Governed Collaborative Memory as Artificial Selection in LLM-Based Multi-Agent Systems

A new paper from researchers at arXiv (ID: 2605.04264) tackles a critical question for LLM-based multi-agent systems: which candidate memories should become shared institutional state? Authors Diego F. Cuadros, Abdoul-Aziz Maiga, Helen Meskhidze, and Andre Curtis-Trudel frame this as "governed collaborative memory," arguing that memory governance functions as a selection regime—determining which memories persist, remain private, or are rejected/superseded. They distinguish four regimes: ungoverned persistence, constitutional or hybrid selection, automatic metric-based selection, and human-ratified artificial selection. Notably, they emphasize these are not a ranking but design choices over target properties like epistemic quality and role preservation.

The paper describes a layered architecture that separates agent-local memory, shared institutional memory, archive memory, and project-continuity memory—each with provenance and version lineage for inspectability. Using documented traces from one running LLM-based multi-agent ecosystem, they illustrate unmanaged false-memory persistence, ratified institutional memory, rejection and revision, identity-preserving expansion, and governance-as-learning. The contribution is a design agenda: persistent LLM-based multi-agent systems should evaluate memory not only for recall and performance, but also for provenance fidelity, selection traceability, epistemic quality, correction pathways, and role preservation. This work provides a structured approach to a previously unaddressed design problem in multi-agent AI.