Researchers propose Multi-Agent Transactive Memory (MATM), a framework enabling decentralized LLM agents to share and retrieve trajectories—recorded problem-solving paths—from a shared repository. Experiments in interactive environments demonstrate that agents retrieving stored trajectories improve task performance and efficiency without requiring coordination or joint training.
Multi-Agent Transactive Memory addresses a critical inefficiency in distributed AI systems: agents repeatedly solving identical problems despite solutions already existing within their population. The framework draws inspiration from human knowledge infrastructure like search engines, extending retrieval-augmented generation from individual agent use cases to population-level knowledge sharing. This represents a meaningful shift toward treating agent-generated artifacts as reusable infrastructure rather than disposable outputs.
The research builds on growing recognition that LLM agents operate within social structures requiring coordination mechanisms. Prior work demonstrated retrieval-augmented generation's value for individual agents; MATM scales this principle to heterogeneous agent populations with diverse capabilities. Testing on ALFWorld and WebArena—interactive environments where trajectories encode substantial procedural complexity—shows performance gains and reduced computational overhead without explicit coordination between producer and consumer agents.
For the AI infrastructure sector, MATM establishes a design pattern for open agent ecosystems that could drive adoption and efficiency improvements. As deployed agents increase, knowledge accumulation becomes increasingly valuable, potentially creating network effects favoring platforms implementing trajectory retrieval systems. This framework also implies new opportunities for specialized services: trajectory curation, quality assessment, and indexed knowledge management tailored to agent populations.
The significance lies not in breakthrough capability but in architectural thinking. As AI systems transition from isolated applications to distributed populations, infrastructure for knowledge sharing becomes foundational. MATM's demonstration that passive retrieval—without coordination overhead—improves outcomes suggests scalable approaches to agent ecosystem maturation, positioning it as relevant infrastructure for production deployments.
- →MATM enables agent populations to share problem-solving trajectories, reducing redundant computation and improving task efficiency without coordination overhead.
- →Agent-generated artifacts contain reusable procedural knowledge that transactive memory systems can systematically preserve and leverage across populations.
- →Performance improvements in interactive environments suggest trajectory retrieval creates network effects as agent populations grow larger.
- →The framework establishes a design pattern for open agent ecosystems, with potential commercial applications in trajectory management and curation services.
- →Infrastructure for population-level knowledge sharing represents an emerging category of AI systems architecture distinct from individual agent optimization.