Multi-Agent Transactive Memory

• Multi-Agent Transactive Memory (MATM) is a framework where diverse agents collaboratively store, index, and retrieve knowledge, enabling robust and scalable cognitive systems.
• It incorporates distributed memory storage, expertise indexing via embeddings, and coordinated retrieval protocols to achieve efficient and dynamic task execution.
• MATM supports long-term knowledge consolidation and has demonstrated significant performance gains in LLMs, reinforcement learning teams, and multimodal systems.

Multi-Agent Transactive Memory (MATM) is an organizational and computational framework wherein a population of agents collectively stores, indexes, retrieves, and updates knowledge through distributed mechanisms, enabling synergistic reasoning, robust task performance, and efficient knowledge transfer. MATM extends classical transactive memory concepts—originally developed in cognitive and social sciences—to multi-agent artificial intelligence, including LLMs, reinforcement-learning teams, multimodal systems, and open agent populations. Central to MATM is the codification of "who knows what," dynamic coordination protocols for memory interaction, and memory architectures that scale to heterogeneity, open-ended environments, and long-term knowledge consolidation.

1. Formal Principles and Definitions

MATM systems instantiate a shared cognitive architecture over agent collectives. Classical transactive memory (TMS) is formalized as a tuple $(\{M_i\}_{i=1}^{n}, M_{\text{shared}}, I)$, where $M_i$ is private memory for agent $i$, $M_{\text{shared}}$ is shared team memory, and $I$ is an indexing or directory mechanism encoding expertise allocation ("who knows what"). In modern MATM, these elements are realized via:

• Distributed Knowledge Storage: Each agent and the system collectively hold persistent knowledge elements, ranging from private state vectors, fact repositories, episodic buffers, to external key-value stores (Zhou et al., 2019, Wang et al., 10 Jul 2025, Huang et al., 28 Jan 2026).
• Expertise Indexing: Memory slots, exemplars, or trajectories are annotated (often via embeddings or metadata) for efficient retrieval keyed to agent competence, subdomain, or context (Kim et al., 18 Jun 2026, Michelman et al., 7 Mar 2025).
• Coordinated Retrieval and Write: Agents exploit attention, cross-referencing, or consensus/summarization routines to access, integrate, and update relevant knowledge (Wang et al., 2024, Michelman et al., 7 Mar 2025).

Formally, in MATM for population-level artifact sharing, given sets of producers $\mathcal{P}$ and consumers $\mathcal{C}$ acting over a repository $\mathcal{R}$, two central operations are defined:

• Contribution: $$(p, T) \mapsto \mathcal{R} \leftarrow \mathcal{R} \cup \{T\}$$ (producer $p$ adds trajectory $M_i$0)
• Retrieval: $M_i$1 (consumer $M_i$2 retrieves chunk $M_i$3 given query $M_i$4 summarizing state/task) (Kim et al., 18 Jun 2026). Indexing and retrieval are frequently realized via vector similarity, dense indexing, or learning-to-rank rerankers built over agent and trajectory features.

2. Memory Architectures and Roles

State-of-the-art MATM systems manifest a range of architectural decompositions, driven by domain, modality, and coordination demands.

• Hierarchical and Modular Memory: Memory is explicitly partitioned by semantic granularity (raw event, fact, summary), by memory type (episodic, semantic, procedural), or by trajectory segment (Huang et al., 28 Jan 2026, Wang et al., 10 Jul 2025, Kim et al., 18 Jun 2026). For example, MIRIX segregates a MetaMemoryManager and six specialized sub-agents responsible for orthogonal knowledge spaces (Core, Episodic, Semantic, Procedural, Resource, Vault) (Wang et al., 10 Jul 2025).
• Role-Specialized Agents: Execution is distributed across agents responsible for construction (extracting facts/events), retrieval (query routing/adaptive selection), judgment (consistency, relevance verification), and refreshing/updating (conflict resolution, targeted forgetting) (Huang et al., 28 Jan 2026).
• Cooperative Reasoning: LLM systems deploy teams of varied-context agents (distinct in-context exemplars per agent) whose outputs are aggregated by a summarizer agent, operationalizing transactive decision-making and synthesis (Michelman et al., 7 Mar 2025).

Table 1 summarizes canonical MATM agent roles:

Role	Function in MATM	Example System
Constructor	Decompose input, insert new memory	AMA (Huang et al., 28 Jan 2026)
Retriever	Route/select/retrieve memory at query	AMA, MIRIX
Judge/Evaluator	Assess relevance, detect inconsistency	AMA
Refresher/Updater	Modify/delete for consistency, capacity	AMA
Summarizer	Aggregate, coordinate team output	LLM MATM (Michelman et al., 7 Mar 2025)
MetaManager	Orchestrate/route events and queries	MIRIX

3. Retrieval, Indexing, and Coordination Mechanisms

• Indexing: Memory elements are mapped to vector embeddings (content-plus-metadata), enabling efficient, content-sensitive retrieval via nearest neighbor search, FAISS indices, or cosine similarity (Wang et al., 10 Jul 2025, Huang et al., 28 Jan 2026, Kim et al., 18 Jun 2026). Expert domain allocation is thereby encoded as an attribute in the memory's "directory" (Kim et al., 18 Jun 2026).
• Retrieval Protocols:
  • Fixed: Pre-select constant exemplars for all queries (Michelman et al., 7 Mar 2025).
  • Random: Uniformly sample, maximizing context diversity (Michelman et al., 7 Mar 2025).
  • Similarity-based: Top-k closest matches given embedding similarity (e.g., cosine) (Michelman et al., 7 Mar 2025, Kim et al., 18 Jun 2026).
  • Learning-to-rank reranking: Feature-rich models (e.g., FFN, LambdaMART) prioritize memory chunks via expected downstream utility, incorporating metadata, task type, and agent ID (Kim et al., 18 Jun 2026).
• Coordination: Queries and memory updates flow via explicit message-passing (AMA), meta-manager orchestration (MIRIX), or consensus/summarizer voting (LLM MATM) (Michelman et al., 7 Mar 2025, Huang et al., 28 Jan 2026, Wang et al., 10 Jul 2025).

4. Empirical Results and System Comparisons

MATM methods yield robust gains in real-world and synthetic domains compared to single-agent or monolithic baselines.

• LLM Reasoning: Varied-context agents with distributed exemplars and a summarizer outperform identical-agent or static-shot baselines. Random retrieval surpasses similarity-based selection, suggesting that maximizing exemplar diversity yields richer reasoning (Michelman et al., 7 Mar 2025).
• Long-Term Conversational/Multimodal Memory: MIRIX achieves a $M_i$5 accuracy increase and a $M_i$6 reduction in storage on ScreenshotVQA vs. RAG-style baselines; on LOCOMO, an $M_i$7 accuracy improvement over best existing agentic-memory (Wang et al., 10 Jul 2025).
• LLM Collaboration (AMA): Shows $M_i$8 reduction in token usage and higher accuracy versus full-context baselines, via multi-granularity retrieval and coordination (Huang et al., 28 Jan 2026).
• Population-Level Experience Sharing: In ALFWorld, MATM with reranking increases success rate from $M_i$9 (no retrieval) to $i$0, and reduces required interaction steps. Cross-task retrieval confirms transferability across unseen agents and novel goals (Kim et al., 18 Jun 2026).
• Multi-Agent RL: Attention-based shared memory architectures achieve faster or more stable learning in partially observable or communication-constrained environments, outperforming memoryless or uncoordinated teams (Zhou et al., 2019).

5. Theoretical Underpinnings and Analyses

Transactive memory formalism extends to dynamic expertise models, as in cooperative learning where agents' competence levels ($i$1) and appraisal networks ($i$2) co-evolve. When agents are non-stubborn (openness parameter $i$3), all asymptotically acquire competence equal to the initially best member, and appraisals converge to uniform trust. Conversely, stubborn or disconnected agents stymie knowledge diffusion, isolating expertise (Pasquale et al., 2022). In reinforcement learning MATM (Zhou et al., 2019), memory mechanisms are cast in composable neural units: per-agent private state, differentiable shared memory, attention-based read/write, and embedding-based slot indexing, establishing a direct computational parallel to TMS theory.

6. Design Lessons, Limitations, and Deployment Considerations

MATM effectiveness hinges upon the careful engineering of memory abstraction, coordination protocols, and agent specialization:

• Diversity Maximization: Allocating exemplars and memory regions across agents or time steps amplifies model performance relative to temperature-based or naive sampling (Michelman et al., 7 Mar 2025).
• Memory Granularity and Adaptivity: Adaptive partitioning (by abstraction/type) better matches reasoning demands (AMA), as opposed to rigid chunking (Huang et al., 28 Jan 2026).
• Specialization and Meta-Coordination: Maintaining explicit role separation (creator, retriever, judge, refresher) outperforms monolithic agentic memories that entangle all phases (Huang et al., 28 Jan 2026).
• Population Scalability: MATM supports population-level utility, with consumer-producer marketplaces, context-conditioned retrieval, and cross-agent trust modeling providing natural avenues for generalization (Kim et al., 18 Jun 2026).
• Limitations: Overly large memory banks may yield spurious reads; state-based bookkeeping can defeat interpretability if agents "cheat" by storing shared state in environment variables (Zhou et al., 2019). Transactive memory convergence demands connectivity in interaction graphs and openness in appraisal dynamics (Pasquale et al., 2022). In LLM-based settings, summarizer agents improve weak base models but may be redundant for high-capacity models (Michelman et al., 7 Mar 2025).

7. Future Directions

Key avenues in MATM research include:

• Incorporating dynamic producer-consumer trust and incentivization for fair artifact contribution (Kim et al., 18 Jun 2026).
• Refining reranking/selection models for richer context-specific retrieval (Kim et al., 18 Jun 2026, Wang et al., 10 Jul 2025).
• Developing hybrid consolidation schemes (cross-agent attention, meta-learning) for long-run knowledge integration (Wang et al., 10 Jul 2025).
• Scaling to open-ended, cross-modality settings including vision, speech, and action planning (Wang et al., 10 Jul 2025, Wang et al., 2024).
• Formalizing guarantees for bounded forgetting, avoidance of spurious coordination, and generalization across unseen domains (Pasquale et al., 2022, Zhou et al., 2019).

MATM encapsulates the convergence of distributed cognitive science, modern deep learning, and multi-agent systems engineering, operationalizing scalable, robust collective memory for artificial agents in diverse environments.