BrainMesh: A Brain-Science-Inspired Memory Topology Architecture for AI Agents

BrainMesh is a memory architecture for AI agents that replaces flat file-based retrieval with a weighted, multi-layer topological network inspired by neuroscience principles. The architecture implements four layers (Event, Relation, Emotion, Intuition) connected by weighted edges, with six core mechanisms: spreading activation for associative recall, Hebbian learning for connection strengthening, memory decay for adaptive forgetting, involuntary recall (flash memory) triggered by pattern signatures rather than content matching, gamma binding for cross-layer memory assembly, and DMN-inspired background consolidation for pattern emergence. We position BrainMesh against recent graph-based agent memory systems (SYNAPSE, Kumiho, GAAMA, MemArchitect) and identify five genuinely novel contributions: involuntary recall, Hebbian edge dynamics, sleep-like consolidation, multi-dimensional pattern signatures, and small-world topology optimization. A working Python prototype demonstrates all core mechanisms. The architecture is designed for integration with existing MSS (Memory-Soul-Skill) frameworks used in AI companion products.