Ilmu Komputer & AI editorial
Open AccessOA2026
MemoGuard: An Adaptive Runtime for Guarding Against Memory Traps in Communication-Limited Robot Navigation
A lightweight validation layer that prevents unsafe episodic memory reuse in communication-limited robots, reducing battery safety violations by 76.6% while cutting fallback reasoning overhead by 21.4%.
Rajat Bhattacharjya; Hyeonjong Ju; Sing-Yao Wu; Eli Bozorgzadeh; Nikil Dutt· 2026· DOI 10.48550/arXiv.2607.15589
The core problem
Communication-limited robots operating in mission-critical scenarios—such as disaster inspection and search-and-rescue—must make reliable onboard decisions without access to remote operators or high-capacity reasoning services. Episodic memory reuse offers a low-cost fallback: retrieve a past episode similar to the current context and reuse its action. However, retrieval similarity does not guarantee execution validity. A retrieved action may match the current context yet be unsafe due to changed topology, insufficient battery margin, or unreliable prior outcomes. The authors term such high-similarity but execution-invalid episodes **memory traps**. This creates a safety-efficiency design space: similarity-only reuse minimizes fallback cost but can be unsafe, while always invoking local reasoning improves safety at high computational and energy cost. The paper presents **MemoGuard**, a lightweight adaptive runtime that validates episodic memories against topology, resource, and outcome contracts before reuse, invoking fallback only when validation fails. The core problem is formalized as a trade-off between safety and efficiency, with the goal of achieving safe reuse without excess
Innovation
In the graph-based corridor-inspection simulator, MemoGuard reduces battery safety violations by **76.6%** compared to similarity-only top-1 reuse. Simultaneously, it reduces fallback calls by **21.4%** compared to always invoking local reasoning. On an NVIDIA Jetson AGX Xavier with local `llama3.2:3b` fallback reasoning, these reductions correspond to **3.67 seconds** and **36.97 Joules** of avoided fallback-reasoning overhead per trial. The results demonstrate that contract-based validation effectively filters out memory traps without incurring the full cost of always reasoning. The safety improvement is substantial, addressing the critical risk of unsafe actions in communication-limited settings. The efficiency gain is also significant, as fallback reasoning is computationally and energetically expensive. The combination of safety and efficiency makes MemoGuard suitable for mission-critical robots with limited onboard resources.
Communication-limited robots operating in mission-critical scenarios—such as disaster inspection and search-and-rescue—must make reliable onboard decisions without access to remote operators or high-capacity reasoning services. Episodic memory reuse offers a low-cost fallback: retrieve a past episode similar to the current context and reuse its action. However, retrieval similarity does not guarantee execution validity. A retrieved action may match the current context yet be unsafe due to changed topology, insufficient battery margin, or unreliable prior outcomes. The authors term such high-similarity but execution-invalid episodes **memory traps**. This creates a safety-efficiency design space: similarity-only reuse minimizes fallback cost but can be unsafe, while always invoking local reasoning improves safety at high computational and energy cost. The paper presents **MemoGuard**, a lightweight adaptive runtime that validates episodic memories against topology, resource, and outcome contracts before reuse, invoking fallback only when validation fails. The core problem is formalized as a trade-off between safety and efficiency, with the goal of achieving safe reuse without excessive reasoning overhead.
MemoGuard operates as an adaptive runtime layer that intercepts memory retrieval and validates candidate episodes before action execution. The validation is based on three contract types:
Why it matters
The paper highlights a fundamental tension in episodic memory reuse: similarity is a necessary but insufficient condition for safe action reuse. MemoGuard addresses this by introducing explicit contracts that capture execution validity beyond similarity. The 76.6% reduction in battery safety violations indicates that many high-similarity episodes are indeed memory traps, and that topology, resource, and outcome checks are effective filters. The 21.4% reduction in fallback calls shows that validation can avoid unnecessary reasoning, saving time and energy. The overhead of validation itself is not detailed, but the net benefit suggests it is low. The approach is generalizable to other domains where memory reuse is used under resource constraints. Limitations include the reliance on accurate contract definitions and the potential for false negatives (rejecting safe episodes) or false positives (accepting unsafe ones). Future work could explore learning contracts or adapting thresholds online. The open-source release at https://github.com/hetheiin/memoguard facilitates reproducibility and further development. Overall, MemoGuard provides a practical solution for safe and efficient memory reuse in communication-limited robot navigation.
Who should read this
CS practitioners and researchers
Opening member content…