of Ca ⁺ or other ions) was proposed to show the membrane potential of neuronal activity (Hindmarsh and Rose, 1982), which has rich dynamical behaviors (Wang et al. 2021a; Wang et al. 2021b; Wang et al. 2022). d x d t = y − a x 3 + b x 2 + I ext − z , d y d t = c − d x 2 − y , d z d t = r s x − x r − z . (1)

The synchronization and bifurcation (Song and Xu, 2013; Liebovitch et al., 2011; Kumar et al., 2016; Zheng et al. 2020a; Zheng et al. 2024; Zheng et al. 2023) of the HR model were often studied to demonstrate the dynamical mechanism of chaotic bursting or spikes (Shi and Wang, 2012; Wu et al., 2016; You, 2023a). The interplay between neurons was analyzed to present the effect of the parameters and coupling strength on the appropriate functioning of the system (Lepek and Fronczak, 2018; Rajagopal et al., 2019; You, 2023b). Energy is necessary for neuron activity (Attwell and Laughlin, 2001). The Hamiltonian energy function is a vital tool to evaluate energy consumption when neurons are active (Torrealdea et al., 2006). The average energy consumption of the HR model was given to display the energy consumption ratio in different situations, which could help optimize energy use (Torrealdea et al., 2009; Song et al., 2015; Usha and Subha, 2019). The Hamilton energy balance of different functional neurons was discussed through the coupling strength, which contributes to designing functional assistive devices (Zhang et al., 2022; Yu et al., 2023). Although the HR model could explain the generation of short-term memory (Zheng et al., 2022), the utilization of energy should be further stated in short-term memory.

Short-term memory results from neuronal activity coming with a change in energy, and a physical neuron circuit plays a vital role in the synergistic effect of neurons and the generation of short-term memory. In this paper, the pattern formation of a modified HR model is investigated to find the dynamical mechanism of how the coupling strength and links (degree) affect the generation of short-term memory. The Hamiltonian energy function is derived in the HR model with a single node, which means the energy consumption varies at different states of neuronal activity. Then, the degree and stimuli from other neurons are studied through bifurcation, which means the energy is necessary to generate short-term memory. Finally, the related dynamical and biological mechanisms are obtained.

As the membrane potential of neurons is often coupled with others, the following network-organized HR model is introduced: d x i d t = y i − a x i 3 + b x i 2 + I ext − z i + d 1 ∑ j = 1 n L i j t x j , d y i d t = c − d x i 2 − y i , d z i d t = r s x i − x r − z i , (2)

where x _i is the membrane potential, I represents the ith neuron and i = 1, … , n, y _i is the recovery variable of the fast current of K ⁺ or Na ⁺, and z _i is the adaptation variable of the slow current of Ca ⁺ or other ions. D ₁ is the coupling strength between neurons. L _ij (t) = A _ij – δ _ij k _i , where A _ij is the adjacent matrix and k _i is the degree of the ith node.

In order to obtain the Hamiltonian energy function of system (2), we consider a simplified model with a single node through the mean-field approach (McCullen and Wagenknecht, 2016). The reduced system is d x i d t = y i − a x i 3 + b x i 2 + I ext − z i + d 1 k i x 0 − x i , d y i d t = c − d x i 2 − y i , d z i d t = r s x i − x r − z i , (3)

where (x ₀, y ₀, z ₀) (Zheng et al., 2022) is the equilibrium point of system (2) without a network, and x ₀ makes I _ext = f(x) = −c + dx ² + ax ³ − bx ² + s (x − x _r ) + I _ext + d ₁ k _i (x ₀ − x _i ) hold. In addition, x ₀ is the external stimulus from other neurons.

In this paper, we mainly investigate the dynamical behaviors of system (3) and its Hamiltonian equation. According to Helmholtz’s theorem (Donald and Rose, 1986), an autonomous ordinary differential equation X ̇ = F ( X ) (F(X) can be treated as the velocity vector field) can be described in the usual forms of a Hamiltonian equation: X ̇ = G X ∇ H X ,

where G(X) is a skew-symmetric matrix in the Hamiltonian system. If G(X) is not a skew-symmetric matrix in a generalized Hamiltonian system, G(X) can be divided into two parts G(X) = G ₁(X) + G ₂(X): a skew-symmetric matrix G ₁(X) and a symmetric matrix G ₂(X) (Sarasola et al., 2004). H(X) is an energy function. Then, we have X ̇ = G 1 X + G 2 X ∇ H X .

For the network-organized HR model (3), it can be written as (Torrealdea et al., 2006; Torrealdea et al., 2009; Song et al., 2015; Usha and Subha, 2019) x ̇ i y ̇ i z ̇ i = G 1 x i , y i , z i + G 2 x i , y i , z i ∇ H x i , y i , z i , where F 1 x i , y i , z i = G 1 x i , y i , z i ∇ H x i , y i , z i = y i − z i + d 1 k i x 0 − x i − d x i 2 r s x i , F 2 x i , y i , z i = G 2 x i , y i , z i ∇ H x i , y i , z i = − a x i 3 + b x i 2 + I ext − k i x i c − y i − r s x r − r z i . Furthermore, we obtain the Hamiltonian energy function as (Zhang et al., 2022) ∇ H x i , y i , z i T F 1 x i , y i , z i = 0 , H ̇ x i , y i , z i = ∇ H x i , y i , z i T F 2 x i , y i , z i . Namely, y i − z i + d 1 k i x 0 − x i ∂ H x i , y i , z i ∂ x i − d x i 2 ∂ H x i , y i , z i ∂ y i + r s x i ∂ H x i , y i , z i ∂ z i = 0 , where a general solution can be expressed as H x i , y i , z i = y i − z i + d 1 k i x 0 − x i 2 + 2 3 d x i 3 + r s x i 2 , and H t = H ̇ x i , y i , z i = 2 r s x r + 2 r z i + 2 c − 2 y i d 1 k i x 0 − x i − 2 a d x i 5 + − 2 a r s + 2 b d x i 4 + 2 b r s − 2 d k i x i 3 + − 2 k i r s + 2 I ext d x i 2 + 2 I ext r s x i + 2 r s x r y − 2 r s x r z i + 2 r y i z i − 2 r z i 2 + 2 c y i − 2 c z i − 2 y i 2 + 2 y i z i .

In this section, the finite difference method is applied to find numerical solutions for the network-organized HR model (3) with time step dt = 0.01. These parameters a = 1, b = 3, c = 1, d = 5, r = 0.01, s = 4 are set (Zheng et al., 2022). The small-world network is constructed with W (n, K, p) (the number of node n, nearest neighbor K, and reconnection probability p), which can be found can be found at https://github.com/zhengqianqian35/network-code. We give the concept of the average Hamiltonian energy for time and nodes: H 1 = ∫ t 0 t 0 + T H x i , y i , z i d t T , H 2 = m a x H ̇ x i , y i , z i , m i n H ̇ x i , y i , z i ,

where the integration period is set at T = 5,000 time units. In order to exclude the influence of initial conditions, t ₀ is the starting time of the cycle after the system tends to a stable state. In addition, we assume H = H (x _i , y _i , z _i )/100. From Figure 1, only one real equilibrium point (x ₀, y ₀, z ₀) exists in system (1) when a = 1, b = 3, c = 1, d = 5, r = 0.01, s = 4, which guarantees the uniqueness of system (3).

In general, the generation of neural function results from the collaboration of multiple neurons. Therefore, we consider the strength d ₁ of the coupling between neurons and the number of links K. First, a small-world network with W (100, 8, 0.01) is given. The pattern formation is chaotic (Figure 2A) when the strength d = 0.01 is weak, which means the nervous system does not work. The pattern formation starts to become clear and tends to sync with the increase of d ₁ (Figure 2b,c). Ultimately, the pattern formation becomes synchronized; namely, all the neurons become completely phase synchronized (Figure 2D). The short-term memory needs to be clarified when d ₁ is weak. Only when all the neurons work perfectly together is a clear short-term memory formed (Figure 2), which is also the mechanism by which adequate short-term memory is produced.

Then, the number of cooperative neurons will be considered to generate short-term memory. The links between neurons can be treated as the number of collaborative neurons in our analysis, which could be measured by K. When K is small, the pattern formation is chaotic, and every neuron is relatively independent (Figure 3A). This condition is not suitable for the generation of short-term memory. When K = 2, the pattern formation shows some neurons are in sync (Figure 3B); namely, multiple short-term memories are produced simultaneously. In this case, short-term memory is often fuzzy. The short-term memory gradually becomes clear when K becomes large (Figure 3C). Eventually, multiple neurons work together to form a clear short-term memory when all the neurons are in sync (Figure 3D).

Finally, it is found that the link probability does not work because p cannot change the number of cooperative neurons and the coupling strength.

Energy plays a vital role in neuronal activity, which is the basis of the generation of short-term memory. In this paper, the pattern formation could represent the collecting dynamics of short-term memory through the Hamiltonian energy, showing the neuronal activity in generating short-term memory. Therefore, the interplay between neurons is considered through a simple network to show the effect of the external stimulus and coupling strength (degree) on the dynamical behaviors. It is found that the Hamiltonian energy, change in the Hamiltonian energy, and membrane potential are consistent. The excessive influence of other neurons can make the ith neurons lose their dominance and align with the dynamic behaviors of other neurons, which could show the synergistic effect of neurons through a physical neuron circuit. In addition, the energy from other neurons is necessary for short-term memory, proving that short-term memory results from multiple neuronal activities. Generating short-term memory requires much energy, and energy requires a process of accumulation to maintain a complete spike. Meanwhile, the coupling strength could further reduce energy consumption, which provides a novel way to reduce the energy consumption in information storage and processing. However, more short-term memory descriptions should be completed next.