ISSRE 2026 Dual-Memory Diagnosis
1Nankai University 2Tsinghua University 3Huawei Technologies Co., Ltd.
ISSRE 2026
* Corresponding author
01 THE PAPER
Failure diagnosis in modern software systems requires iterative evidence acquisition and hypothesis reasoning guided by operational experience. Existing LLM-based methods improve diagnosis through agentic reasoning or knowledge augmentation, but they often lack a mechanism to coordinate the evolving diagnostic state with operational experience during iterative diagnosis. We propose OpsMem, a dual-memory framework that maintains a short-term memory for the current diagnostic state and a long-term memory for reusable operational experience. OpsMem uses cross-memory resonance to activate state-relevant long-term memory, conditions multi-agent diagnosis on the short-term and activated long-term memories, and consolidates reusable experience from solved incidents back into long-term memory. Experiments on a real-world Huawei microservice failure diagnosis dataset show that OpsMem outperforms representative agentic-reasoning and knowledge-augmented baselines, improving Match and Relevant by up to 46.88% and 18.39% over the strongest baseline, respectively.
A graph-structured memory of the current diagnostic state: observed symptoms, acquired evidence, and candidate hypotheses with their relationships.
Reusable operational experience distilled from past incidents: diagnostic patterns, cases, and procedures that future incidents can draw on.
The coupling mechanism that activates state-relevant long-term memory as the diagnostic state evolves, so retrieved experience matches what the diagnosis needs now.
02 THE FRAMEWORK
OpsMem coordinates two memories during iterative, multi-agent failure diagnosis, and learns from every solved incident.
Cross-memory resonance couples the short-term state with long-term experience through signal coupling, pattern activation, and memory propagation, retrieving the cases and procedures that the current hypotheses actually need.
Multi-agent diagnosis is conditioned on both the short-term memory and the activated long-term memory, so evidence acquisition and hypothesis updates stay aligned with relevant operational experience.
After an incident is solved, consolidation distills reusable experience back into long-term memory, so the knowledge available to future diagnoses keeps growing.
03 EVALUATION
Evaluation on 120 real-world failure incidents from Huawei production microservice systems, against agentic-reasoning and knowledge-augmented baselines (ReAct, GoS, VectorRAG, GraphRAG, LinearRAG). Improvements are relative to the strongest baseline (Table I).
Match
up to +46.88%
Relative improvement over strongest baseline
Match measures agreement with the reference diagnosis. The figure is a relative improvement, not an accuracy value.
Table IRelevant
up to +18.39%
Relative improvement over strongest baseline
Relevant measures how much of the reference diagnosis the answer covers. Also a relative improvement, not accuracy points.
Table IEvaluation data
120 incidents
Real-world Huawei failures
Production microservice incidents. The public release contains one sanitized case and a small sanitized seed knowledge base, not the full dataset.
Experiments04 START FROM THE CODE
The public release contains the framework code, one sanitized example case, and a small sanitized seed knowledge base. The full dataset and knowledge base are not publicly released.
# From the repository's code directory, after configuring config.yaml
python main.py
The pipeline runs failure diagnosis over the configured case and writes evaluation outputs.
Repository entry point; local data and model setup are required.
05 FREQUENTLY ASKED
OpsMem is a dual-memory reasoning framework for failure diagnosis, accepted at ISSRE 2026. It maintains a short-term memory for the current diagnostic state — observed symptoms, acquired evidence, and candidate hypotheses — and a long-term memory of reusable operational experience, and connects the two through cross-memory resonance.
Existing LLM-based diagnosticians either reason without operational experience or retrieve knowledge that ignores where the diagnosis currently stands. OpsMem's cross-memory resonance activates state-relevant long-term memory as the diagnostic state evolves, conditions multi-agent diagnosis on both memories, and consolidates reusable experience from solved incidents back into long-term memory for future incidents.
OpsMem was evaluated on 120 real-world failure incidents from Huawei production microservice systems, against agentic-reasoning and knowledge-augmented baselines including ReAct, GoS, VectorRAG, GraphRAG, and LinearRAG. It improves Match and Relevant by up to 46.88% and 18.39% over the strongest baseline respectively (Table I). These are relative improvements, not accuracy points.
The framework code is open at github.com/gaorch85/OpsMem under Apache-2.0, including one sanitized example case and a small sanitized seed knowledge base. The full dataset and knowledge base cannot be publicly released due to confidentiality and compliance restrictions, so results on the private Huawei data are not reproducible from the public release alone.
06 CITE THIS WORK
@article{sun2026opsmem,
title={OpsMem: Dual-Memory Reasoning with Cross-Memory Resonance for Failure Diagnosis},
author={Sun, Yongqian and Gao, Rongchen and Luo, Yu and Gu, Wenwei and Zhang, Shenglin and Guo, Qingyi and Fu, Qiuai and Wu, Yaoliang and Pei, Dan},
journal={arXiv preprint arXiv:2607.11357},
year={2026}
}