AuthorsDac Duy Anh Nguyen, Zhangchi Qiu, Shigeng Chen, Alan Wee-Chung Liew
TL;DR
Graph-Based Personalized Memory for LLM Agents uses a lifecycle taxonomy over representation, evolution, retrieval, and evaluation to unify fragmented graph-memory designs.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
LLM agents are evolving into long-term assistants, but dialogue logs and flat summaries cannot capture how user facts interrelate or change over time.
This breaks sustained personalization for tasks like dynamic preference tracking, where agents misinterpret evolving goals and constraints, degrading user modeling and decision quality.
HOW IT WORKS
Graph-Based Personalized Memory for LLM Agents structures personalized memory into memory representation, memory evolution, memory retrieval, and memory evaluation as explicit lifecycle stages.
Think of it like a card catalog plus history log for an LLM assistant, where each user fact is linked, versioned, and reachable instead of buried in a giant text archive.
This lifecycle lets Graph-Based Personalized Memory for LLM Agents support multi-hop reasoning, preference drift, and provenance-aware updates that a plain context window or vector store cannot manage.
DIAGRAM
This diagram shows how Graph-Based Personalized Memory for LLM Agents organizes personal memory dimensions and graph representation patterns.
DIAGRAM
This diagram shows how Graph-Based Personalized Memory for LLM Agents groups benchmarks by what they test in personalized memory.
PROCESS
Memory Representation
Graph-Based Personalized Memory for LLM Agents defines node roles, relation types, and lifecycle metadata to encode semantic, temporal, relational, experiential, and affective user information.
Memory Evolution
Graph-Based Personalized Memory for LLM Agents analyzes admission, integration, conflict resolution, consolidation, and removal to maintain an accurate active user state over time.
Memory Retrieval
Graph-Based Personalized Memory for LLM Agents categorizes similarity-based, structure-based, and adaptive retrieval to map queries to the right personalized graph context.
Memory Evaluation
Graph-Based Personalized Memory for LLM Agents surveys benchmarks and metrics to assess answer quality, evidence retrieval, user-state modeling, evolution, structure, and efficiency.
KEY CONTRIBUTIONS
Lifecycle Taxonomy
Graph-Based Personalized Memory for LLM Agents introduces a lifecycle taxonomy over memory representation, memory evolution, memory retrieval, and memory evaluation for personalized graph memory.
Evaluation Survey
Graph-Based Personalized Memory for LLM Agents summarizes benchmarks like LoCoMo, PersonaMem-v2, EngramaBench, and EvoMemBench, highlighting gaps in graph quality and user control.
Future Directions
Graph-Based Personalized Memory for LLM Agents outlines challenges in scalable lifelong memory, multimodal personal graphs, and causal and counterfactual user modeling.
RESULTS
Benchmark Count
12 benchmarks
covers more settings than single-benchmark surveys
Lifecycle Stages
4 stages
representation, evolution, retrieval, evaluation
Memory Dimensions
5 dimensions
semantic, temporal, relational, experiential, affective
Representation Patterns
5 patterns
flat, hierarchical, hypergraph, hybrid, multiple graphs
Graph-Based Personalized Memory for LLM Agents aggregates at least twelve named benchmarks across long-horizon recall and personalization. This breadth shows how the taxonomy spans answer quality, retrieval, user modeling, evolution, and structure for LLM agents.
BENCHMARK
Graph-Based Personalized Memory for LLM Agents aggregates at least twelve named benchmarks across long-horizon recall and personalization. This breadth shows how the taxonomy spans answer quality, retrieval, user modeling, evolution, and structure for LLM agents.
BENCHMARK
Count of core lifecycle components discussed by Graph-Based Personalized Memory for LLM Agents.
KEY INSIGHT
Graph-Based Personalized Memory for LLM Agents notes that adopting a graph alone does not guarantee better retrieval or personalization quality.
This is counterintuitive because many assume structured graphs automatically improve memory, but the survey shows lifecycle design and evaluation practices are equally critical.
WHY IT MATTERS
Graph-Based Personalized Memory for LLM Agents gives builders a clear blueprint for designing personalized memory across representation, evolution, retrieval, and evaluation.
With this structure, developers can systematically compare graph-based memories, choose appropriate benchmarks, and design agents that maintain trustworthy long-term personalization rather than ad hoc context hacks.
Jingbiao Mei, Jinghong Chen et al.
arXiv 2026 · 2026
ATM-Bench structurally evaluates long-term multimodal personal memory using Memory Ingestion, Retrieval, and Answer Generation with Schema-Guided Memory and Descriptive Memory variants. On ATM-Bench-Hard, Oracle with SGM reaches 47.3% QS while the best full system stays under 20% accuracy, revealing a large gap.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
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.
Answers use this explainer on Memory Papers.
Checking…