AuthorsChang Yang, Chuang Zhou, Yilin Xiao et al.
TL;DR
Graph-based Agent Memory uses a unified graph-centric taxonomy and lifecycle (extraction, storage, retrieval, evolution) to systematize dozens of agent memory techniques and applications.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
LLM-based agents suffer from knowledge cutoff, tool incompetence, and performance saturation, repeatedly failing long-horizon tasks without learning from past interactions.
These failures break complex applications like multi-turn dialogue, game playing, and scientific discovery, where agents must retain experience and knowledge across sessions to avoid repeated errors.
HOW IT WORKS
Graph-based Agent Memory centers on a lifecycle with Memory Extraction, Memory Storage, Memory Retrieval, and Memory Evolution, grounded in Knowledge vs Experience Memory and structural graph choices.
You can think of this as upgrading agents from a flat log file to a rich card catalog plus timeline, where graphs act like a relational database of experiences and facts.
This graph-centric mechanism lets Graph-based Agent Memory support relational reasoning, hierarchical organization, and temporal evolution that a plain context window or vector store alone cannot provide.
DIAGRAM
This diagram shows how Graph-based Agent Memory categorizes memory types and maps them to structural vs non-structural implementations.
DIAGRAM
This diagram shows how Graph-based Agent Memory organizes representative systems along the extraction, storage, retrieval, and evolution stages.
PROCESS
Memory Extraction
Graph-based Agent Memory transforms raw dialogue, trajectories, and multimodal data into structured entities, relations, embeddings, and summaries, initializing Knowledge Memory and Experience Memory contents.
Memory Storage
Graph-based Agent Memory organizes extracted units into Knowledge Graphs, Hierarchical Structures, Temporal Graphs, Hypergraphs, and hybrid memories tailored to long-term and short-term roles.
Memory Retrieval
Graph-based Agent Memory applies similarity, rule, temporal, graph, RL, and agent-based operators to recall relevant subgraphs and episodes for reasoning over current queries.
Memory Evolution
Graph-based Agent Memory updates nodes, edges, and summaries via internal self-evolving and external self-exploration, consolidating experiences and resolving outdated or conflicting information.
KEY CONTRIBUTIONS
Taxonomy of Agent Memory
Graph-based Agent Memory proposes a multi-dimensional taxonomy: Short-term vs Long-term, Knowledge vs Experience, and Non-structural vs Structural Memory, grounded in graph implementations.
Lifecycle-based Technique Survey
Graph-based Agent Memory systematically analyzes memory Extraction, Storage, Retrieval, and Evolution, mapping named systems like Mem0, Zep, GraphRAG, and MemGPT to each stage.
Libraries, Benchmarks, and Applications
Graph-based Agent Memory summarizes open-source libraries, benchmarks, and application scenarios, and identifies challenges and future directions for self-evolving graph-based agent memory systems.
RESULTS
Surveyed systems count
70+ systems
covers MemGPT, GraphRAG, Zep, HyperGraphRAG and many others
Lifecycle stages
4 stages
Extraction, Storage, Retrieval, Evolution
Memory categories
3 dimensions
Temporal, functional, structural taxonomy
Resource hub
1 GitHub repo
Awesome GraphMemory collection for papers, data, projects
Graph-based Agent Memory is a survey and framework paper, so its main result is unifying dozens of graph-based memory systems into a single lifecycle and taxonomy rather than reporting benchmark scores.
BENCHMARK
Graph-based Agent Memory is a survey and framework paper, so its main result is unifying dozens of graph-based memory systems into a single lifecycle and taxonomy rather than reporting benchmark scores.
BENCHMARK
Relative emphasis on different graph structures (knowledge, hierarchical, temporal, hypergraph, hybrid) in Graph-based Agent Memory's taxonomy.
KEY INSIGHT
Graph-based Agent Memory argues that even plain linear memory can be viewed as a degenerate graph, making graph-based memory a general framework rather than a niche structure.
This is counterintuitive because many builders treat graphs as heavyweight add-ons, but the paper reframes them as the underlying abstraction for almost all agent memory designs.
WHY IT MATTERS
Graph-based Agent Memory gives researchers a shared vocabulary and lifecycle to compare and compose memory systems across extraction, storage, retrieval, and evolution.
Builders can now systematically choose between knowledge graphs, temporal graphs, hypergraphs, and hybrid memories for specific agent tasks instead of improvising ad hoc buffers and vector stores.
Peilin Feng, Suorong Yang, Soujanya Poria
arXiv 2026 · 2026
Σ-Mem maintains historical competence evidence matrices M_p, a peer relationship matrix G, a memory readout r_{p,t}, and residual steering δ_{p,t} to guide multi-agent trust decisions. On the mixed counterfactual benchmark, Σ-Mem raises Qwen3-0.6B accuracy at CF@90 from 46.22% with the base model to 71.10% with Σ w/ G.
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.
Xiaohui Zhang, Zequn Sun et al.
· 2026
ActMem transforms dialogue history into atomic facts via Memory Fact Extraction, groups them with Fact Clustering, links them through a Memory KG Construction module, and uses Counterfactual-based Retrieval and Reasoning for action-aware answers. On ActMemEval, ActMem reaches 76.52% QA accuracy with DeepSeek-V3, beating LightMem’s 63.97% by 12.55 points and NaiveRAG’s 61.54%.
Answers use this explainer on Memory Papers.
Checking…