Multi-Agent Transactive Memory

arXiv:2606.19911 · cs.AI, cs.CL, cs.IR · Submitted 2026-06-18 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Multi-Agent Transactive Memory".

Jane: The gist The decentralized deployment of LLM agents with diverse capabilities across diverse tasks motivates infrastructure for knowledge sharing across heterogeneous agent populations.

Tom: First, who's behind it and why it matters.

Paper summary: Tom: Alright team, so we're looking at this paper called Multi-Agent Transactive Memory. The main idea here is setting up infrastructure for knowledge sharing across different types of AI agents.

Jane: So, what is the core problem they are trying to solve?

Tom: They point out that just like search engines index stuff for humans, retrieval systems can organize things agents generate so other agents can reuse them and extend retrieval-augmented generation beyond just human documents.

Lu: Specifically, agent trajectories are this thing that encodes reusable procedural knowledge, but right now those artifacts are usually tossed after one use or only kept by the agent that made them.

Tom: So, it forces new agents to keep rediscovering the same solutions over and over again because they don't have access to what others have already figured out.

Meng: That sounds like a major headache for deployment at scale. If every agent has to re-solve problems from scratch, the computational cost just explodes.

Jane: But this paper proposes Multi-Agent Transactive Memory, which is a framework for storing and retrieving those agent trajectories at the population level.

Lalam: So producer agents contribute their trajectories to a shared repository and consumer agents go out and retrieve them to help improve their task execution.

Tom: That's the gist of it. It’s moving from individual memory to a collective experience bank for the entire agent ecosystem.

Lu: The paper frames this as an exchange where producer agents give their artifacts, and consumer agents take them back, which creates a kind of two-sided marketplace for procedural knowledge with clear attribution between what's retrieved and who produced it.

Meng: I like that structure because it implies some kind of incentive or at least a clear way to track where the knowledge came from, which is important for debugging if something goes wrong at scale.

Jane: It sounds like they are trying to fix the problem where knowledge gets siloed with the person who first found it.

Tom: Exactly. And they claim that in interactive environments, specifically ALFWorld and WebArena, retrieving these trajectories actually improves task performance without needing any extra coordination or joint training between agents.

Jane: That's a big claim because usually when you introduce something new, you need to retrain the whole system or coordinate complex training routines.

Paper summary: Tom: The paper shows that even without that joint training, just adding this retrieval step helps things out in terms of task success rates.

Lu: They are using specific indexing and retrieval mechanisms for these trajectories. It's a state-conditioned key-value indexing scheme where the recent interaction history serves as the retrieval key and the next segment is stored as the value.

Meng: That sounds computationally heavy, though I want to see how efficient that actually is when you're dealing with long, procedural interactions.

Tom: They talk about defining a window size l, so the key is a function of x and the sequence tau t-l+one through tau t <ref:2606.19911#pg1>.

Jane: And then retrieval isn't just one simple search; they introduce a trajectory retriever, R, which can be either a dense retriever or a pipeline combining an initial retriever with a reranker.

Tom: And to make that better, they add this learning-to-rank stage for trajectory reranking.

Lu: That LTR stage uses a feature map phi to capture different aspects of usefulness, like metadata from the producer agent for trust modeling and metadata from the consumer agent for personalization.

Meng: So it's not just about finding what is similar; it's about ranking what is *useful* for this specific consumer in this specific context.

Tom: They train a scoring function g parameterized by theta to predict the helpfulness of a document, and they use supervision over marginal utility.

Jane: The training data is constructed by looking at retrieved chunks d(j)t at branching points t, recording the tuple (q t, d(j)t, s(j)t - s base), where s is the label capturing the chunk’s marginal utility relative to a no-retrieval baseline.

Lu: They test three different reranker families for training: a feed-forward network FFN, a pairwise LambdaMART method, and pairwise SVMRank.

Meng: The paper mentions that they're evaluating this on thirty-five to thirty-seven producer agents and thirty-four consumer agents in ALFWorld and WebArena.

Tom: And the results show that retrieval consistently improves task outcomes across both benchmarks. For ALFWorld, they saw a success rate increase from forty-seven percent to fifty-five percent, while WebArena went from eighteen percent to twenty percent.

Jane: That's a noticeable jump in performance for these agent systems.

Lu: The LTR stage really boosts things; for ALFWorld specifically, the SVMRank method achieved the strongest results across all metrics, reaching a success rate of sixty-four point three percent and an RPP of zero point one five.

Paper summary: Meng: That's interesting because it suggests that the quality of ranking matters a lot in this system.

Jane: And they found something quite broad about the benefit: retrieval helps agents regardless of whether the producer agent is weaker or stronger than the consumer. They state that MATM’s value cannot be reduced to a single strong producer.

Tom: That's important because it means you don't need to have the best agent in your system to get a return on this infrastructure.

Lu: Furthermore, they found that the retrieval benefit extends across different tasks, even when agents are looking for trajectories from structurally different task types; ALFWorld SR reaches fifty-nine point nine percent.

Meng: So it’s not just about reusing solutions for the exact same problem an agent faced before.

Tom: And they noted that the correlation between a capability gap and the retrieval advantage stays small and insignificant across both benchmarks.

Jane: That suggests the knowledge in MATM is more generally useful than just specialized knowledge held by a single elite agent.

Lu: In terms of scalability, ALFWorld shows monotonic improvement in both success rate and step efficiency as the index size grows. But WebArena's success rate has a non-monotonic pattern; it dips at fifty percent index before recovering sharply to twenty point nine percent at full scale.

Meng: That dip makes me wonder what happens when you try to scale things up too much on that benchmark.

Jane: So, the paper concludes that Multi-Agent Transactive Memory provides a scalable mechanism for population-level experience reuse, letting agents leverage collective trajectories instead of just rediscovering solutions in isolation.

Tom: It supports the idea that transactive memory is a viable architecture for agent memory.

Meng: The authors also pointed out some limitations. They admit that the scope is restricted to these two interactive benchmarks and a consumer population of thirty-four models, which limits how much we can generalize their claims right now.

Jane: And they also noted that they don't account for adversarial producers who might contribute malicious trajectories, which could put consumers at risk.

Lu: They also mentioned that the scaling experiments suggest that incentives or remuneration for producer contributions will be an important challenge as these platforms grow.

Tom: So, it’s not just a technical solution; there’s a whole economic layer to consider if this is going to actually get deployed widely.

Conclusion: Tom: So we've been talking about Multi-Agent Transactive Memory, or MATM. Basically, it’s this new way for AI agents to share knowledge without having to constantly retrain everyone together.

Jane: Right, so the core idea is that producer agents can drop their successful solutions into a shared memory bank and consumer agents can pull those solutions out when they need them.

Tom: Exactly. And what makes it interesting is that these trajectories are procedural knowledge, like how an agent navigated a specific task or environment.

Lu: What's key is the structure they set up—producer agents contribute, consumer agents retrieve—which turns the knowledge into a kind of marketplace.

Meng: I'm trying to picture this in production. It sounds like we’re building a decentralized system where different AI models are trading experience artifacts with each other at scale.

Jane: That’s right, and it shifts the focus from just having one super-smart agent to building an entire population that can leverage collective experience.

Tom: The authors of this paper are showing that this system works well in real interactive environments like ALFWorld and WebArena.

Lu: They're proving that you don't need joint training or complex coordination for the agents to benefit from this retrieval step.

Jane: So what does this mean for us listening? It suggests a path toward building more resilient and efficient agent systems by making knowledge sharing a fundamental part of their architecture.

Tom: It really boils down to moving away from individual silos of knowledge and towards a shared, accessible experience pool for the entire agent ecosystem.

Jane: And that’s what we're going to dig into next—we'll look at exactly how they ranked those retrieved trajectories using learning-to-rank techniques.

Carnegie Mellon University · University of California, Berkeley

cs.AI, cs.CL, cs.IR

Submitted: 2026-06-18

Updated: 2026-10-15

Code: https://github.com/kimdanny/matm

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 90/100

The gist: The gist The decentralized deployment of LLM agents with diverse capabilities across diverse tasks motivates infrastructure for knowledge sharing across heterogeneous agent populations.

Key concepts

Agent Trajectories
These are sequences of actions and observations recorded by an LLM agent during a task. They encode reusable procedural knowledge—how an agent successfully solved a problem—which is often discarded after one use, making it valuable for other agents.
Multi-Agent Transactive Memory (MATM)
MATM is a framework where producer agents contribute their trajectories to a shared repository, and consumer agents retrieve them when performing new tasks. This creates a 'twosided marketplace' for knowledge where producers share and consumers benefit from collective experience.
Learning-to-Rank Trajectories (LTRT)
This is a stage in the retrieval pipeline that uses machine learning to rank retrieved trajectories based on their predicted helpfulness. It considers features like producer and consumer agent metadata to personalize results, ensuring the most useful knowledge is prioritized for each request.

Terminology

Summary

The gist The decentralized deployment of LLM agents with diverse capabilities across diverse tasks motivates infrastructure for knowledge sharing across heterogeneous agent populations.

Multi-Agent Transactive Memory (MATM)

MATM is a framework designed for population-level storage and retrieval of agent-generated trajectories, where producer agents contribute artifacts to a shared repository and consumer agents retrieve them to improve task execution. This system addresses the problem that agent trajectories, which encode reusable procedural knowledge, are typically discarded after single use or retained only by the producing agent, forcing newly instantiated agents to repeatedly rediscover existing solutions The decentralized deployment of LLM agents with diverse capabilities across diverse tasks motivates infrastructure for knowledge sharing across heterogeneous agent populations.

The framework operates on a producer-consumer structure where roles are not mutually exclusive, allowing an agent to both produce trajectories in one context and consume them in another This structure induces a twosided marketplace for agent-generated procedural knowledge, with clear attribution between retrieved artifacts and their sources (Page 2). As more agents interact with the repository, MATM grows organically, accumulating a corpus across an increasingly diverse set of tasks and environments

Indexing and Retrieval Mechanisms

The system employs a state-conditioned key-value indexing scheme for action-observation trajectories, where recent interaction history serves as the retrieval key and the subsequent interaction segment as the stored value Specifically, given a window size l, the key is defined as e(t) = f (x, τt−l+1,..., τt) and the associated value is (τt,..., τt+l−1), which serves as the document d that an agent retrieves at inference time

Retrieval is performed by a trajectory retriever R, which may be instantiated as a dense retriever or a cascaded retrieval pipeline combining an initial retriever with a reranker To further improve retrieval quality, the paper introduces an efficient yet powerful learning-to-rank (LTR)-based trajectory reranking stage This stage uses a feature map ϕ to capture multiple complementary aspects of trajectory usefulness, including producer agent metadata for trust modeling and consumer agent metadata for personalization

Learning to Rank Trajectories (LTRT)

The LTRT pipeline consists of a retrieval stage followed by a feature-based ranking stage where a parameterized scoring function gθ: Rz 7→ R indicates the predicted helpfulness of a document d value for task q Features are categorized into six types, including producer agent metadata and consumer agent metadata, which enable trust modeling and personalization Training requires supervision over marginal utility, where a chunk is labeled as helpful to the extent that injecting it improves task outcome relative to running the same agent with no retrieval

The LTRT dataset S is constructed by evaluating retrieved chunks d(j)t at branching points t and recording the tuple (qt, d(j)t, s(j)t − sbase), where the label l = s(j)t − sbase captures the chunk’s marginal utility relative to the no-retrieval baseline The training involves three reranker families, including a feed-forward network (FFN), pairwise LambdaMART, and pairwise SVMRank

Experimental Results and Scope

Experiments are conducted in two interactive benchmarks: ALFWorld and WebArena, utilizing 35 to 37 producer agents and 34 consumer agents respectively The results demonstrate that retrieving trajectories from MATM improves downstream task performance without requiring additional coordination or joint training

Key findings include:

  1. Retrieval consistently improves task outcomes across both benchmarks, with success rate increases observed on ALFWorld (from 47% to 55%) and WebArena (from 18% to 20%)

  2. The Learning-to-Rank stage further boosts welfare; for ALFWorld, SVMRank achieves the strongest results across all metrics, reaching a success rate of 64.3% and an RPP of 0.15

  3. Retrieval benefit is distributed across the agent population regardless of whether the producer is weaker or stronger than the consumer, showing that MATM’s value cannot be reduced to a single strong producer (Page 7)

  4. MATM retrieval benefit extends to both weaker and stronger agents, generalizes across tasks, and continues to improve as the repository grows

  5. Retrieval benefit is not primarily driven by producer-to-consumer competence transfer, as the correlation between capability gap and retrieval advantage remains small and insignificant across benchmarks

Scalability and Generalization

The system exhibits scalability with memory size, where ALFWorld shows monotonic improvement in both success rate and step efficiency with index size WebArena's success rate exhibits a non-monotonic pattern, dipping at the 50% index before recovering sharply to 20.9% at full scale

Furthermore, MATM retrieval benefit is not exclusively confined to same-task trajectories; even under cross-task retrieval, ALFWorld SR reaches 59.9%, showing that trajectories from structurally different task types still carry transferable utility (Page 8)

The paper concludes that MATM provides a scalable mechanism for population-level experience reuse, enabling agents to leverage collective trajectories rather than repeatedly rediscovering solutions in isolation The results support the hypothesis that transactive memory provides a viable architecture for population-level agent memory

Limitations and Future Directions

Limitations include the scope being restricted to two interactive benchmarks and a 34-model consumer population, which limits the scope of empirical claims Future work could draw on attribution fairness in RAG or marketplace evaluation frameworks to address producer-side welfare Additionally, the current framework does not account for adversarial producers who may contribute malicious trajectories, potentially placing consumer agents at risk The scaling experiments suggest that incentives or remuneration for producer contributions will be an important challenge for such platforms The paper is supported by NSF grant 2402874

--- Page 1 ---

To Eun Kim and Fernando Diaz. 2025 Towards fair rag: On the impact of fair ranking in retrieval-augmented generation. In Proceedings of the 2025 International ACM SIGIR Conference on Innovative Concepts and Theories in Information Retrieval (ICTIR), ICTIR’25, page 33–43 Association for Computing Machinery.

--- Page 1 ---

To Eun Kim, Alireza Salemi, Andrew Drozdov, Fernando Diaz, and Hamed Zamani. 2024 Retrievalenhanced machine learning: Synthesis and opportunities. arXiv preprint arXiv:2407.12982.

--- Page 1 ---

To Eun Kim, Alireza Salemi, Hamed Zamani, and Fernando Diaz. 2026 Evaluation of agents under simulated ai marketplace dynamics. arXiv preprint arXiv:2604.14256.

--- Page 1 ---

Index Retrieval Model documents images videos human LLM agent TRADITIONAL RETRIEVAL RETRIEVAL-AUGMENTED GENERATION MULTI-AGENT TRANSACTIVE MEMORY INDEX RETRIEVAL MODEL documents images videos INDEX RETRIEVAL MODEL trajectories Figure 1: Multi-Agent Transactive Memory (MATM).

--- Page 2 ---

Artifact sharing and reuse are essential for enabling scalable, efficient, and continually improving agent populations. As agents operate across environments, they produce a number of intermediate artifacts, which contain rich procedural knowledge, such as action-observation trajectories (Muennighoff et al., 2025). Yet these artifacts are typically discarded after a single use or retained only by the producing agent (Zheng et al., 2024).

--- Page 3 ---

In contrast with RAG, agentgenerated artifacts can be more suitable for agent consumption compared to human-authored documents (Chen et al., 2026). The need for populationlevel reuse is further amplified by practical considerations. Many modern agents rely on inferencetime scaling and generate a number of intermediate artifacts, incurring substantial computational cost (Kaplan et al., 2020; Yao et al., 2023; Wu et al., 2024; Welleck et al., 2024).

--- Page 5 ---

We first show that agents consistently benefit from a simple single-stage retrieval pipeline: retrieving relevant trajectories from a MATM repository populated by diverse agents not only improves downstream task performance without requiring additional coordination or joint training, but also improves task efficiency as measured by a reduced number of interaction steps

--- Page 6 ---

Table 1 Summary Results for ALFWorld and WebArena under no-retrieval and single-stage retrieval from MATM. Across both benchmarks, retrieval from the shared repository consistently improves task outcomes.

--- Page 6 ---

Table 2 MATM retrieval scope ablation results across three candidate pool restrictions. Number of tasks in parentheses.

--- Page 6 ---

--- Page 7 ---

Improvements for AI systems

  1. Bold header: Population-level knowledge sharing infrastructure

This system can store agent-generated artifacts such as interaction trajectories in a shared repository to enable population-level storage and retrieval of agent-generated trajectories. This allows consumers to retrieve them to improve task execution, facilitating continual learning and cross-agent transfer, reducing redundant exploration.

  1. Bold header: State-conditioned trajectory retrieval

The system can employ a state-conditioned key-value indexing scheme where recent interaction history serves as the retrieval key and the subsequent interaction segment as the stored value, allowing agents to retrieve continued guidance conditioned on their current state rather than only on the original task instruction.

  1. Bold header: Learning to Rank trajectory reranking

The system can use a learning-to-rank (LTR)-based trajectory reranking stage with features like producer agent metadata and consumer agent metadata to select the most relevant artifacts, which is shown to improve performance, as evidenced by SVMRank achieving a success rate of 64.3% on ALFWorld.

Sources

Related papers