A Topology-Aware, Memory-Centric Architecture that Separates Root-Cause Derivation from Root-Cause Explanation

Researchers present OpsCortex, a multi-agent system that uses persistent operational memory and dependency graphs to automatically derive root causes of microservice failures, then leverages LLMs only for explanation rather than diagnosis. The architecture separates root-cause derivation from explanation, addressing a critical gap in autonomous operations by maintaining structured system knowledge that typical monitoring stacks discard.

OpsCortex tackles a fundamental challenge in modern cloud infrastructure: when cascading failures propagate across service dependencies, engineers spend more time reconstructing system context than solving problems. The research identifies that current monitoring solutions excel at detection but fail at diagnosis, creating a bottleneck where scarce expert knowledge becomes the constraint. This work addresses that gap by proposing a tiered memory architecture that learns system topology and historical failure patterns, enabling deterministic root-cause identification independent of machine learning uncertainty.

The approach represents a meaningful shift in how the industry thinks about observability and AI-assisted diagnosis. Rather than applying increasingly powerful language models to the diagnosis problem—a common industry trend—the authors argue for separating concerns: deterministic systems handle what can be mechanically derived from dependency graphs and temporal data, while LLMs focus on explaining and contextualizing findings. This mirrors broader maturation in AI application design, where task decomposition often outperforms end-to-end black-box approaches.

For DevOps teams and platform engineers, this work validates that operational memory and topology-awareness are critical infrastructure gaps. Cloud platforms that embed such capabilities gain competitive advantages in reducing mean-time-to-resolution (MTTR) and enabling smaller teams to manage larger systems. The validation on e-commerce benchmarks with eight failure scenarios provides practical grounding, though real-world deployment at scale remains to be demonstrated. The research implications extend beyond incident response into capacity planning and architectural decision-making, where understanding service dependencies informs optimization.

• →Root cause derivation should be deterministic and graph-based, not delegated to probabilistic models or LLMs
• →Operational memory—structured knowledge of system topology and failure history—is the missing layer in autonomous operations
• →Separating root-cause discovery from explanation improves both accuracy and explainability in microservice diagnostics
• →Multi-tier memory architecture enables engineering teams to scale beyond human expert bottlenecks in incident response
• →LLMs are more effective at explanation and recommendation tasks than at causal inference from distributed system data