How a 12-Agent AI Fleet Solved Persistent Memory Loss Across Runs

Empire Labs built and runs a fleet of twelve autonomous AI agents in production that handle tasks such as lead generation, compliance monitoring, and code review. A critical incident occurred when an agent lost track of an external submission it had made, with no durable record surviving after its session ended. The team identified three core gaps: lack of durable state storage, no recall path for past sessions, and no enforced discipline for writing state at the moment of action. To fix this, they developed a four-layer memory architecture that separates hot in-prompt memory from persistent per-thread state files stored on disk. The system allows any agent to resume a workflow days later with full context on where things stand and what needs to happen next.