PaperlayerView exact revision on arXiv

Source-linked AI summary

Agentic Memory Enhanced Recursive Reasoning for Root Cause Localization in Microservices

Lingzhe Zhang, Tong Jia, Yunpeng Zhai, Leyi Pan, Chiming Duan, Minghua He, Mengxi Jia, Ying Li

arXiv:2601.02732v1cs.SEcs.AI

TL;DR

Complex microservice dependencies make accurate root cause localization difficult, while existing approaches suffer from shallow reasoning and limited cross-alert reuse. AMER-RCL derives recursive, multi-dimensional, and cross-modal reasoning from SRE practice, combining recursive multi-agent analysis with agentic memory. It consistently outperforms state-of-the-art methods in localization accuracy and inference efficiency.

  • Problem

    Accurate root cause localization is difficult in complex microservices, while existing methods struggle with evolving contexts, shallow symptom-centric reasoning, and redundant per-alert analysis.

  • Method

    AMER-RCL combines a Recursive Reasoning RCL engine with Agentic Memory that accumulates and reuses reasoning across alerts within a time window.

  • Results

    AMER-RCL consistently outperforms state-of-the-art methods in localization accuracy and inference efficiency.

  • Takeaways & Limitations

    Recursive reasoning and reusable alert history provide the framework’s supported route to more accurate and efficient root cause localization.

  • Takeaways & Limitations

    The empirical study’s reliance on professional SRE judgments may introduce bias because diagnostic processes are inherently subjective.

Abstract

from arXiv · show

As contemporary microservice systems become increasingly popular and complex-often comprising hundreds or even thousands of fine-grained, interdependent subsystems-they are experiencing more frequent failures. Ensuring system reliability thus demands accurate root cause localization. While many traditional graph-based and deep learning approaches have been explored for this task, they often rely heavily on pre-defined schemas that struggle to adapt to evolving operational contexts. Consequently, a number of LLM-based methods have recently been proposed. However, these methods still face two major limitations: shallow, symptom-centric reasoning that undermines accuracy, and a lack of cross-alert reuse that leads to redundant reasoning and high latency. In this paper, we conduct a comprehensive study of how Site Reliability Engineers (SREs) localize the root causes of failures, drawing insights from professionals across multiple organizations. Our investigation reveals that expert root cause analysis exhibits three key characteristics: recursiveness, multi-dimensional expansion, and cross-modal reasoning. Motivated by these findings, we introduce AMER-RCL, an agentic memory enhanced recursive reasoning framework for root cause localization in microservices. AMER-RCL employs the Recursive Reasoning RCL engine, a multi-agent framework that performs recursive reasoning on each alert to progressively refine candidate causes, while Agentic Memory incrementally accumulates and reuses reasoning from prior alerts within a time window to reduce redundant exploration and lower inference latency. Experimental results demonstrate that AMER-RCL consistently outperforms state-of-the-art methods in both localization accuracy and inference efficiency.

1 Introduction

Microservice root cause localization is difficult because complex, evolving dependencies challenge traditional methods, while LLM approaches still reason shallowly and reuse little across alerts. AMER-RCL models SRE practices with recursive, multi-dimensional, cross-modal reasoning and agentic memory, outperforming existing methods in accuracy and efficiency.

  • 1 Introduction: Traditional graph-based and deep learning approaches struggle with interpretability, cross-platform generality, and adaptation to evolving system states.
  • 1 Introduction: Existing LLM methods often conflate downstream symptoms with true causes because their reasoning remains shallow and symptom-centric.
  • 1 Introduction: SRE root cause analysis is recursive, expands vertically and horizontally, and validates trace-based candidates with correlated metrics and logs.
  • 1 Introduction: AMER-RCL combines recursive multi-agent reasoning for each alert with memory that reuses prior reasoning and focuses exploration on alert divergences.
  • 1 Introduction: AMER-RCL surpasses state-of-the-art methods in localization accuracy and inference efficiency across three datasets.It exceeds the second-best approach by approximately 16% in accuracy and achieves more than 3.5× speedup over existing LLM-based methods for batch alert analysis.

2 Background

The background establishes root cause localization as identifying the responsible component in complex distributed systems. Distributed traces record request execution across services and operations, while traditional methods construct dependency graphs and rank likely causes.

  • 2 Background: Root cause localization identifies the specific node, service, or pod responsible for a failure rather than merely classifying its category.
  • 2 Background: Traditional methods construct component dependency graphs from traces or logs and apply graph algorithms to rank likely root causes.
  • 2 Background: Distributed traces record request execution, timing, dependencies, and performance through spans representing individual operations.

3 Empirical Study

The empirical study characterizes SRE root cause analysis as recursive narrowing, multi-dimensional expansion, and cross-modal validation. Practitioners recursively follow anomalous dependencies, correlate services and infrastructure across alerts, and use metrics and logs to distinguish causes from propagated symptoms.

  • 3 Empirical Study: The study combines researcher analysis with semi-structured interviews involving 22 professional developers and SREs across four organizations.The process produced a validated characterization organized around how practitioners narrow, expand, and differentiate root causes.
  • 3.1 Recursiveness: SREs recursively follow anomalous downstream operations from an alert’s entry span until the candidate root cause is isolated.In the example, this narrows candidates to frontend2-0 and recommendationservice2-0 after downstream timeouts disappear at productcatalogservice-0.
  • 3.2 Multi-Dimensional Expansion: SREs expand analysis vertically across service, instance, pod, and host layers and horizontally across related alerts and traces.These complementary expansions expose infrastructure-level and cross-service dependencies within the same anomaly window.
  • 3.3 Cross-Modal Reasoning: Cross-modal reasoning distinguishes causes from candidate anomalies by comparing metrics and logs across pods, services, and nodes.The study identifies the component with the most significant or earliest fluctuation as the likely true cause.

4 AMER-RCL

AMER-RCL turns the study’s SRE practices into an architecture that combines causal graph extraction, recursive multi-agent reasoning, and agentic memory. Its modules process multiple alerts in a time window while preserving prior reasoning for reuse.

  • 4 AMER-RCL: AMER-RCL transforms alerts into causal graphs capturing anomalous request paths and contextual dependencies before comparing alerts.
  • 4 AMER-RCL: Agentic Memory records prior reasoning outcomes and intermediate deliberations so similar alerts can skip redundant analysis while divergences receive recursive exploration.
  • 4 AMER-RCL: Its Recursive Reasoning RCL module coordinates specialized trace, log, and metric agents through a MetaAgent and consolidates their findings.

4.1 Agentic Memory

Agentic Memory stores causal graphs with their reasoning trajectories, retrieves similar prior alerts, and uses similarity thresholds to reuse or restart reasoning for new alerts.

  • 4.1 Agentic Memory: Agentic Memory couples each alert’s causal graph with its associated reasoning trajectory in a structured, queryable repository.Each memory entry includes a graph fingerprint, embedding, full graph, metadata, and node-linked reasoning transcript.
  • 4.1 Agentic Memory: Each alert is deterministically converted into a causal graph whose nodes carry service, pod, operation, metric, and log attributes, while edges encode invocations or causal dependencies.Edge attributes include call latency and response status.
  • 4.1 Agentic Memory: After Recursive Reasoning RCL executes, the graph and reasoning steps are stored as a memory entry for reuse in subsequent executions.
  • 4.1 Agentic Memory: Memory retrieval first uses embedding-based nearest-neighbor search, then evaluates structural and attribute similarity between the new alert graph and stored graphs.Structural similarity and node-attribute agreement jointly determine which prior entries are relevant.
  • 4.1 Agentic Memory: Similarity thresholds determine whether prior reasoning is reused, continued from a divergence point, or re-initiated for low-similarity alerts.High similarity triggers skipping or reuse, while low similarity triggers reasoning from the beginning.

4.2 Specialized Agents

AMER-RCL uses specialized agents to filter and analyze traces, logs, and metrics, while a Consolidator integrates their observations into structured evidence for recursive root-cause reasoning.

  • 4.2 Specialized Agents: AMER-RCL primarily uses Trace, Log, Metric, and Consolidator agents, each assigned a focused role in the multi-agent reasoning process.The specialized agents analyze distinct data sources, while the Consolidator combines their outputs for the MetaAgent.
  • 4.2 Specialized Agents: Trace Agent filtering reduces the enormous invocation-path context by retrieving only trace data relevant to the current root-cause analysis.Trace data records service-call sequences, and the agent returns selected child spans with metadata.
  • 4.2 Specialized Agents: The Trace Agent returns child spans for a given span together with timestamps, service and operation names, durations, and status codes.
  • 4.2 Specialized Agents: The Log Agent filters logs by a time window, target component, and relevance criterion to retain entries most indicative of anomalies.Relevance considers message type, error codes, and correlations with other alert or trace events, reducing context while preserving informative logs.
  • 4.2 Specialized Agents: The Metric Agent selectively retrieves metrics with notable deviations, reflecting the observation that most metrics remain stable during anomalies.It examines metrics associated with the target component and related infrastructure using historical means and standard deviations.
  • 4.2 Specialized Agents: The Consolidator aggregates intermediate inferences into a unified representation containing evidential support for each hypothesized root cause.

4.3 MetaAgent

The MetaAgent coordinates recursive, multi-agent root cause analysis that explores trace candidates, verifies suspicious spans across modalities, and consolidates results into a ranked cause set.

  • Recursive Reasoning: Recursive reasoning iteratively generates instructions conditioned on prior outcomes, activates subordinate agents, and backtracks when an analytical path is exhausted.This contrasts with linear Chain-of-Thought reasoning and supports broader exploration of candidate causes.
  • Multi-Modal Analysis: Trace data drives the recursion from an alert’s entry span, while logs and metrics provide cross-modal verification of suspicious candidates.Only spans corroborated across these modalities enter the potential root cause set.
  • Depth-Assured Reasoning: The depth-assured framework uses Initial Reasoning, Critical Reflection, and Final Review to prevent premature termination and recency-biased ranking.Critical Reflection continues analysis of suspicious spans, while Final Review consolidates outputs from all recursive calls.
  • Initial Reasoning: Initial Reasoning broadens the trace search with only the MetaAgent and TraceAgent, producing a coarse-grained frontier for later refinement.The Metrics and Log Agents are withheld during this initial expansion stage.
  • Final Review: Final Review invokes the Consolidator to synthesize recursive outputs into the final ranked root cause set, retaining coverage of unexplored candidates.Reviewing the complete trajectory is intended to prevent premature candidate dismissal.

5 Evaluation

The evaluation examines AMER-RCL across three datasets, comparing its accuracy with LLM- and non-LLM-based baselines under specified localization metrics and experimental settings.

  • Evaluation Design: AMER-RCL is evaluated through accuracy, inference-latency, backbone, and component-contribution research questions.These questions cover comparisons with baselines, efficiency against LLM methods, LLM-backbone effects, and ablations.
  • Datasets: Experiments use the AIOPS 2022, TrainTicket, and FAMOS-Mall datasets, spanning real-world, benchmark, and production-demonstration microservice settings.AIOPS 2022 is a large-scale real-world e-commerce platform, while TrainTicket uses injected faults in a 90-instance deployment.
  • Experimental Setup: Accuracy is measured with Recall@k and mean reciprocal rank, using Claude-3.5 Sonnet, a 3-sigma metric threshold, and a 60-second anomaly window unless otherwise specified.The study compares both LLM-based and non-LLM-based methods.
  • Accuracy: AMER-RCL achieves an average MRR improvement of 15.75% over the second-best method across all three datasets and consistently outperforms SOTA baselines.Table 1 reports the root cause localization accuracy comparison; the cited prose also gives dataset-specific MRR values and margins.
  • Accuracy: AMER-RCL combines prior-reasoning reuse with recursive analysis to avoid redundant computation while addressing heterogeneous-alert context and attention constraints.The comparison discusses batch analysis, independent alert processing, and AMER-RCL’s memory-based reuse strategy.

5.3 Efficiency

AMER-RCL improves inference efficiency through Agentic Memory and shows accuracy gains from deeper reasoning stages and stronger LLM backbones.

  • 5.3 Efficiency: AMER-RCL is on average 3.53× faster than RCAgent, 31.05× faster than mABC, and 14.61× faster than CoT in inference speed.These comparisons are reported in Table 2’s seconds-per-query evaluation.
  • 5.3 Efficiency: Agentic Memory reduces redundant computation by reusing prior reasoning outcomes, with the largest speedup occurring when alerts are largely homogeneous.This allows AMER-RCL to skip repeated reasoning for similar alerts.
  • 5.4 LLM Backbone Impact: Claude-3.5-Sonnet achieves the best backbone results, outperforming the second-best model by an average of 18.14%.DeepSeek-R1-Qwen ranks second overall but underperforms Qwen-2.5-Plus on subsets A and B.
  • 5.5 Ablation Study: Agentic Memory improves MRR from 71.13% to 73.89% on subset A and from 72.41% to 74.92% on subset Z.The gains reflect reuse of prior reasoning and knowledge consolidation across alerts.
  • 5.5 Ablation Study: Critical Reflection improves average MRR by 13.21%, Final Review adds 2.96%, and Agentic Memory adds 2.61%.The ablation evaluates Initial Reasoning, Critical Reflection, Final Review, and Agentic Memory as successive design components.

6 Related Work

Related work spans metrics-based, trace-based, and LLM-based approaches to failure management and root cause localization, each using different representations and reasoning strategies.

  • Metrics-Based Approaches: Metrics-based methods analyze response time, throughput, and resource utilization to infer anomalous or problematic components.Examples include causal graphs, Bayesian networks, time-series forecasting, and personalized PageRank.
  • Trace-Based Approaches: Trace-based methods follow execution paths and intercomponent interactions to localize abnormal services, instances, or latency-critical operations.Representative techniques include trace coverage trees, spectrum analysis, critical-path analysis, and sequential pattern mining.
  • LLM-Based Approaches: LLM-based failure-management research covers anomaly detection, diagnosis, and automated mitigation, using foundation models, fine-tuning, and prompting strategies.Prompt-based examples structure diagnosis with in-context learning or Chain-of-Thought reasoning.

7 Conclusion

AMER-RCL combines recursive reasoning with agentic memory to improve root cause localization accuracy and inference efficiency. The authors identify smaller-scale models and broader failure-management coverage as future directions.

  • AMER-RCL consistently outperforms state-of-the-art methods in both root cause localization accuracy and inference efficiency.
  • Recursive reasoning refines candidate causes for each alert, while agentic memory reuses prior-alert reasoning within a time window to reduce redundant exploration and latency.
  • Future work will investigate accurate and efficient localization with smaller-scale models and extend the approach across the entire failure-management process.
Loading 2601.02732v1…