AuthorsDianxing Zhang, Wendong Li, Kani Song et al.
TL;DR
Memory in Large Language Models uses a four-way memory taxonomy and a write–read–inhibit causal chain to unify mechanisms, evaluation, and governance without new benchmarks.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Memory in Large Language Models notes that definitions of parametric, contextual, and external memory are often conflated, causing irreproducible designs and non comparable results.
Memory in Large Language Models shows that RAG tasks and long context settings suffer from mid sequence drop and biased LLM as a Judge evaluations, directly impacting safety critical deployments.
HOW IT WORKS
Memory in Large Language Models centers on a unified definition of memory and a memory quadruple across parametric memory, contextual memory, external memory, and procedural episodic memory.
Memory in Large Language Models treats parametric weights like long term storage, contextual windows like working RAM, external stores like disk, and procedural timelines like a transaction log.
Memory in Large Language Models’ key write read inhibit chain enables systematic evaluation and governance that plain context windows and ad hoc RAG pipelines cannot provide.
DIAGRAM
This diagram shows how Memory in Large Language Models runs parametric only, offline retrieval, and online retrieval regimes on the same data slice.
DIAGRAM
This diagram shows how Memory in Large Language Models coordinates DAPT TAPT, PEFT, model editing, and RAG in the DMM Gov closed loop.
PROCESS
Parametric Only setting
Memory in Large Language Models runs the Parametric Only setting to measure closed book recall and privacy risks using the memory quadruple and parametric memory definitions.
Offline Retrieval setting
Memory in Large Language Models fixes the index snapshot in the Offline Retrieval setting to decouple retrieval quality and faithfulness for external memory.
Online Retrieval setting
Memory in Large Language Models enables Online Retrieval to measure freshness hits, outdated answers, and refusal slices under temporal governance.
DMM Gov dynamic governance
Memory in Large Language Models applies DMM Gov dynamic governance to coordinate DAPT TAPT, PEFT, model editing, and RAG with reversible rollback and change audit certificates.
KEY CONTRIBUTIONS
Unified definitions and taxonomy
Memory in Large Language Models defines memory as persistent and addressable state and introduces the memory quadruple across parametric, contextual, external, and procedural episodic memory.
Three regime evaluation protocol
Memory in Large Language Models proposes parallel Parametric Only, Offline Retrieval, and Online Retrieval regimes to decouple model capability from information availability on the same data slice.
DMM Gov dynamic governance
Memory in Large Language Models introduces DMM Gov dynamic governance coordinating DAPT TAPT, PEFT, ROME MEND MEMIT SERAC, and RAG into an auditable closed loop for updating and forgetting.
RESULTS
Freshness hits
N A
vs RAGBench baseline N A
Unsupported Claim Rate
N A
vs RAGAS baseline N A
Edit Success Rate
N A
context ESR defined with target suppression and locality
Verbatim reproduction rate
N A
context linked to de duplication thresholds and extraction risk
Memory in Large Language Models is a framework and survey paper without new benchmark numbers, so the key result is the unified taxonomy, metric families, and governance protocols rather than a specific leaderboard score.
BENCHMARK
Relative emphasis across parametric, contextual, external, and procedural memory in the unified framework.
KEY INSIGHT
Memory in Large Language Models argues that enlarging context windows alone does not fix long context issues because mid span information still suffers from a mid sequence drop.
This is counterintuitive because many practitioners assume longer windows automatically improve memory, but Memory in Large Language Models shows positional bias and attention dilution can dominate.
WHY IT MATTERS
Memory in Large Language Models unlocks a shared coordinate system for memory mechanisms, evaluation metrics, and governance processes across parametric, contextual, external, and procedural memory.
Builders can now design RAG and long context systems with aligned regimes, metric families, and DMM Gov governance, making deployments more reproducible, comparable, and auditable.
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.
Yijie Zhong, Yunfan Gao, Haofen Wang
· 2026
HingeMem combines Boundary Guided Long-Term Memory, Dialogue Boundary Extraction, Memory Construction, Query Adaptive Retrieval, Hyperedge Rerank, and Adaptive Stop to segment dialogues into element-indexed hyperedges and plan query-specific retrieval. On LOCOMO, HingeMem achieves 63.9 overall F1 and 75.1 LLM-as-a-Judge score, surpassing the best baseline Zep (56.9 F1) by 7.0 F1 without using category-specific QA formats.
Answers use this explainer on Memory Papers.
Checking…