An empirical study of memory collapse and belief contamination in multi-agent LLM systems. Compares structured incremental memory updates against baseline architectures under varied network topologies. Demonstrates how memory curation and topology govern the spread of non-transferable lessons and long-term agent efficacy.
reinforcement-learning multi-agent-systems ai-safety network-topologies ollama llm-agents agentic-memory self-evolving-agents memory-collapse belief-contamination
-
Updated
Sep 6, 2026 - Python