Dynamic Collective Memory

• Dynamic Collective Memory is the emergent phenomenon where distributed agents (biological, social, or artificial) collectively store, sustain, and retrieve past information through structured interactions.
• Mathematical models use agent-based reinforcement, Hopfield associative memory, and network contagion to capture phase transitions and mnemonic convergence, with key thresholds dictating memory persistence.
• DCM has practical applications in designing multi-agent AI, optimizing social network communication, and understanding biological collective behavior through temporal decay, triggering models, and critical phase transitions.

Dynamic Collective Memory (DCM) refers to the emergent, system-level mechanisms by which distributed agents—biological, social, or artificial—store, sustain, and utilize information about past states or stimuli in ways that shape future collective behavior. DCM emphasizes that memory is not simply the sum of individuals’ memories, but arises from structured interactions, co-evolving with network topology, environmental context, and external signals. Across disciplines, DCM is formalized in models that include agent-based reinforcement, associative or Hopfield-like coupling, bifurcation-induced hysteresis, network contagion, and mean-field phase transitions, all with explicit mathematical and algorithmic detail.

1. Foundational Models: Agent-Based Dynamics and Associative Memory

The core mechanism of DCM in opinion dynamics is built on coupled stochastic differential equations for agent preference fields $u_i(t)$. Each agent’s opinion is updated according to: $$\dot{u}_i = -u_i + I_i + \sum_{j \neq i} J_{ij}(t) g_j + \eta_i(t)$$ where $u_i$ mean-reverts, $I_i$ encodes external fields (news), $g_j$ is the bounded expressed opinion, $J_{ij}(t)$ encodes memory as an exponentially weighted average of past co-opinions, and $\eta_i$ adds white noise (Boschi et al., 2019).

The evolution of the coupling matrix is central: $$J_{ij}(t) = \left(\frac{J_0 \gamma}{N}\right) \int_0^t e^{-\gamma(t-s)} g_i(s) g_j(s) ds$$ This formalism is equivalent to the temporal Hebbian rule underlying Hopfield associative memory models, resulting in the embedding of past opinion patterns as attractor states. For repeated exposure to $p$ patterns,

$$J_{ij}^\infty \approx \frac{J_0}{Np} \sum_{\mu=1}^p \xi^\mu_i \xi^\mu_j$$

retrieval proceeds via a self-consistent overlap equation equivalent to Hopfield/Mattis neural network retrieval conditions.

Phase transitions govern memory formation: persistent memory emerges only if the time-averaged coupling strength or the duty-cycle of exposure exceeds a critical value $$\dot{u}_i = -u_i + I_i + \sum_{j \neq i} J_{ij}(t) g_j + \eta_i(t)$$0. Below this threshold, memories fade; above it, spontaneous retrieval occurs even after signal removal (Boschi et al., 2019).

2. Network Structure, Communication Patterns, and Mnemonic Convergence

DCM in social and experimental networks demonstrates that the emergence and structure of collective memory depend acutely on the temporal ordering of interactions and the architecture of inter-agent ties. In controlled networks of human subjects, mnemonic convergence (mean Jaccard similarity of individual memory vectors) increases more when “weak ties”—cross-cluster connectors—are activated early during conversational sequences, as opposed to “strong ties” (intra-clique links) (Momennejad et al., 2017).

Mathematically, the reachability of memory states is modeled by a convolution over contact matrices: $$\dot{u}_i = -u_i + I_i + \sum_{j \neq i} J_{ij}(t) g_j + \eta_i(t)$$1 where $$\dot{u}_i = -u_i + I_i + \sum_{j \neq i} J_{ij}(t) g_j + \eta_i(t)$$2 encodes the network at round $$\dot{u}_i = -u_i + I_i + \sum_{j \neq i} J_{ij}(t) g_j + \eta_i(t)$$3, $$\dot{u}_i = -u_i + I_i + \sum_{j \neq i} J_{ij}(t) g_j + \eta_i(t)$$4 is a recency parameter, and $$\dot{u}_i = -u_i + I_i + \sum_{j \neq i} J_{ij}(t) g_j + \eta_i(t)$$5 is a practice-decay parameter. Simulation and empirical calibration show that early weak-tie activation maximizes cross-community integration of memories, while strong-tie-first schedules lead to local echo chambers and fragmented mnemonic landscapes.

3. Memory Webs, Cognitive Networks, and Group-Level Phase Structure

Individual and collective memories are also modeled as directed event-graphs—memory webs—where nodes are events and edges indicate inferred causal or temporal links (Lee et al., 2010). Agents repeatedly reinforce event associations via stochastic communication and update social tie strengths. Forgetting and noise are implemented via uniform arc decay and random rewiring.

The system exhibits bifurcating phase regimes in the mean-field limit: $$\dot{u}_i = -u_i + I_i + \sum_{j \neq i} J_{ij}(t) g_j + \eta_i(t)$$6 with consensus threshold $$\dot{u}_i = -u_i + I_i + \sum_{j \neq i} J_{ij}(t) g_j + \eta_i(t)$$7, where $$\dot{u}_i = -u_i + I_i + \sum_{j \neq i} J_{ij}(t) g_j + \eta_i(t)$$8 is communication rate and $$\dot{u}_i = -u_i + I_i + \sum_{j \neq i} J_{ij}(t) g_j + \eta_i(t)$$9 is decay. Above $u_i$0, global consensus emerges; below, multiple stable subgroups develop distinct memory webs. Empirically and in simulations, the number of such subgroups saturates as $u_i$1 grows, establishing a bounded-memory regime: large populations yield only a finite number of persistent collective memories (Lee et al., 2010).

4. Decay, Triggering, and Recurrence: Temporal Mechanisms of DCM

Quantitative modeling of DCM accounts for the rapid loss and long-tail persistence of social attention. Two-step and two-phase decay models decompose the normalized attention $u_i$2 into fast (communicative) and slow (cultural) terms: $u_i$3 (Candia et al. (Candia, 2022)) or

$u_i$4

(Igarashi et al. (Igarashi et al., 2022)).

The crossover (switching) time $u_i$5, where long-term cultural memory overtakes communicative memory, is analytically tractable: $u_i$6 Empirical fitting across event types (earthquakes, deaths, aviation disasters, mass/casualty events) yields a universal transition at $u_i$7 days (Igarashi et al., 2022). This quantitatively defines a collective attention span and frames the time window for policy intervention or sustained engagement.

Triggering models demonstrate that new events cause cascades of recall for prior, similar events via multiplicative flows: $u_i$8 where $u_i$9 is exposure to the new event, $I_i$0 captures pre-existing memory strength, and $I_i$1 is a function of temporal, categorical, and hyperlink similarity (García-Gavilanes et al., 2016). Total secondary attention frequently exceeds primary attention, generating large-scale network cascades of dynamic memory.

5. DCM in Artificial and Biological Multi-Agent Systems

In decentralized AI systems and natural collectives, DCM emerges through the interplay of internal agent memory and environmental or stigmergic traces (Khushiyant, 10 Dec 2025), and through macroscopic state-retention in noisy phase transitions (Chan et al., 21 Jul 2025). In multi-agent AI, the joint dynamical system $I_i$2 tracks agent-local and environmental traces, both obeying category-specific decay and update rules. The system exhibits a mean-field phase transition: at low agent density $I_i$3, individual memory dominates, while above $I_i$4, stigmergic coordination via persistent environmental traces vastly improves group-level performance (36–41% gains) and resilience. This identifies DCM as the performance envelope of agent-trace coupling (Khushiyant, 10 Dec 2025).

In biological systems, algorithms derived from collective animal motion (e.g., schooling fish) display DCM via noisy bifurcation-induced hysteresis: group polarization $I_i$5 obeys a normal form with additive noise,

$I_i$6

Memory emerges not from deterministic bistability but as exponentially extended group states near the bifurcation due to Kramers-type escape times, yielding history-dependent, path-retentive collective behavior (Chan et al., 21 Jul 2025).

6. Generalities and Logical Frameworks for DCM

Abstract models formalize DCM as logical processes on agent neighborhoods and transition relations, establishing that every DCM protocol corresponds to a semi-linear predicate and can be modeled with population protocols and linear-time temporal logic (Ramanujam, 2021). The emergence of stable collective memory is associated with reachability and stability properties in such logics, making model-checking tractable for finite configurations.

Discrete-time renewal-resetting processes further exemplify DCM: population-level coupling (rank-based reset bias) imparts non-Markovian macroscopic memory even when microscopic resets are Poissonian. A first-order phase transition at critical bias $I_i$7 separates stationary from aging regimes, with persistent macroscopic localization and slow algebraic growth of “condensate” memory fractions (Vilk, 20 Jan 2026).

7. Applications, Quantitative Analysis, and Design Principles

DCM models yield specific operational guidelines. In communication networks, early cross-cluster dialogue (weak ties) minimizes polarization and maximizes shared recollection (Momennejad et al., 2017). In multi-agent AI, critical density thresholds and analytic eigenvalue conditions determine when to allocate resources to agent memory versus environmental traces (Khushiyant, 10 Dec 2025). In information campaigns, the time-dependent decay law quantitatively defines intervention windows for maximizing social memory or rapid amnesia (Igarashi et al., 2022, Candia, 2022).

Table: Representative DCM Mechanisms Across Domains

Domain	DCM Implementation	Phase Transition / Criticality
Social/opinion networks	Exponential Hebbian updating, Hopfield retrieval	Critical coupling $I_i$8
Experimental human groups	Mnemonic convergence, weak/strong tie scheduling	Temporal order governs convergence
Web/Wikipedia attention	Two-phase decay, cascading recall	Universal switching point $I_i$910d
Multi-agent AI	Memory/trace coupling, consensus weighting	Density threshold $g_j$0
Animal collectives	Noisy bifurcation, hysteresis	Kramers escape, stochastic memory timescale
Renewal resetting	Rank-biased resets, localization	Critical bias $g_j$1, first-order dynamic PT

These systems all exhibit DCM through reinforcement, memory decay, coupling, and critical phase structure, with model parameters and empirical predictions available for system-level tuning and control.

• (Boschi et al., 2019) Opinion dynamics with memory: how a society is shaped by its own past
• (Momennejad et al., 2017) The Ties that Bind Networks: Weak Ties Facilitate the Emergence of Collective Memories
• (Lee et al., 2010) Emergence of collective memories
• (Igarashi et al., 2022) A two-phase model of collective memory decay with a dynamical switching point
• (Candia, 2022) Quantifying Collective Memories
• (Khushiyant, 10 Dec 2025) Emergent Collective Memory in Decentralized Multi-Agent AI Systems
• (Chan et al., 21 Jul 2025) Noise-Induced Collective Memory in Schooling Fish
• (Vilk, 20 Jan 2026) Macroscopic localization and collective memory in Poisson renewal resetting
• (Ramanujam, 2021) Reasoning about Emergence of Collective Memory
• (García-Gavilanes et al., 2016) Memory Remains: Understanding Collective Memory in the Digital Age
• (Boschi et al., 2020) Opinion dynamics with emergent collective memory: the impact of a long and heterogeneous news history
• (Miz et al., 2017) Wikipedia graph mining: dynamic structure of collective memory
• (Graus et al., 2017) The Birth of Collective Memories: Analyzing Emerging Entities in Text Streams
• (Yu et al., 28 Jan 2026) Remember Me, Not Save Me: A Collective Memory System for Evolving Virtual Identities in Augmented Reality

Dynamic Collective Memory constitutes a quantitatively tractable, mathematically rich, and cross-disciplinary field, with broad applicability to social computation, AI, network science, and beyond.