AuthorsAnbi Guo, Mahfuza Farooque
TL;DR
DM-RAG uses dual memories plus Bayesian fusion to push recall to 98.70% on UNSW-NB15 with Phi-4-mini.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Structured security logs for APT detection span long periods, but DM-RAG’s target domain exceeds typical LLM context windows and domain priors, causing missed attacks.
When DM-RAG is not used, Phi-4-mini-based systems struggle on UNSW-NB15, leading to low recall and undetected multistage attacks, harming intrusion detection and threat response.
HOW IT WORKS
DM-RAG’s core mechanism combines Short-Term Memory (STM), Long-Term Memory (LTM), logistic regression confidence model, and Bayesian fusion around an instruction-tuned Phi-4-mini.
You can think of STM as RAM and LTM as disk, with FAISS acting like an indexed card catalog for historical attack patterns.
This dual-memory design lets DM-RAG reason over days of logs, combining recent and historical context in ways a plain context window cannot.
DIAGRAM
This diagram shows how DM-RAG processes each UNSW-NB15 log entry, updates STM and LTM, and applies Bayesian fusion when STM is full.
DIAGRAM
This diagram shows how DM-RAG is trained and evaluated on UNSW-NB15 with logistic regression, instruction tuning, and test-time dual-memory prompting.
PROCESS
Step 1: Confidence Model Preparation
DM-RAG trains the logistic regression confidence model on normalized UNSW-NB15 features, defining score(x) as P(y = 1 | x) for anomaly confidence.
Step 2: Online Log Analysis
DM-RAG encodes each log with the Encoder E, retrieves from Long-Term Memory (LTM) via FAISS, merges STM and LTM into a prompt, and queries Phi-4-mini.
Step 3: Memory Generation
DM-RAG parses summary, confidence, and label, then appends them into Short-Term Memory (STM) as a sliding window of K = 10 recent summaries.
Step 4: Memory Compression and Promotion
When STM is full, DM-RAG compresses summaries with Phi-4-mini, applies Bayesian fusion to compute conffused, and promotes high-confidence summaries into Long-Term Memory (LTM).
KEY CONTRIBUTIONS
Dual-Memory Retrieval-Augmented Generation (DM-RAG)
DM-RAG introduces interacting Short-Term Memory (STM) and Long-Term Memory (LTM) modules around Phi-4-mini, enabling high-recall log anomaly detection with 98.70% recall and 69.59% F1 on UNSW-NB15.
Bayesian Fusion for Memory Promotion
DM-RAG uses Bayesian fusion over Beta-modeled features to compute fused anomaly confidence, deciding which STM summaries are compressed and promoted into LTM.
Instruction-Tuned Phi-4-mini for Structured Logs
DM-RAG instruction-tunes Phi-4-mini with a strict JSON prompt design, combining STM and LTM context to produce interpretable summaries and attack labels for structured security logs.
RESULTS
Accuracy
53.64%
-3.60 over Phi-4 + RAG (MITRE)
Precision
53.74%
+8.82 over LoRA Fine-tuned
Recall
98.70%
+57.13 over Phi-4 + RAG (MITRE)
F1 Score
69.59%
+17.89 over Phi-4 + RAG (MITRE)
On the UNSW-NB15 intrusion detection benchmark, DM-RAG is evaluated on accuracy, precision, recall, and F1. The 98.70% recall and 69.59% F1 show that DM-RAG captures almost all attacks while maintaining balanced precision compared to Phi-4-mini baselines.
BENCHMARK
On the UNSW-NB15 intrusion detection benchmark, DM-RAG is evaluated on accuracy, precision, recall, and F1. The 98.70% recall and 69.59% F1 show that DM-RAG captures almost all attacks while maintaining balanced precision compared to Phi-4-mini baselines.
BENCHMARK
F1 Score on UNSW-NB15 for DM-RAG and Phi-4-mini baselines.
KEY INSIGHT
DM-RAG achieves 98.70% recall with only 53.74% precision, while zero-shot Phi-4-mini reaches 98.91% precision but just 0.20% recall.
This is surprising because high precision is usually desirable, yet DM-RAG shows that for security logs, aggressively maximizing recall is more valuable than ultra-conservative precision.
WHY IT MATTERS
DM-RAG unlocks high-recall, interpretable, long-horizon reasoning over structured security logs using a compact Phi-4-mini backbone and dual-memory RAG.
Builders can now deploy lightweight, memory-augmented LLM detectors that track multistage attacks over time without massive external corpora or heavy fine-tuning.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Okan Bursa
· 2026
Adaptive RAG Memory (ARM) augments a standard retriever–generator stack with a Dynamic Embedding Layer and Remembrance Engine that track usage statistics and apply selective remembrance and decay to embeddings. On a lightweight retrieval benchmark, ARM achieves NDCG@5 ≈ 0.9401 and Recall@5 = 1.000 with 22M parameters, matching larger baselines like gte-small while providing the best efficiency among ultra-efficient models.
Wei Zhou, Xuanhe Zhou et al.
arXiv 2026 · 2026
Are We Ready For An Agent-Native Memory System? analyzes Memory Representation and Storage, Memory Extraction, Memory Retrieval and Routing, and Memory Maintenance across 12 real systems like Mem0, Zep, MemTree, LightMem, MemOS, MemoryOS, and A-MEM. The study’s main result is that structured systems such as Zep reach 48.0 LLM Judge Accuracy on LongMemEval while Long Context reaches 19.0, and that localized maintenance strategies like LightMem achieve 48.3% normalized utility at only 3.67 s per query.
Answers use this explainer on Memory Papers.
Checking…