Amnesia and déjà vu: how our AI agents failed after every restart (and how we fixed it)

The developers behind the open-source agent runtime ApowerB have shared their technical journey in solving reliability issues for AI agents in production. They identified two primary failure modes: 'amnesia,' where agents lost all context after restarts due to in-memory session storage, and 'déjà vu,' where agents blindly repeated side-effect-heavy actions like sending emails upon relaunch. To fix amnesia, the team migrated session and memory management to a persistent PostgreSQL database. To address déjà vu, they implemented a strict safety protocol that records every tool call before execution. By tracking these calls, the system can now distinguish between safe-to-replay tasks and those that have already interacted with the real world. This approach ensures that agents can recover from crashes without duplicating sensitive actions, providing a robust framework for production-grade AI deployments that prioritize safety and auditability over blind automation.