From Private Experience to Collective State: A Survey of Multi-Agent Memory Formation and Maintenance

Yiqi Wang, Jiaqi Zhang, Zhangkai Wu, Mingkai Zheng, Zequn Sun, Yiqun Duan, Zirui Liu, Zhihao Hao, Taotao Cai

📌 Contents

• Overview
• Why Collective State?
• Core Concepts
• Survey Contributions
• OTGE Framework
• Corpus at a Glance
• Key Findings
• Evaluation
• Research Agenda
• Survey Roadmap
• Representative Works by Theme
• Figures
• Citation

💡 Overview

Research on memory for large language model (LLM) agents is usually organized around representation, storage, retrieval, and use. That framing is useful for a single agent, but it is incomplete for a team. Persistence tells us whether information survives, and an access architecture tells us whether another agent can reach it. Neither tells us whether that agent should rely on the information, who is authorized to rely on it, how disagreements should be resolved, when reliance should stop, or how downstream damage should be repaired.

We call this missing layer the reliance gap and shift the unit of analysis from memory stores to the decisions through which agent teams form and maintain persistent collective state.

Shared Access ≠ Collective State. A shared store is neither necessary nor sufficient for a team to have a governed, reliable memory.

🤔 Why Collective State?

Three properties vary independently in an agent team:

Question	What it establishes
Does the information survive?	Persistence
Can another agent reach it?	Accessibility
Should another agent act on it, and for how long?	Governed reliance

This survey focuses on the third property. It defines memory-mediated collective state as persistent, memory-represented information that stands in a cross-agent relation and is collectively consequential:

$$ S_C^M(t)={m\mid \mathrm{Persistent}(m)\wedge\mathrm{MemoryMediated}(m)\wedge\mathrm{CrossAgentRelation}(m)\wedge\mathrm{CollectivelyConsequential}(m)}. $$

Collective state need not be centralized, globally visible, or identically replicated. It can be maintained through role-scoped digests, handoff summaries, governed shared stores, or divergent per-agent replicas with reconciliation.

🤓 Core Concepts

Concept	Meaning in this survey
Memory-mediated collective state	Persistent memory-represented information whose use, non-use, or failure affects agents beyond its producer.
Shared memory	An access architecture. It overlaps with collective state but does not determine what may be relied upon.
Persistent agent state	The broader space that also includes workflow state, permissions, credentials, tool state, and external commitments.
Admission	The decision that makes candidate information eligible for collective reuse.
Authorized reach	Who may or should rely on an item.
Realized reach	Who actually receives or can retrieve the item.
Reconciliation	Resolution of divergent views across agents, roles, or replicas.
Revision	Temporal change along one informational lineage.
Revocation	Withdrawal of the validity or reliance authority of previously admitted state.
Recovery	Repair of affected state, holders, and decisions after an error has propagated.

Six distinctions keep the survey boundary precise:

• Shared memory ≠ collective state
• Scoping ≠ propagation
• Reconciliation ≠ revision
• Scope restriction ≠ revocation
• Collective state ⊂ persistent agent state
• Communication ≠ collective-state formation

🎉 Survey Contributions

This survey makes four main contributions:

1. A framework with defensible boundaries. It gives a membership criterion for memory-mediated collective state, six non-negotiable distinctions, and an attribute model that makes the seven operators disjoint and comparable.
2. A coded corpus and a quantitative landscape. It codes 89 LLM-agent papers on six facets, counting an operator only when a decision is present rather than merely a data structure.
3. Four conceptual constructs not supplied by the literature. It develops a collective-utility decomposition, a representation-centered conflict typology, an account of which distributed-systems guarantees transfer to natural-language state, and a three-part model of recovery closure.
4. A vector-valued evaluation view and evidence-derived agenda. It proposes nine evaluation dimensions and five research programs derived from operator × guarantee × measurement gaps.

🧭 OTGE Framework

The Object-Transition-Guarantee-Evaluation (OTGE) framework models collective state as

$$ S_C^M \xrightarrow{;T;} {S_C^M}' \quad \text{subject to } G,\ \text{evaluated by } Q. $$

The operators are composable and non-sequential: a running system can alternate among them and apply several to one item in the same episode.

Phase	Operator	Core question	Typical failure
Formation	Admission	Should this become eligible for collective reuse?	Polluted or low-utility memory
Formation	Scoping	Who may or should rely on it?	Overexposure, underexposure, or scope drift
Formation	Propagation	Who actually receives it?	Missing, delayed, stale, or unauthorized delivery
Formation	Reconciliation	What should the group believe?	Persistent or misclassified conflict
Maintenance	Revision	How does one informational lineage evolve?	Stale or superseded state reuse
Maintenance	Revocation	When should agents stop relying on it?	Zombie memory and incomplete invalidation
Maintenance	Recovery	How is propagated damage repaired?	Incomplete state, holder, or decision closure

Six guarantees constrain these operators:

Guarantee	Requirement
Provenance	Preserve a recoverable record of origin, transformation, and propagation.
Attribution	Assign an item or responsibility for it to an agent, source, or human principal.
Authority	Specify which principal may apply which operator to which item.
Consistency	State how agents' views of the same underlying fact are allowed to agree or diverge.
Privacy	Prevent reliance outside the item's appropriate context.
Confidence	Attach and propagate epistemic uncertainty with the item.

Cost is not a seventh guarantee. Tokens, latency, storage, communication, verifier calls, and repair overhead are charged against every operator as resources consumed.

📊 Corpus at a Glance

The paper conducts a structured curated review covering January 2019 through August 21, 2026. Searches span DBLP, Semantic Scholar, the ACM Digital Library, IEEE Xplore, OpenReview, and arXiv, followed by backward and forward snowballing to closure.

Corpus statistic	Count
Primary coded papers	89
System papers	62
Analyses or attacks	16
Benchmarks	11
Papers reporting a task metric	66
Papers reporting a retrieval metric	14
Papers reporting any collective-state property metric	31
Papers using a named benchmark	32
Papers not stating who controls implemented memory operations	80

Transition coverage

Operator	Papers	Share of corpus
Admission	77	86.5%
Revision	52	58.4%
Propagation	36	40.4%
Scoping	25	28.1%
Reconciliation	22	24.7%
Revocation	12	13.5%
Recovery	6	6.7%

Widest collective scope

Scope	Papers	Share of corpus
Individual	48	53.9%
Team	23	25.8%
Organization	10	11.2%
Role	5	5.6%
Dyadic	3	3.4%
Global	0	0.0%

Operator coding is multi-label, whereas scope is assigned as a single widest level. Foundational pre-LLM work is cited for theory but is not included in the 89-paper coded corpus.

🔎 Key Findings

• Coverage thins by an order of magnitude from formation to repair. Admission appears in 77 papers, but revocation in only 12 and recovery in 6.
• The gap is structural, not merely chronological. Revocation and recovery depend on lineage, validity, scope, and derivation records that must be established while collective state is formed.
• Shared access does not imply managed reliance. More than half of the corpus places collective state at individual scope, while global scope appears in no paper.
• Decision authority is usually silent. 80 of 89 papers do not state who controls the memory operations they implement.
• Reconciliation is the sparsest formation operator. It is studied mainly through concurrency control and analyses of collective-belief failure rather than systems for semantic reconciliation.
• Evaluation remains outcome-heavy. Task success and retrieval accuracy can stay unchanged while scope, propagation, provenance, revocation, and recovery quality regress.

Four constructs introduced by the survey

1. Collective utility separates value to the producer from value to the team:

   $$ U_{\mathrm{team}}(m)=\sum_{j\in A}\pi_j,[u_j(m)-r_j(m)]-c(m). $$

   Here, $u_j$ is the usefulness of item $m$ to agent $j$, $r_j$ is the risk of misapplication, $\pi_j$ weights the relevant audience, and $c(m)$ is the standing cost of maintaining the item.

2. A seven-type conflict taxonomy distinguishes factual, semantic, temporal, contextual, policy, role-dependent, and causal/derived conflicts. Temporal, contextual, and role-dependent conflicts are often representation failures rather than genuine disagreements.

3. A transfer analysis for distributed-systems guarantees shows that ordering, session guarantees, availability-consistency trade-offs, and concurrency control transfer usefully. Convergence does not transfer without an equality relation over natural-language state; commutative merge fails separately because semantic summarization is order-dependent and non-associative.

4. Recovery closure separates whether repair reaches affected state, all holders, and downstream decisions.

📐 Evaluation

The survey proposes a nine-dimensional vector instead of a scalar score:

$$ Q(S_C^M)=[Q_{adm},Q_{scope},Q_{prop},Q_{recon},Q_{prov},Q_{revise},Q_{revoke},Q_{recover},Q_{cost}]. $$

Dimension	Representative measurements
Admission quality	Admission precision, collective-utility recall, harmful-admission rate
Scope quality	Scope accuracy, violation rate, authorized-reach coverage
Propagation quality	Coverage, latency, stale exposure, handoff sufficiency
Reconciliation quality	Resolution accuracy, contradiction persistence, conflict-type errors
Provenance quality	Provenance completeness, attribution fidelity, transformation traceability
Revision quality	Freshness, update lag, superseded-state reuse
Revocation quality	Revocation completeness, propagation lag, zombie-memory rate
Recovery quality	Repair accuracy, state/holder/decision closure, rollback consistency
Cost	Tokens, latency, storage, communication, verification, and repair cost

Three operator-quality quantities highlighted by the survey remain unmeasured in the coded corpus: collective-utility recall, handoff sufficiency, and recovery closure. The corpus likewise reports no direct measure of repair cost.

❔ Research Agenda

The operator × guarantee × measurement gaps motivate five research programs:

1. Dynamic scope and governance - content-bound enforcement, scope-drift detection, risk-aware promotion, and delegable per-operator authority.
2. Semantic consistency and transactions - checkable semantic-consistency predicates, scope-relative guarantees, atomic related writes, and typed conflict resolution.
3. Provenance, invalidation, and recovery - provenance that survives compaction, graded dependencies, complete revocation, tombstones, and recovery-closure guarantees.
4. Minimal coordination and memory-world alignment - behaviorally sufficient handoff, cost-constrained propagation, effect-linked memory, and compensation after repair.
5. Benchmarks for collective-state management - scope transitions, typed conflicts, stale propagation, revocation-repair, and long-horizon organizational scenarios.

🗺️ Survey Roadmap

Section	Focus
1	Introduction and the reliance gap
2	Scope, definitions, conceptual boundaries, and transition operators
3	Review protocol, corpus landscape, and coded evidence
4	Formation: admission, scoping, propagation, and reconciliation
5	Maintenance: revision, revocation, and recovery
6	Cross-cutting guarantees and trade-offs
7	Nine-dimensional evaluation and benchmark gaps
8	System-level synthesis and five research programs
9	Conclusion
Appendices	Search protocol, coding schema, mechanism tables, extended definitions, guarantee analysis, and full-text recoding

📖 Representative Works by Theme

The 42 works below reproduce the grouping and ordering of the representative paper × transition matrix in Table 4. They are not presented as the complete 89-paper corpus. † denotes an analysis, attack, or benchmark rather than a system, following the paper's notation.

Theme	Representative works
Single-agent memory substrates	4
Experience and procedural memory	5
Multi-agent shared and collaborative memory	12
Governance layers and memory operating systems	6
Consistency, concurrency, and collective belief	6
Temporal validity and versioning	3
Provenance, repair, and attack	6
Total	42

Single-Agent Memory Substrates: What Multi-Agent Systems Inherit

Work	Venue / Year	Focus
A-Mem: Agentic Memory for LLM Agents	NeurIPS 2025	Agentic-memory design
Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory	arXiv, 2025	Scalable long-term memory
Memory OS of AI Agent	EMNLP 2025	Memory operating-system design
Text2Mem: A Unified Memory Operation Language for Memory Operating System	Findings of ACL 2026	Unified memory operations

Experience and Procedural Memory

Work	Venue / Year	Focus
Agent Workflow Memory	ICML 2025	Workflow-level memory
Memp: Exploring Agent Procedural Memory	Findings of ACL 2026	Procedural memory
What Deserves Memory: Adaptive Memory Distillation for LLM Agents	ACL 2026	Adaptive memory distillation
Adaptive Memory Admission Control for LLM Agents	ICLR 2026 MemAgents Workshop	Admission control
Memory-R1: Enhancing Large Language Model Agents to Manage and Utilize Memories via Reinforcement Learning	ACL 2026	Reinforcement-learning-based memory management

Multi-Agent Shared and Collaborative Memory

Work	Venue / Year	Focus
Memory Sharing for Large Language Model Based Agents	arXiv, 2024	Memory sharing
Experiential Co-Learning of Software-Developing Agents	ACL 2024	Experiential co-learning
G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems	NeurIPS 2025	Hierarchical multi-agent memory
Collaborative Memory: Multi-User Memory Sharing in LLM Agents with Dynamic Access Control	arXiv, 2025	Dynamically controlled memory sharing
RCR-Router: Efficient Role-Aware Context Routing for Multi-Agent LLM Systems with Structured Memory	arXiv, 2025	Role-aware context routing
MIRIX: Multi-Agent Memory System for LLM-Based Agents	arXiv, 2025	Multi-agent memory system
MemIndex: Agentic Event-Based Distributed Memory Management for Multi-Agent Systems	ACM Transactions on Autonomous and Adaptive Systems, 2025	Event-based distributed memory
LLM-Powered Decentralized Generative Agents with Adaptive Hierarchical Knowledge Graph for Cooperative Planning	AAAI 2025 Workshop on Multi-Agent AI in the Real World	Decentralized knowledge-graph memory
LEGOMem: Modular Procedural Memory for Multi-Agent LLM Systems for Workflow Automation	AAMAS 2026 Extended Abstracts	Modular procedural memory
ESAA-Conversational: An Event-Sourced Memory Layer for Continuity, Handoff, and Curation Across Heterogeneous LLM Coding Agents	arXiv, 2026	Event-sourced continuity and handoff
Multi-Agent Transactive Memory	arXiv, 2026	Transactive memory
LatentMem: Customizing Latent Memory for Multi-Agent Systems	arXiv, 2026	Latent multi-agent memory

Governance Layers and Memory Operating Systems

Work	Venue / Year	Focus
† GateMem: Benchmarking Memory Governance in Multi-Principal Shared-Memory Agents	arXiv, 2026	Memory-governance benchmark
Governed Shared Memory for Multi-Agent LLM Systems	arXiv, 2026	Governed shared memory
† Governed Collaborative Memory as Artificial Selection in LLM-Based Multi-Agent Systems	arXiv, 2026	Governed collaborative memory
MemArchitect: A Policy Driven Memory Governance Layer	arXiv, 2026	Policy-driven memory governance
SuperLocalMemory 4.0: The Governed Memory Operating System for AI Agents	arXiv, 2026	Governed memory operating system
† Governing Evolving Memory in LLM Agents: Risks, Mechanisms, and the Stability and Safety Governed Memory Framework	arXiv, 2026	Stability and safety governance

Consistency, Concurrency, and Collective Belief

Work	Venue / Year	Focus
CoAgent: Concurrency Control for Multi-Agent Systems	arXiv, 2026	Concurrency control
Verified Detection and Prevention of Concurrency Anomalies in Multi-Agent Large Language Model Systems	arXiv, 2026	Verified concurrency-anomaly handling
S-Bus: Automatic Read-Set Reconstruction for Multi-Agent LLM State Coordination	arXiv, 2026	Read-set reconstruction
† When Agents "Misremember" Collectively: Exploring the Mandela Effect in LLM-Based Multi-Agent Systems	ICLR 2026	Collective misremembering
† Emergence of Biased Consensus in Multi-Agent LLM Debates	arXiv, 2026	Biased consensus
† When Truth Is Distributed: Misinformation Derails Collective Fact Recovery in LLM-Based Multi-Agent Systems	arXiv, 2026	Misinformation and collective fact recovery

Temporal Validity and Versioning

Work	Venue / Year	Focus
ChronoMem: Version Control and Semantic Rollback for Large Language Model Agent Memory	arXiv, 2026	Version control and semantic rollback
† STALE: Can LLM Agents Know When Their Memories Are No Longer Valid?	arXiv, 2026	Memory-validity awareness
Beyond Dialogue Time: Temporal Semantic Memory for Personalized LLM Agents	Findings of ACL 2026	Temporal semantic memory

Provenance, Repair, and Attack

Work	Venue / Year	Focus
MAP-Graph: Provenance-Aware Shared Memory for Multi-Agent Workflows	arXiv, 2026	Provenance-aware shared memory
Memory Provenance Laundering in LLM Agents: A Non-Amplification Firewall for Persistent Memory	arXiv, 2026	Provenance laundering and containment
Memory Poisoning Propagation and Repair Mechanism in Multi-Agent Collaborative Environments	Proceedings of the 2nd International Conference on Artificial Intelligence, Digital Media Technology and Social Computing, 2026	Poisoning propagation and repair
Goal-Aware Identification and Rectification of Misinformation in Multi-Agent Systems	ICLR 2026	Misinformation identification and rectification
† Aegis: Automated Error Generation and Attribution for Multi-Agent Systems	ICLR 2026	Automated error attribution
† AgentPoison: Red-Teaming LLM Agents via Poisoning Memory or Knowledge Bases	NeurIPS 2024	Memory-poisoning attack

🎥 Figures

These are high-resolution PNG exports (300 DPI) of the four vector-PDF figures used in the manuscript. Because the source files are named independently of their order in the paper, manuscript Figures 1–4 correspond to Figure1.pdf, Figure3.pdf, Figure2.pdf, and Figure5.pdf, respectively. See the archived manuscript on Zenodo for the figures in context.

Figure 1: Object-Transition-Guarantee-Evaluation Framework

Locally held information becomes memory-mediated collective state through formation operators and is kept valid through maintenance operators. Guarantees constrain every transition, while multidimensional evaluation tests the resulting state.

Figure 2: Collective-State Transition Operators

The seven operators transform different attributes of a collective-state item and can be composed in any order.

Figure 3: Scope and Conceptual Boundaries

Collective state is a subset of persistent agent state, collective memory is a subset of collective state, and shared memory only partially overlaps with it.

Figure 4: Evaluation View

Task success and retrieval accuracy remain necessary, but they are insufficient for measuring formation quality, lifecycle integrity, provenance, and cross-cutting efficiency.

📝 Citation

If you find this survey useful, please consider citing it:

@article{wangprivate,
  author={Yiqi Wang and
          Jiaqi Zhang and
          Zhangkai Wu and
          Mingkai Zheng and
          Zequn Sun and
          Yiqun Duan and
          Zirui Liu and
          Zhihao Hao and
          Taotao Cai},
  title={From Private Experience to Collective State: A Survey of Multi-Agent Memory Formation and Maintenance}
}