Invalidation Contracts Enhance LLM Agent Memory Efficiency

Large Language Model (LLM) agents often cache solutions to API errors, allowing them to reuse these "recovery suggestions" in subsequent interactions and save on token costs and model calls. However, when server-side data changes, these cached fixes can become outdated, leading to silent failures. The traditional solution of re-deriving fixes every time negates any efficiency gains. Researchers propose "invalidation contracts" as a novel protocol layer to address this. This system embeds version stamps and cacheability hints directly into each recovery suggestion. This enables the client-side agent to intelligently evict stale entries without needing trial-and-error, while retaining valid ones. The effectiveness of this approach is measured by two factors: "validity," which is the fraction of cached suggestions that remain correct after data drift, and "compliance," which is the fraction of times the agent's planner successfully applies a cached fix on the first attempt. The study found that validity, being protocol-dependent, remained consistently high. However, compliance varied significantly across different LLM planner models. For instance, Claude Haiku 4.5 achieved 100% first-try compliance, while Claude Sonnet 5 showed much lower rates, often refusing fixes that added new fields. Row-level invalidation, which precisely targets individual stale entries, proved highly effective, recovering 29-33% of baseline token costs on several models and significantly raising compliance compared to less precise table-level invalidation. The contract adds a modest 15% to the response payload.