---
title: Agent-based Social Memory Architecture
url: https://www.emergentmind.com/topics/agent-based-social-memory-architecture-asma
type: topic
---

# Agent-based Social Memory Architecture

Searching arXiv for recent papers on social memory architectures, multi-agent memory, and related agent systems.
{"query": "\"social memory\" multi-agent memory architecture arXiv agent-based", "max_results": 10}
{"query": "\"Attention Schema in Neural Agents\" arXiv", "max_results": 5}
{"query": "\"AIvilization v0\" arXiv", "max_results": 5}
Agent-based Social Memory Architecture (ASMA) is best understood as an *Editor’s term* for a family of agent designs in which memory is treated as the substrate through which agents preserve identity, regulate attention, consolidate experience, share knowledge, and coordinate with other agents over time. In recent arXiv literature, ASMA-like systems appear in recurrent internal models of attention for multi-agent reinforcement learning, dual-process identity architectures for large artificial societies, hierarchical individual and collective memory systems for agent simulation, governed shared memory layers for enterprise workflows, memory-anchored belief dynamics in social simulation, lifelong semantic-episodic memory for embodied agents, and context-selective multimodal memory for social robots [2305.17375][2602.10429][2507.20215][2603.17787][2605.13725][2506.24019][2604.12081][2511.04235][2203.12446].

## 1. Conceptual scope and lineage

The central premise of ASMA is that memory is not an auxiliary cache but an organizing principle of agency. One line of work formulates this at the level of internal control: attention is a dynamical selector of information, whereas an Attention Schema is “a descriptive and predictive model of attention,” and “a good controller contains a model of the item being controlled” [2305.17375]. Another line formulates it at the level of persistent identity: AIvilization v0 separates short-term execution traces from long-term semantic consolidation so that agents remain “teleologically stable,” “reactively correct,” and “identity-consistent” while still evolving [2602.10429]. A third line formulates it as social infrastructure rather than purely intrapsychic structure: Governed Memory identifies a “memory governance gap” in enterprise multi-agent systems and treats memory as a shared, organization-scoped layer with explicit routing, isolation, schema enforcement, and refinement [2603.17787].

Taken together, these works suggest that ASMA has three recurring commitments. First, memory is stateful and temporally extended rather than prompt-local. Second, memory is socially consequential: it supports coordination, knowledge dissemination, belief persistence, or personalization. Third, memory is operational: it affects planning, action selection, retrieval, governance, and adaptation rather than merely recording prior events. This suggests that ASMA is less a single canonical architecture than a research program spanning internal self-models, per-agent long-horizon autobiographical memory, collective memory repositories, and governed institutional memory.

A further commonality is the rejection of flat, indiscriminate storage. MLC-Agent explicitly argues for hierarchical memory rather than a flat store, using an individual memory set, a memory buffer pool, and a collective memory set [2507.20215]. SUMMER stores only moments with emotional salience or scene novelty [2604.12081]. Ella distinguishes graph-structured semantic memory from spatiotemporal episodic memory [2506.24019]. In each case, the architecture assumes that relevance, selectivity, and consolidation are constitutive properties of memory, not post hoc retrieval heuristics.

## 2. Architectural substrates and memory topologies

ASMA-like systems differ sharply in where memory is located and how it is structured. In the most minimal form, memory can be an internal recurrent control state. In “Attention Schema in Neural Agents,” the internal control module is implemented as an RNN with GRU units and functions as a memory-like predictive representation of attentional dynamics [2305.17375]. In this formulation, memory is not an external database but a recurrent latent state that predicts and regulates attention over time.

A more explicit autobiographical design appears in AIvilization v0. Its Adaptive Agent Profile integrates Dynamic State, Short-Term Memory, and Long-Term Memory. STM is a high-frequency buffer of recent successful and failed actions, outcomes, and reflections; LTM stores values, personality traits, habits, summaries of social interaction records, and other identity features, forming a “stable yet evolving semantic core” [2602.10429]. The architecture therefore separates tactical adaptation from identity-level persistence while coupling them bidirectionally.

MLC-Agent provides a more explicitly social topology. Its memory system comprises an individual memory repository, a group memory repository, and a memory buffer pool, with a memory-learning collaboration mechanism deciding when behavior should be memory-based and when it should rely on learning [2507.20215]. The shared store is not simply pooled replay; it is a filtered collective memory repository collecting experiences, rules, and patterns contributed by multiple agents.

Governed Memory shifts the locus again, from agent memory to organizational memory. Its dual memory model combines open-set atomic facts with schema-enforced typed property values. The first preserves long-tail, temporally anchored, coreference-resolved facts; the second yields structured, queryable memory for downstream workflows, CRM synchronization, conditional routing, and analytics [2603.17787]. In ASMA terms, this is a shared memory commons rather than an autobiographical store.

Other works elaborate richer cognitive taxonomies. ScioMind uses four namespaces—episodic, semantic, procedural, and reflection memory—so that recent interaction context, long-term beliefs, response routines, and after-tick self-reflections coexist in a single social simulation framework [2605.13725]. Ella organizes lifelong memory into a name-centric semantic memory and a spatiotemporal episodic memory, the former grounded in a hierarchical scene graph and the latter storing when, where, and what happened, including text and egocentric images [2506.24019]. SMEMO, by contrast, uses an external differentiable episodic working memory with one write head and multiple read heads, allowing agent-specific information to be continuously written, updated, and recalled during trajectory forecasting [2203.12446].

These memory topologies imply different notions of the “social” in ASMA. In some systems, sociality lies in modeling other agents’ hidden or attentional states; in others, in sharing memories across agents; in others, in binding memory to human users or institutional entities. The architecture category therefore spans internal social cognition, distributed collective memory, and governed cross-agent memory infrastructure.

## 3. Memory operations: encoding, consolidation, retrieval, and forgetting

The defining mechanics of ASMA lie in how memories are created, evaluated, promoted, retrieved, and pruned. MLC-Agent makes these operations explicit. A new memory item is represented as
$$
m_t^i = \{type, o_{t-1}^i, a_t^i, o_t^i\},
$$
and is appended both to the individual store and to the buffer pool. Promotion from buffer to collective memory depends on value error or rarity exceeding thresholds:
$$
m_{\text{selected}} = \{ m_i \in M_t^{buffer} \mid |\delta_t| > \theta_{value} \vee R(m_i) > \theta_{rare} \},
$$
with pruning governed by
$$
M_{t+1} = f_{update}(M_t,k).
$$
Action selection then depends on memory credibility, environmental similarity, success rate, and decay, rather than on unconditional replay [2507.20215]. The result is a memory system that is selective, decaying, and socially disseminative.

AIvilization v0 implements a different but related consolidation logic. Planned action sequences terminate in execution traces that are summarized into concrete experiences; successful patterns refine mappings from sub-goals to actions, whereas failure patterns prime re-planning. Repeated successful STM patterns can consolidate into LTM, while LTM provides top-down context that shapes how STM experiences are interpreted [2602.10429]. This makes memory the mechanism through which identity persists across failures without collapsing into prompt resets.

Ella operationalizes episodic retrieval through a three-factor score over spatial proximity, content relevance, and temporal recency:
$$
\text{proximity}(e,q)=\frac{1}{\|\mathbf{p}_e-\mathbf{p}_q\|+\epsilon},
$$
$$
\text{Relevance}(e,q)=\frac{\cos(T_e,T_q)+\cos(I_e,I_q)}{2},
$$
$$
\text{Recency}(e)=\exp(t_e-t_q).
$$
The normalized scores are averaged and the top-$k$ events are retrieved [2506.24019]. Conversation is also a memory-formation mechanism: after a conversation ends, the summary is stored in episodic memory and extracted knowledge is added to semantic memory.

SUMMER makes selectivity itself the primary mechanism. Emotional salience is computed from emotion-specific thresholds, scene novelty from cosine distance in scene-embedding space, and a moment is stored if either criterion is met:
$$
\text{Memorable}=
\begin{cases}
1, & \text{if } e>0 \text{ or } n>T_n \\
0, & \text{otherwise.}
\end{cases}
$$
Stored traces preserve both image and caption representations, and multimodal retrieval fuses text and image similarity with timestamp alignment [2604.12081]. This design turns episodic storage into a context-selective process rather than an indiscriminate transcript archive.

Governed Memory emphasizes retrieval and write-side governance. It uses cosine duplicate filtering thresholds of 0.92 on writes and 0.95 for background consolidation, reflection-bounded retrieval with a default maximum of 2 rounds, entity-scoped isolation via CRM keys, and progressive context delivery that excludes already delivered governance variables within a session [2603.17787]. A recurrent theme across these systems is that memory quality depends as much on filtering, typing, and isolation as on raw recall capacity.

## 4. Social inference, coordination, and belief dynamics

A major motivation for ASMA is that memory should support social inference, not only private continuity. Attention Schema Theory supplies one of the clearest formulations: an agent’s model of its own attention may be reused to infer the attentional states of other agents. The paper states that “a strong prediction of this theory is that an agent can use its own AS to also infer the states of other agents’ attention and consequently enhance coordination with other agents” [2305.17375]. In the tested architectures, the most AST-consistent design is Hypothesis 5, especially 5.4, where internal control learns to predict attention outputs via contrastive loss and exerts control directly on the attention mechanism by applying a binary mask to attention scores. The reported interpretation is that recurrent internal control provides memory over time, enabling a predictive model of attention that improves control in dynamic multi-agent environments.

ScioMind extends the social-memory idea from attention to opinion dynamics. Its belief update rule combines self-weight, trust-weighted social influence, and memory-derived anchor pull:
$$
b_i^{k,t+1} = (1-\rho_i)\big[(1-\lambda_i)b_i^{k,t}+\lambda_i S_i^{k,t}\big]+\rho_i m_i^{k,t}.
$$
Here the anchor $m_i^{k,t}$ is dynamic and memory-derived rather than a fixed initial opinion, and anchoring strength $\rho_i$ is conditioned on Big Five traits through a sigmoid mapping [2605.13725]. This produces persistent yet adaptive belief trajectories. The reported findings are that dynamic profiles increase opinion diversity, memory and reflection reduce unstable oscillation, and anchoring yields persistent trajectories aligned with patterns reported in political psychology. Removing anchoring causes overly rapid convergence and reduced persistence, whereas anchoring yields the highest polarization at $0.35 \pm 0.04$, moderate extremization at $0.60 \pm 0.05$, and stable radicalization at $0.56 \pm 0.04$ [2605.13725].

Shared spatial memory introduces another route from memory to coordination. In “Shared Spatial Memory Through Predictive Coding,” agents minimize mutual predictive uncertainty through an information bottleneck objective, learning not only who and what to communicate but also when [2511.04235]. A grid-cell-like metric scaffold emerges from self-supervised motion prediction, and specialized units encode partner locations as an artificial analogue of hippocampal social place cells. This turns shared memory into a distributed alignment of internal world models rather than a centralized store.

SMEMO shows that external social memory can also serve as an interpretable interaction model. Its memory segmentation variant assigns memory segments to individual agents, and read weights over those segments define explicit inter-agent attention. Cause-Effect Accuracy on the synthetic SSA dataset reaches 0.71, compared with 0.48 for SR-LSTM and 0.39 for AgentFormer, supporting the claim that memory access patterns can recover which agent influenced which prediction [2203.12446]. ASMA, in this sense, is not only about long-term continuity; it is also about causal structure in social interaction.

## 5. Empirical domains and reported performance

ASMA-like architectures have been evaluated in markedly different problem domains, from cooperative reinforcement learning to enterprise operations. In multi-agent reinforcement learning, “Attention Schema in Neural Agents” tests five hypotheses about the relation between attention and internal control in GhostRun and MazeCleaners, under IID testing, OOD testing, and continual learning. The paper reports that Hypothesis 5 shows marginal advantages over other hypotheses, that “hypothesis five, corresponding to the attention schema theory, achieves the best performance,” and that in MazeCleaners, Hypothesis 5.4 gives the best performance among all architectures [2305.17375].

In artificial society simulation, AIvilization v0 deploys tens of thousands of agents in a resource-constrained sandbox economy and analyzes 600,000+ transaction records, including 400,000 transactions from the mature phase. The environment exhibits stable markets with heavy-tailed returns and volatility clustering, as well as structured wealth stratification driven by education and access constraints [2602.10429]. The architecture’s significance for ASMA lies in showing that persistent identity, dual-process memory, and memory-mediated steering can operate at large social scale rather than only in small interactive demos.

MLC-Agent reports that, compared with existing memory modeling methods, its memory-learning collaboration mechanism yields better decision-making quality and adaptability, improving average daily profit, number of completed orders, effective working time, and system-level order completion rate in a delivery-agent scenario [2507.20215]. Governed Memory validates each of its four mechanisms through controlled experiments on 250 samples across five content types. It reports 99.6% extraction recall, 92% governance routing precision, 88% governance routing recall, 50.3% token reduction from progressive delivery, zero true cross-entity leakage across 3,800 retrieval results, 100% adversarial governance compliance, and 74.8% overall accuracy on LoCoMo [2603.17787]. These results situate ASMA not only in cognitive agent design but also in production workflow infrastructure.

Embodied and social-interactive settings show a parallel pattern. Ella operates in a 3D open world with 15 agents across New York, London, and Detroit scenes; agents undergo a first-day memory-building phase of 34,200 steps and are later evaluated in Influence Battle and Leadership Quest [2506.24019]. The reported outcome is that Ella can influence, lead, and cooperate well, and that structured memory yields more stable and organized growth than Generative Agents’ memory. SUMMER, evaluated on a curated pilot dataset of 81 socially grounded images rated by 25 human raters, achieves a Spearman correlation of 0.506 with human memorability ratings, exceeding the reported human consistency of 0.415, while multimodal fusion improves Recall@1 by up to 13% over unimodal retrieval and remains within real-time latency constraints [2604.12081].

Finally, memory-centered designs also improve classical multi-agent prediction tasks. SMEMO achieves ADE/FDE of 0.169/0.244 and Kendall $\tau=0.827$ on SSA, outperforming GRU encoder-decoder, SR-LSTM, and AgentFormer, and remains competitive or state of the art on SDD and ETH/UCY [2203.12446]. Shared Spatial Memory demonstrates graceful degradation under bandwidth constraints on Memory-Maze: success declines from 73.5% to 64.4% as bandwidth shrinks from 128 to 4 bits/step, whereas a full-broadcast baseline collapses from 67.6% to 28.6% [2511.04235]. Across domains, the common empirical pattern is that structured memory improves robustness under partial observability, long horizons, or social interdependence.

## 6. Boundaries, misconceptions, and open problems

A persistent misconception is that ASMA is equivalent to retrieval-augmented generation or long chat history. The literature argues against this repeatedly. Governed Memory states that RAG is not enough because retrieval relevance does not provide shared write-side governance, schema enforcement, session-aware delivery, or quality feedback [2603.17787]. AIvilization treats memory as the medium through which steering propagates, rather than as a brittle prompt override [2602.10429]. Attention Schema work treats the recurrent internal model not as passive storage but as inner regulation of attention itself [2305.17375]. ASMA therefore refers to architectures in which memory has causal authority over behavior.

A second boundary concerns the meaning of “social.” In some systems, social memory is explicit group-level storage and dissemination, as in MLC-Agent’s collective memory repository [2507.20215]. In others, it is memory about users, as in SUMMER’s user-indexed multimodal traces [2604.12081]. In others, it is memory-shaped belief persistence and influence susceptibility, as in ScioMind [2605.13725]. In yet others, it is entity-scoped organizational memory across workflows [2603.17787]. This heterogeneity suggests that ASMA is a cross-cutting architectural category rather than a single ontology.

The current literature also exposes several limitations. The Attention Schema work mainly models attention rather than a full social memory of agent interactions, and it does not explicitly model other agents’ hidden internal states beyond the theoretical motivation [2305.17375]. AIvilization notes that its long-term identity model remains relatively coarse, that guidance and social mobility analyses are correlational rather than causal, and that computational cost limits scaling to millions of agents [2602.10429]. MLC-Agent does not deeply formalize trust relations, network topology, or explicit inter-agent memory governance [2507.20215]. Ella acknowledges that its graph-structured semantic memory is currently retrieved mainly by text and image similarity rather than more sophisticated graph reasoning [2506.24019]. SUMMER’s pilot dataset is synthetic, and social relevance is inferred indirectly rather than explicitly modeled [2604.12081]. ScioMind captures structural controversy in policy-debate simulations but underestimates emotional intensity in the Roe v. Wade case study [2605.13725].

These limitations indicate that a mature ASMA theory would need to unify several strands that are currently separate: internal self-modeling, autobiographical memory, group memory promotion, retrieval governance, multimodal grounding, and social-state inference. A plausible implication is that future ASMA systems will be evaluated less by isolated recall metrics than by whether memory yields persistent yet revisable identity, robust coordination under uncertainty, bounded and governable sharing, and interpretable social causality across long horizons.

Source: https://www.emergentmind.com/topics/agent-based-social-memory-architecture-asma