Transactive Memory Systems (TMS)

Updated 9 April 2026

Transactive Memory Systems (TMS) are group-level cognitive architectures that distribute and update knowledge using social communication.
TMS integrates private memory stores, shared directories, and coordinated retrieval methods to enhance group performance and adaptive learning.
Applied in human and artificial multi-agent systems, TMS models provide measurable improvements in team coordination and decision accuracy.

A Transactive Memory System (TMS) is a group-level cognitive architecture in which distributed knowledge is codified and dynamically updated via social communication, enabling collectives (human or artificial) to organize, store, and retrieve expertise more effectively than the sum of individual memories alone. Modern TMS research formalizes how individual competencies, memory partitions, and inter-agent coordination strategies integrate to shape overall group performance. TMS models are pervasive across domains, from human teams to large-scale multi-agent artificial systems, with recent work providing precise mathematical and algorithmic characterizations.

1. Formal Structure and Core Components

A TMS consists of private memory stores, a shared directory of “who knows what,” and mechanisms for communication, updating, and recall. Formally, in agent-based settings, a TMS can be expressed as: $\mathrm{TMS} = \left\{ M_1,\dots,M_N; D \right\}$ where each $M_i$ is the private memory (knowledge, skills) of agent $i$ , and $D: K \to 2^{\{1,\dots,N\}}$ is a directory mapping each knowledge item $k$ to the subset of agents believed to possess it (Hu et al., 2023).

In quantitative models of cooperative learning, the state is characterized by:

Expertise levels $y_i(t)\in[0,1]$ for each agent $i$ (where $y_i=1$ represents perfect mastery).
Row-stochastic influence matrix $M(t)=[m_{ij}(t)]$ , with $m_{ij}(t)\geq0$ and $M_i$ 0, denoting agent $M_i$ 1’s belief about $M_i$ 2’s task responsibility.
Stubbornness parameters $M_i$ 3, where $M_i$ 4 (stubborn, never updates beliefs), $M_i$ 5 (non-stubborn, open to new evidence) (Pasquale et al., 2022).

Algorithmically, TMS-based systems—human or artificial—typically exhibit three “phases”:

Encoding (discovery and labeling of expertise),
Storage (maintenance and update of the directory and private knowledge),
Retrieval (locating and combining expertise in response to tasks) (Hu et al., 2023).

2. Mathematical Modeling of Cooperative Learning

Discrete-time and continuous-time models describe the joint evolution of opinion (memory allocation) and skill (expertise) in TMS-enabled teams (Pasquale et al., 2022):

Discrete-time dynamics:

$M_i$ 6

$M_i$ 7

where $M_i$ 8, $M_i$ 9 (learning rates), and $i$ 0.

Continuous-time consensus:

$i$ 1

representing a standard linear consensus process over an interaction graph defined by $i$ 2.

Main Theorems and Implications

Theorem 1: If all agents are non-stubborn ( $i$ 3 for all $i$ 4), expertise converges to the maximal initial proficiency: $i$ 5 where $i$ 6, and the influence matrix converges to uniform mixing, $i$ 7.

Theorem 2: If all agents are stubborn ( $i$ 8), the update freezes at $i$ 9. The group decomposes into “communication classes” according to Frobenius form. Each class $D: K \to 2^{\{1,\dots,N\}}$ 0 achieves an expertise bounded by the maximal initial expertise $D: K \to 2^{\{1,\dots,N\}}$ 1 of all upstream classes $D: K \to 2^{\{1,\dots,N\}}$ 2 with a directed path to $D: K \to 2^{\{1,\dots,N\}}$ 3.

Spectral conditions: Non-stubborn convergence requires $D: K \to 2^{\{1,\dots,N\}}$ 4 to be irreducible (graph strongly connected); the Laplacian must have a simple zero eigenvalue. Stubbornness localizes learning, with matrix reducibility partitioning the flow of knowledge (Pasquale et al., 2022).

3. TMS Implementations in Artificial Multi-Agent Systems

TMS principles have been instantiated in both human-machine and fully artificial agent collectives.

Multi-Agent Memory Architectures

Collective memory banks: Each agent draws from a communal set $D: K \to 2^{\{1,\dots,N\}}$ $D : K \to 2^{{1, \dots, N}}$ 5 of exemplars, with different retrieval policies (fixed, random, similarity-based) approximating distinct “memory directories” (Michelman et al., 7 Mar 2025).
- Varied-context agents each retrieve distinct subsets, operationalizing memory specialization analogous to human TMS.
- Summarizer agents coordinate and synthesize outputs from distributed memory sources, paralleling the coordination role in TMS.
Hierarchical graph architectures: G-Memory structures agentic recall across three explicit graph layers (Zhang et al., 9 Jun 2025):
- Insight Graph: distilled high-level organizational knowledge analogous to the shared expertise store in TMS.
- Query Graph: nodes represent past user queries, storing outcomes and solution trajectories, corresponding to “who knows what.”
- Interaction Graphs: detailed agent-agent dialogue traces, resembling private memory stores in TMS.

Bidirectional memory traversal combines top-down recall of general insights with bottom-up access to prior detailed interaction episodes. Updates propagate experiences through all three memory tiers, ensuring rapid knowledge transfer and continual organizational learning.

Empirical Outcomes

G-Memory enhances performance in multi-agent LLM systems by up to $D: K \to 2^{\{1,\dots,N\}}$ 6 on embodied action and $D: K \to 2^{\{1,\dots,N\}}$ 7 on factual QA tasks compared to memoryless baselines (Zhang et al., 9 Jun 2025).
Random exemplar retrieval improves LLM group reasoning more than similarity-based selection, with distributed recall increasing ensemble accuracy over single-agent or uncoordinated approaches (Michelman et al., 7 Mar 2025).

4. Human-AI Partnerships and TMS

TMS theory has been adapted to analyze and design workflows in human-GenAI partnerships, especially within education (Islam et al., 27 Mar 2026). Key constructs and metrics are operationalized as:

Weighted credibility (CR-S): factor-weighted composite of survey responses quantifying trust and reliance on the AI partner.
Specialization–Coordination composite (SP-COR): measures clarity and smoothness of division of labor and integration between human and AI.

Workflow manipulations (e.g., reflection-first vs verification-required approaches) modulate credibility without significantly altering perceived specialization or coordination. For instance, structuring tasks so students reason independently before consulting AI produces the largest reduction in AI credibility reliance (CR-S adj. mean $D: K \to 2^{\{1,\dots,N\}}$ 8 in “reflection-first” vs $D: K \to 2^{\{1,\dots,N\}}$ 9 in control, $k$ 0 at post) (Islam et al., 27 Mar 2026).

A plausible implication is that instructional sequencing and explicit role scaffolding can be strategically varied to modulate transactive reliance (credibility) without destabilizing the benefits of role clarity or smooth integration.

5. TMS in Socially-Aware Robotics

TMS-inspired frameworks formalize memory and decision-making in socially assistive robots operating among multiple human stakeholders (Hu et al., 2023).

Phases and Architecture

Encoding: Extraction of stakeholder expertise through dialogue and behavioral observation, populating a directory $k$ 1 mapping expertise to agents.
Storage: Updating a knowledge base or graph store using probabilistic confidence values $k$ 2 and provenance logs for each $k$ 3 pair.
Retrieval: Querying the directory and ranking candidates via a confidence-recency tradeoff. Robots explain choices with reference to memory provenance, enhancing transparency and trust.

This TMS deployment is realized in pipeline architecture:

Perception and dialogue → 2. Directory update (Encoding) → 3. Memory maintenance (Storage) → 4. Decision/Retrieval → 5. Explanatory feedback.

Expected outcomes include increased group decision accuracy and higher perceived transparency, as evidenced by prototype trust ratings (mean trust $k$ 4 for TMS-enabled robot vs $k$ 5 baseline, on a $k$ 6– $k$ 7 scale, in a hypothetical experiment) (Hu et al., 2023).

6. Design Principles, Limitations, and Future Directions

Design Guidelines:

Increase “open-mindedness” (lower stubbornness) in teams and MAS by incentivizing opinion update.
Ensure strong connectivity in the collaboration graph—rotate pairings, insert “knowledge brokers” (agents with high $k$ 8 and many ties).
For artificial TMS, leverage random or context-optimized memory retrieval and compositional summarization.

Limitations:

Full observability and uniform persuasiveness are simplifying assumptions in most formal TMS models.
In multi-agent LLM systems, quality of memory retrieval and update is limited by LLM accuracy, risking the persistence of hallucinations or misinformation (Zhang et al., 9 Jun 2025).
Classical TMS frameworks focus mostly on single-task, static networks; extensions include handling multiple tasks, dynamic connections, and noisy or partial observations (Pasquale et al., 2022).

Research Directions:

Weighted or authority-modulated influence models; graph-neural architectures for memory.
Robustness to adversarial or contradictory agent memory entries.
Direct measurement of TMS health (e.g., expertise diffusion speed, coordination robustness) in large and dynamic agent collectives.

Transactive Memory System models unify and operationalize a set of principles and mechanisms by which distributed groups—of humans, robots, or artificial agents—organize, maintain, and dynamically evolve collective knowledge. Rigorous mathematical, algorithmic, and experimental work continues to elaborate the foundations and extend applicability to ever larger and more complex collectives (Pasquale et al., 2022, Michelman et al., 7 Mar 2025, Hu et al., 2023, Islam et al., 27 Mar 2026, Zhang et al., 9 Jun 2025).