Mnesion: A Knowledge Accumulation and Consensus Protocol for LLM Agent Organizations

LLM agents increasingly collaborate in multi-agent systems, yet existing protocols address task delegation (A2A) or tool connectivity (MCP), leaving collective intelligence formation unaddressed: hallucination propagation, knowledge evolution, and consensus formation. We present Mnesion, a protocol operating above the messaging layer that provides: (1) a confidence model mitigating hallucination propagation through behavioral conventions at different trust levels; (2) a knowledge management layer built on context efficiency for finite context windows; and (3) a consensus protocol enabling agents to propose, vote on, and enact organizational rules. We describe the four-layer architecture, its relationship to FIPA-ACL, and report on a working implementation serving 5–10 agents. Mnesion complements A2A and MCP as an upper knowledge/consensus layer for persistent agent organizations.