BrainBox: Hebbian Memory Architecture for AI Coding Agents

We present BrainBox, a novel memory architecture for AI coding agents that learns behavioral patterns using Hebbian learning and synaptic myelination. Unlike declarative memory systems (Mem0, SuperMemory, OpenMemory, Zep, Letta, LangMem), BrainBox implements procedural memory — learning how agents work rather than what they know. The system records file co-access patterns, error-fix associations, and tool usage sequences, strengthening neural pathways through repeated use. We survey seven prominent agent memory systems and demonstrate that all operate at Layer 2 (declarative knowledge). BrainBox operates at Layer 3 (agent behavioral learning) — a layer where no prior published work exists. Production evaluation shows 8.9% gross token savings in 5 hours, with benchmark evaluation achieving 67% top-1 recall accuracy (5x over baseline) on a 2,276-neuron production network.