AI agents face a fundamentally different crash-recovery problem than databases. Unlike databases that replay write-ahead logs deterministically, LLM-based agents are non-deterministic — restarting from scratch can produce different decisions from the same inputs, causing 'decision drift' that corrupts all downstream reasoning. The solution is checkpoint-and-resume: periodically saving the agent's full state, accumulated decisions, and reasoning chain. Google Cloud's Agent Platform addresses this natively via three mechanisms: Memory Bank (persistent cross-session state), Resume Agents (checkpoint-and-resume built into the ADK), and Ambient Agents (event-driven execution). The broader insight is that agent memory is fundamentally a consistency problem, not just a retrieval problem, and requires platform-level guarantees similar to database ACID properties.
