MemoGuard adaptive runtime slashes robot safety violations by 76.6%

MemoGuard, presented at IEEE/ACM ESWEEK 2026, tackles a critical gap in autonomous robot navigation: the reliance on episodic memory reuse in communication-denied environments. Similarity-based retrieval can return actions that match the current context but are unsafe due to topology changes, battery limits, or unreliable outcomes—dubbed 'memory traps.' The lightweight runtime introduces three validation contracts (topology, resource, outcome) that check each retrieved episode before execution. If validation fails, MemoGuard falls back to local reasoning using an onboard LLM (Llama 3.2). In a graph-based corridor-inspection simulator, this approach achieves 76.6% fewer battery safety violations than similarity-only top-1 reuse.

On real hardware (NVIDIA Jetson AGX Xavier), each avoided fallback saves 3.67 seconds and 36.97 joules, meaning MemoGuard not only improves safety but also cuts energy consumption versus always-reasoning strategies. The 21.4% reduction in fallback calls directly translates to longer mission duration and faster decision cycles. Open-sourced on GitHub, MemoGuard is designed for cost-sensitive, low-power robots operating in disaster inspection and search-and-rescue scenarios where remote connectivity is unavailable. Its contracts can be adapted to other constraints like time or sensor health, making it a versatile safety layer for autonomous systems.