Capture Notes

arXiv:2605.22721.

Why Collected

Directly relevant to MAS misevolution propagation because it frames self-evolving MAS as experience-improving agents built on persistent memory and contrasts centralized shared memory with decentralized per-agent dual-pool memory.

Key Metadata

• Submitted: 2026-05-21
• Subject: Multiagent Systems
• DOI: https://doi.org/10.48550/arXiv.2605.22721

Collection Summary

The paper proposes DecentMem, where each agent maintains a dual-pool memory: an exploitation pool of consolidated past trajectories and an exploration pool of LLM-generated candidates for unseen contexts. Pools are reweighted online based on stage-wise feedback from an LLM-as-a-judge. It evaluates across AutoGen, DyLAN, and AgentNet.

Security Relevance

• Provides a concrete self-evolving MAS memory architecture.
• Identifies centralized repository sharing as a design pattern with coordination/privacy/diversity tradeoffs.
• Gives concrete artifact types for threat modeling: past trajectories, generated candidates, stage-wise feedback, LLM-as-a-judge signals, memory pool weighting.
• Useful for comparing centralized memory propagation risk vs decentralized memory isolation and diversity.

Suggested Ingest Focus

• Link to [[04_Research_Questions/RQ - MAS Misevolution Propagation Control]] as a primary architectural source.
• Extract evidence about shared-memory centralization and per-agent memory pools as possible propagation/containment variables.
• Treat performance claims as preprint claims until replicated.