Graph-based Agent Memory

Updated 4 July 2026

Graph-based agent memory is a structural paradigm that represents memory as dynamic graphs with nodes storing events, entities, and text and edges encoding various relations.
It enables efficient short-term and long-term storage by distinguishing among semantic, episodic, and experiential data through diverse graph schemas.
It supports targeted retrieval and reasoning via graph traversal, reinforcement learning-based navigation, and semantic reranking, proving effective across multiple application domains.

Graph-based agent memory is a structural memory paradigm for LLM-based agents in which memory at time $t$ is represented as a graph $G_t=(V_t,E_t,A_t)$ whose nodes store entities, events, text chunks, or summaries, and whose edges encode semantic, temporal, causal, or associative relations (Yang et al., 5 Feb 2026). Within the survey taxonomy, it spans short-term and long-term memory, knowledge and experience memory, and a range of implementations including knowledge graphs, temporal knowledge graphs, hierarchical graphs, hypergraphs, and hybrid designs (Yang et al., 5 Feb 2026). The central premise is that explicit graph structure can preserve relational dependencies, hierarchical organization, provenance, and retrieval paths more effectively than non-structural memories such as linear buffers or flat vector stores (Yang et al., 5 Feb 2026).

1. Taxonomy and conceptual scope

Graph-based agent memory is usually situated against three axes. The first is temporal scope: short-term memory is a volatile, limited-capacity store of immediately relevant information, whereas long-term memory is a persistent store of facts, experiences, and user-specific patterns across sessions (Yang et al., 5 Feb 2026). The second is functional role: knowledge memory stores general world knowledge, ontologies, or procedural rules, while experience memory logs an agent’s own interactions, observations, rewards, and user preferences (Yang et al., 5 Feb 2026). The third is representational structure: non-structural memory relies on token windows, vector stores, or key–value logs, whereas structural memory makes nodes and edges first-class carriers of meaning (Yang et al., 5 Feb 2026).

This taxonomy is reflected directly in recent systems. AriGraph partitions memory into semantic vertices and episodic vertices, with semantic edges for triples and episodic hyperedges linking an episodic vertex to all semantic edges extracted at a time step (Anokhin et al., 2024). GAM separates a dynamic Event Progression Graph from a Topic Associative Network and archives earlier event graphs after consolidation (Wu et al., 14 Apr 2026). Zep’s Graphiti organizes memory into episodic, semantic entity, and community subgraphs, with temporal validity encoded on semantic edges (Rasmussen et al., 20 Jan 2025). MAGMA represents each memory item across orthogonal semantic, temporal, causal, and entity graphs (Jiang et al., 6 Jan 2026). GAAMA constructs a typed hierarchy over episode, fact, reflection, and concept nodes (Paul et al., 29 Mar 2026).

A consistent implication across these formulations is that “graph-based agent memory” is not a single schema but a family of memory substrates whose common property is explicit relational organization. The survey’s formal definition is deliberately broad: $V_t$ contains memory nodes, $E_t\subseteq V_t\times V_t$ contains relations, and $A_t$ assigns text, embeddings, timestamps, relation types, or confidence scores to nodes and edges (Yang et al., 5 Feb 2026).

2. Architectural patterns and graph schemas

Recent work exhibits several recurring graph schemas. Some systems are semantic–episodic bipartitions. AriGraph’s long-term memory is a directed, attributed graph $G_t=(V_t,E_t,A_t)$ with $V_t=V_s^t\cup V_e^t$ and $E_t=E_s^t\cup E_e^t$ , where semantic memory stores facts and episodic memory stores when those facts were observed (Anokhin et al., 2024). This design keeps fact storage modular while enabling episode ranking by the relevance of incident semantic edges (Anokhin et al., 2024).

Other systems are hierarchical short-term/long-term decompositions. GAM defines memory as

$H_t=\{G_{\text{topic}}^{(t)},G_{\text{event}}^{(t)},S_{\text{arch}}^{(t)},E_{\text{cross}}^{(t)}\},$

with a small dynamic Event Progression Graph for current dialogue, a sparse Topic Associative Network for long-term topics, archived event graphs, and cross-links from topic nodes to the event graphs that grounded them (Wu et al., 14 Apr 2026). Hybr-Mem couples a temporal-semantic tree with a knowledge graph so that short-term events evolve into long-term summaries while entities and relations remain globally traversable (Yu et al., 15 May 2026).

A third family is multi-view or orthogonal graph memory. MAGMA maintains four relation-specific graphs over the same event-node set: semantic, temporal, causal, and entity views (Jiang et al., 6 Jan 2026). This disentangles relations that would otherwise be entangled in a monolithic store. A plausible implication is that this separation improves alignment between query intent and retrieved evidence when temporal and causal cues diverge.

A fourth family is typed heterogeneous knowledge graphs with abstraction layers. GAAMA stores verbatim episodes, atomic facts, topic-level concepts, and higher-order reflections in one graph, with five structural edge types connecting them (Paul et al., 29 Mar 2026). MRAgent uses a Cue-Tag-Content graph in which associative tags act as semantic bridges between fine-grained cues and episodic, semantic, or topic-level contents (Ji et al., 4 Jun 2026). CompassMem builds an Event Graph whose nodes are event tuples with raw spans, temporal metadata, summaries, and participants, and whose directed labeled edges encode logical relations such as causal, temporal_before, motivation, and part_of (Hu et al., 8 Jan 2026).

A fifth family emphasizes versioning and formal provenance. Kumiho treats every belief or memory unit as a new immutable Revision node, with mutable tag pointers, typed dependency edges, and URI-based addressing in a Neo4j long-term graph plus Redis working memory (Park, 18 Mar 2026). This makes revision itself a graph operation rather than an overwrite.

System	Core graph organization	Reported emphasis
GAM (Wu et al., 14 Apr 2026)	Event Progression Graph + Topic Associative Network + archives	Decoupled encoding and consolidation
AriGraph (Anokhin et al., 2024)	Semantic vertices/edges + episodic vertices/hyperedges	Integrated semantic and episodic memory
Graphiti/Zep (Rasmussen et al., 20 Jan 2025)	Episodic, semantic entity, and community subgraphs	Bi-temporal knowledge graph
MAGMA (Jiang et al., 6 Jan 2026)	Semantic, temporal, causal, and entity graphs	Policy-guided multi-view retrieval
GAAMA (Paul et al., 29 Mar 2026)	Episode, fact, reflection, and concept nodes	Hierarchical associative memory
MRAgent (Ji et al., 4 Jun 2026)	Cue-Tag-Content heterogeneous graph	Active reconstruction rather than static retrieval

These schemas differ in granularity and update policy, but they converge on two recurrent design decisions: explicit separation of rapidly changing interaction traces from more stable abstractions, and typed traversal paths that can be inspected, reranked, or constrained.

3. Extraction, consolidation, and memory evolution

The survey frames the memory lifecycle as Extraction, Storage, Retrieval, and Evolution (Yang et al., 5 Feb 2026). Extraction maps raw data to graph deltas,

$f_{\text{extract}}:\text{RawData}\rightarrow (\Delta V,\Delta E,\Delta A),$

through entity and relation extraction, summarization, event segmentation, embedding assignment, and multimodal captioning when applicable (Yang et al., 5 Feb 2026). For text, this usually means transforming utterances or documents into triples, summaries, or event units; for trajectories, it includes timestamped state or event nodes (Yang et al., 5 Feb 2026).

A central challenge is how to update the graph without contaminating stable memory. GAM addresses this by treating online updates as a boundary-triggered policy. A buffered event graph is consolidated only when a semantic boundary is detected, approximating

$G_t=(V_t,E_t,A_t)$ 0

and then atomically summarizing the buffer, linking the new topic node to candidate topic nodes, archiving the event graph, and resetting the buffer (Wu et al., 14 Apr 2026). The reported motivation is to minimize interference while preserving long-term consistency (Wu et al., 14 Apr 2026).

AriGraph performs per-step graph revision rather than buffered consolidation. Each observation yields new semantic triplets, identifies outdated facts among incident edges, removes contradicted edges, adds semantic nodes and edges, and appends an episodic vertex linked to the extracted semantic edges (Anokhin et al., 2024). Zep’s Graphiti makes temporal validity explicit: if a new semantic edge conflicts with an existing edge over an overlapping interval, the old edge’s invalid time is set to the new edge’s validity time (Rasmussen et al., 20 Jan 2025). This turns contradiction handling into an update over temporal intervals rather than a simple overwrite.

Hierarchical summarization is another common evolution mechanism. Hybr-Mem groups new memory events into temporal windows at level $G_t=(V_t,E_t,A_t)$ 1, checks semantic similarity against threshold $G_t=(V_t,E_t,A_t)$ 2, and recursively creates parent summaries at higher levels with larger time windows $G_t=(V_t,E_t,A_t)$ 3 (Yu et al., 15 May 2026). GAAMA similarly starts from verbatim episode preservation, then extracts facts and concepts, and finally synthesizes reflections (Paul et al., 29 Mar 2026). These systems treat abstraction as a graph-building operation rather than an external compression step.

Kumiho goes further by grounding revision in formal belief revision semantics. Revision creates a new revision node, adds a Supersedes edge, and redirects the current tag; contraction removes tags from revisions whose propositional content contains the contracted belief and soft-deprecates the corresponding item (Park, 18 Mar 2026). The paper proves satisfaction of the basic AGM postulates $G_t=(V_t,E_t,A_t)$ 4– $G_t=(V_t,E_t,A_t)$ 5 and Hansson’s Relevance and Core-Retainment, while deliberately violating Recovery (Park, 18 Mar 2026). This suggests that graph-based memory can be treated not only as data organization but as an operational semantics for belief change.

Graph-based agent memory departs most sharply from flat memory at retrieval time. The survey distinguishes similarity-based retrieval, rule-based and temporal filters, graph traversal, RL-based or agentic retrieval, multi-round retrieval, and hybrid-source retrieval (Yang et al., 5 Feb 2026). In graph memory, retrieval generally begins with anchors—entities, cues, topics, or semantically similar nodes—and then expands structurally.

GAM’s retrieval is explicitly graph-guided and multi-factor. It first selects top- $G_t=(V_t,E_t,A_t)$ 6 topic nodes by vector similarity, expands to their neighbors in the topic graph, drills down through cross-links into archived event graphs, and finally reranks candidate memory units with a score

$G_t=(V_t,E_t,A_t)$ 7

using $G_t=(V_t,E_t,A_t)$ 8, $G_t=(V_t,E_t,A_t)$ 9, and $V_t$ 0 (Wu et al., 14 Apr 2026). The stated purpose is to balance recency, reliability, and interlocutor specificity without over-promoting lexical matches (Wu et al., 14 Apr 2026).

MAGMA formalizes retrieval as an MDP over the combined multi-graph. The state includes the current frontier, visited set, query embedding, and query intent; actions choose neighbors; rewards are transition scores that combine structural alignment and semantic relevance; and the implemented traversal is a heuristic beam search that returns a serialized subgraph under a fixed budget (Jiang et al., 6 Jan 2026). This is a direct instantiation of the survey’s “RL-based and agentic retrieval” category (Yang et al., 5 Feb 2026).

GAAMA combines cosine-similarity $V_t$ 1-nearest-neighbor search with edge-type-aware Personalized PageRank. Seed nodes are retrieved semantically, graph expansion proceeds to depth $V_t$ 2, outgoing edge weights are damped for high-degree nodes, and the final ranking is

$V_t$ 3

with $V_t$ 4 and $V_t$ 5 (Paul et al., 29 Mar 2026). The reported goal is to complement semantic similarity with concept-mediated traversal paths while avoiding hub-dominated retrieval (Paul et al., 29 Mar 2026).

Several systems push retrieval toward active reasoning. CompassMem decomposes a query into subgoals, launches parallel explorers over an Event Graph, and prioritizes frontier nodes according to unsatisfied subgoals rather than static similarity alone (Hu et al., 8 Jan 2026). MRAgent treats memory access as active reconstruction: it initializes from extracted cues, interleaves traversal actions with LLM-based select and route decisions, accumulates evidence iteratively, and stops when enough evidence has been reconstructed (Ji et al., 4 Jun 2026). The paper’s complexity argument contrasts blind $V_t$ 6 expansion with an actively pruned process requiring only $V_t$ 7 memory-tool calls under capped $V_t$ 8 and $V_t$ 9 (Ji et al., 4 Jun 2026).

Embodied systems expose retrieval primitives as tools. eMEM provides ten recall tools, including semantic search, spatial query, temporal query, episode summary, entity query, locate, and recall, over a unified graph backed by SQLite, HNSW, and R-tree indices (Rasheed et al., 2 Jun 2026). EmbodiedLGR similarly exposes semantic, positional, and temporal searches over object nodes and uses a vector database only for semantically complex questions (Riva et al., 20 Apr 2026). This suggests that in embodied settings, graph memory often becomes a tool interface rather than only an internal data structure.

5. Empirical results and application domains

The survey lists application scenarios for graph-based memory in conversational assistants, code agents, recommender systems, financial agents, game and embodied agents, medical agents, and scientific discovery (Yang et al., 5 Feb 2026). Benchmarking spans dialogue and personalization datasets such as LoCoMo and LongMemEval, long-context QA, web and embodied tasks, and continual settings (Yang et al., 5 Feb 2026). The reported metrics are correspondingly heterogeneous: token-level F1, BLEU-1, EM, judge accuracy, recall and relevance, latency, and task success (Yang et al., 5 Feb 2026).

On long-conversation QA, GAM reports that on LoCoMo it achieves the highest average F1 across all four backbones, including 40.00 vs. 35.38 for Mem0 on Qwen2.5-7B, and especially improves Temporal reasoning by +18% F1 over Mem0 (Wu et al., 14 Apr 2026). On LongDialQA, it reports 12.55 F1 vs. 10.27 for Mem0 on Qwen2.5-7B, a +22% gain (Wu et al., 14 Apr 2026). Its ablations also show that removing the Event Progression Graph drops F1 from 40.00 to 25.06, while removing the state-switch mechanism yields 32.58, indicating that both buffering and decoupled consolidation are central to the reported gains (Wu et al., 14 Apr 2026).

GAAMA evaluates on LoCoMo-10 and reports 78.9% mean reward, outperforming a tuned RAG baseline at 75.0%, HippoRAG at 69.9%, A-Mem at 47.2%, and Nemori at 52.1% (Paul et al., 29 Mar 2026). The same study reports an ablation from 78.0% to 78.9% when augmenting semantic search with Personalized PageRank (Paul et al., 29 Mar 2026). CompassMem reports on LoCoMo with GPT-4o-mini an average F1 of 52.18, above HippoRAG’s 47.92 and Mem0’s 45.10, and on NarrativeQA raises F1 from 33.55% for the best baseline to 39.04% (Hu et al., 8 Jan 2026).

Several systems emphasize efficiency as well as accuracy. Zep reports 94.8% vs. 93.4% for MemGPT on the DMR benchmark, and on LongMemEval reports 63.8% accuracy with 3.20 s latency and 1.6 k tokens for Zep with gpt-4o-mini, compared with 55.4% accuracy, 31.3 s latency, and 115 k tokens for full context (Rasmussen et al., 20 Jan 2025). MAGMA reports a LoCoMo judge score of 0.700 and LongMemEval average of 61.2%, with token consumption reduced by over 95% on LongMemEval from $E_t\subseteq V_t\times V_t$ 0 K to $E_t\subseteq V_t\times V_t$ 1– $E_t\subseteq V_t\times V_t$ 2 K tokens (Jiang et al., 6 Jan 2026). MRAgent reports on LoCoMo an LLM-Judge score of 84.21 with Gemini and 88.32 with Claude, and on LongMemEval an overall judge score of 72.95, while using 118 k tokens and 586 s runtime per sample on the Gemini backbone versus 245 k and 533 s for Mem0 (Ji et al., 4 Jun 2026).

Formal or cognitively motivated systems also report strong benchmark results. Kumiho achieves 0.565 overall F1 on LoCoMo including 97.5% adversarial refusal accuracy, and 93.3% judge accuracy on LoCoMo-Plus with 98.5% recall accuracy (Park, 18 Mar 2026). eMEM, designed for embodied agents, scores 80.8 weighted mean over 988 probes on eMEM-Bench v1 and reports that a pure RAG ablation loses 30 pt on context dependent retrieval and 29 pt on DRM lure rejection (Rasheed et al., 2 Jun 2026). AriGraph reports that it solves hard Treasure Hunt settings in TextWorld that baselines fail to complete and is on par with humans in Treasure Hunt, behind in Cleaning, and ahead in Cooking (Anokhin et al., 2024). EmbodiedLGR reports on NaVQA that Florence-2-large with Graph + ReMEmbR reaches 33.90 average accuracy and that graph-only latency is 9.97 s, compared with about 19.79 s for ReMEmbR alone (Riva et al., 20 Apr 2026).

Taken together, these results do not establish a single universal best architecture, because benchmarks, metrics, and backbones differ. They do show that graph structure is repeatedly associated with improvements in long-horizon recall, temporal reasoning, multi-hop evidence recovery, or efficiency under constrained context windows.

6. Formal semantics, security, and open problems

Graph-based memory has increasingly explicit theoretical and operational commitments. Kumiho is the clearest formal example: it maps graph operations to AGM-style revision and Hansson-style belief base contraction, proving satisfaction of $E_t\subseteq V_t\times V_t$ 3– $E_t\subseteq V_t\times V_t$ 4, Relevance, and Core-Retainment, while rejecting Recovery in favor of immutable provenance (Park, 18 Mar 2026). SAGE contributes a different kind of formalization, treating memory as a dynamic graph jointly optimized by a writer and a Graph Foundation Model-based reader, with signal-to-noise and stability analyses for self-evolution under controlled graph drift (Wang et al., 12 May 2026). This suggests that the field is moving beyond ad hoc memory engineering toward explicit semantics and optimization objectives.

At the same time, graph structure creates new attack surfaces. ShadowMerge models the agent’s long-term memory as a directed typed graph and shows that a poisoned relation can be inserted on the same canonicalized relation channel as benign evidence while carrying a conflicting value (Luo et al., 9 May 2026). The attack reports 93.8% average attack success rate, improving the best baseline by 50.3 absolute points, with benign utility of about 97.3%, and reports that representative input-side defenses are insufficient (Luo et al., 9 May 2026). The proposed defense direction is graph-aware rather than text-only: provenance checks, writer authorization policies per anchor–predicate channel, automatic conflict resolution, and retrieval-time consistency screening (Luo et al., 9 May 2026).

Shared or multi-agent memory extends the problem further. HyphaeDB reinterprets HNSW not as a passive search index but as a communication fabric, in which agents are nodes with persistent positions and knowledge propagates through gossip with energy-based attenuation (Halaharvi, 27 Jun 2026). MemGraphRAG similarly uses a hierarchical global memory and a society of agents to resolve conflicts and maintain connectivity during graph construction, reporting 59.3% overall QA accuracy and 0.06 s retrieval latency on its evaluation suite (Wu et al., 30 May 2026). A plausible implication is that as graph memory becomes shared infrastructure rather than per-agent storage, issues of conflict adjudication, trust, and access control become inseparable from retrieval quality.

The survey identifies the broader research agenda as graph quality metrics, scalability and efficiency, privacy and security, dynamic schema learning, interpretability and trust, theoretical foundations, and multi-agent memory coordination (Yang et al., 5 Feb 2026). Individual systems sharpen these points. GAM lists reliance on LLM prompts for boundary detection and relation scoring, unimodality, privacy filters, and adaptive consolidation thresholds as future directions (Wu et al., 14 Apr 2026). MRAgent notes static graph growth and the lack of forgetting (Ji et al., 4 Jun 2026). eMEM identifies HNSW insertion cost and R-tree rebuild cost as scaling pressures in long runs (Rasheed et al., 2 Jun 2026). Across the literature, the open question is not whether graphs can store memory, but how they should evolve, be queried, be revised, and be secured when memory becomes a persistent substrate for long-horizon agency.