PaperlayerView exact revision on arXiv

Source-linked AI summary

Graph-Based Personalized Memory for LLM Agents: Representation, Evolution, Retrieval, and Evaluation

Dac Duy Anh Nguyen, Zhangchi Qiu, Shigeng Chen, Alan Wee-Chung Liew

arXiv:2609.08599v1cs.AI

TL;DR

Long-term personalized agents need memory that captures evolving user information, but graph-based personalized-memory research remains fragmented. This survey organizes the field across representation, evolution, retrieval, and evaluation to clarify design choices and open challenges.

  • Problem

    As LLM agents become long-term personal assistants, they need to represent users’ goals, preferences, constraints, relationships, and prior experiences, while existing graph-memory work remains fragmented.

  • Method

    The survey develops a lifecycle-oriented taxonomy covering graph-based memory representation, evidence evolution, personalized retrieval, and evaluation, while reviewing recent methods and design choices.

  • Results

    The survey synthesizes current benchmarks and metrics, including evaluations of conversational recall, personalization, memory evolution, graph retrieval, reasoning support, multimodal evidence, and structural organization.

  • Takeaways & Limitations

    Graph-based memory can connect scattered user evidence, support multi-hop context retrieval, trace profile claims to interactions, and track changing preferences and circumstances.

Abstract

from arXiv · show

Large Language Model (LLM) agents are evolving from single-session tools toward long-term personal assistants that must adapt to individual users across tasks, contexts, and interactions. This shift makes memory a core requirement for personalization, since user preferences, goals, constraints, relationships, and past experiences are accumulated gradually and often change over time. Graph-based personalized memory provides a structured way to model such user information through explicit relations, temporal context, and evidence links. Such representations can model not only what an agent remembers about a user but also how memories are connected, revised, and retrieved to support personalized decisions. However, existing work remains fragmented across personalized agents and generic graph memory frameworks, making it difficult to understand the design space as a whole. This survey develops a lifecycle-oriented view of graph-based personalized memory for LLM agents. We organize existing studies around memory representation, memory evolution, memory retrieval, and memory evaluation. We further compare key design choices, discuss current evaluation practices, and open challenges in building reliable long-term personalized agents. This survey aims to clarify how graph-based memory can support adaptive, controllable, and user-centric LLM agents.

I. INTRODUCTION

As LLM agents become long-term personal assistants, they need persistent memory that captures changing user-specific information and its relationships. This survey organizes graph-based personalized memory across its lifecycle and reviews current evaluation and future directions.

  • Persistent memory is needed because full interaction histories are impractical to include in every prompt, while user goals, preferences, constraints, and experiences change across sessions.
  • Graph-based memory connects user facts, events, and relations, enabling multi-hop context, evidence tracing, and tracking of evolving preferences and circumstances.
  • Existing research is fragmented across personalized-agent and graph-memory surveys, which use different primary lenses for understanding persistent user memory.
  • The survey’s lifecycle view covers graph construction, evidence updating and consolidation, personalized retrieval, and memory-quality evaluation.
  • The survey contributes a taxonomy, evaluation review, timely method analysis, and open research directions for graph-based personalized memory.

II. PRELIMINARIES

Graph-based personalized memory is persistent, user-specific state stored outside model parameters and organized as an explicit graph. Unlike logs or summaries, it makes remembered items, relations, evidence, and revisions explicit and queryable.

  • Personalized memory stores user-specific facts, preferences, goals, constraints, relationships, and experiences across sessions as explicit graph-structured state.
  • Graph memory represents remembered items as nodes and their relations as edges, making connections, supporting evidence, validity, and revisions queryable.

B. Personal Memory Dimensions

Personal memory spans multiple content dimensions rather than a flat list of traits, and individual memories may occupy several dimensions simultaneously. Graph representations encode these memories through roles, relations, structures, and lifecycle metadata.

  • Personal Memory Dimensions: Personal memory includes facts, preferences, goals, constraints, episodes, feedback, emotions, and corrections to earlier claims beyond the current context window.
  • Personal Memory Dimensions: Systems use flat, hierarchical, hypergraph, hybrid, and multi-graph structures, shaping evidence traceability, state validity, and retrievable granularity.
  • Personal Memory Dimensions: Graph-memory systems represent information through evidence, fact, abstraction, and domain-specific node roles that may overlap within one graph.
  • Personal Memory Dimensions: Common relations encode temporal order, provenance, associations, and domain-specific semantics, while timestamps and validity markers distinguish active state from historical evidence.

B. Memory Structures

The survey organizes graph-based memory into five non-mutually-exclusive patterns, with flat and hierarchical graphs providing distinct ways to structure user information. Hierarchies improve traceability across abstraction levels but introduce consistency risks as depth increases.

  • Five non-exclusive patterns organize graph-based memory: flat, hierarchical, hypergraph, hybrid, and multiple disjoint graphs.
  • Flat Graph: Flat graphs store user facts, preferences, and typed relations without explicit abstraction layers.They can include heterogeneous schemas, provenance, and validity metadata, but do not inherently separate evidence, facts, and abstractions.
  • Hierarchical Graph: Hierarchical graphs expose user state at multiple specificity levels through evidence, intermediate, abstraction, or granularity-preserving layers.Examples retain links among turns, episodes, facts, concepts, personas, sessions, and archived events.
  • Hierarchical Graph: 44? Hierarchical links support traceability between raw evidence and higher-level representations but require cross-layer consistency as depth grows.Personas and reflections may overgeneralize transient evidence or propagate extraction errors, while cross-level links enable checking abstractions against lower-level evidence.

3) Hypergraph:

The supplied passages describe hypergraph and hybrid or multiple-graph designs as alternatives for preserving context and distributing memory across specialized structures. These designs improve organization for particular use cases but introduce coordination and generality questions.

  • Hypergraph: Hypergraphs preserve joint context by grouping connected dialogue, topic, episode, or fact elements into hyperedges.HingeMem emphasizes boundary-triggered segments, whereas HyperMem organizes topic, episode, and fact levels.
  • Hybrid Graph: Hybrid systems combine graph topology with trees, summaries, passages, or evidence buffers that retain long-form, temporal, or source-text information.The graph provides relational access while complementary substrates store content outside the graph.
  • Hybrid Graph: Lightweight graph indexes in hybrid designs separate relational indexing from richer evidence stores but require consistency across components.LiCoMemory and MemWeaver use graphs alongside session summaries, dialogue chunks, experience stores, and passage stores.
  • Multiple Disjoint Graphs: Multiple-graph evidence is limited to DEMENTIA-PLAN, which separates daily routine context from longer-term autobiographical identity.The cited work does not establish whether the graph instances are formally disjoint or whether this pattern generalizes.
  • Multiple Disjoint Graphs: Multiple specialized graphs may support efficient storage and retrieval when the relevant memory location is known in advance.

C. Discussion

Representation determines what evidence remains available to later memory operations, while evolution maintains an active user model through admission, integration, conflict handling, consolidation, and removal. These mechanisms provide structure but do not guarantee reliable personalization.

  • Representation determines which user evidence is retained and which dependencies later lifecycle operations can use.Evidence-rich designs support provenance and correction, whereas compact fact- or abstraction-centered designs may discard context or propagate extraction errors.
  • Memory evolution maintains a usable active user model while preserving historical evidence through five operations: admission, integration, conflict resolution, consolidation, and removal.
  • Admission: Admission filters incoming evidence into persistent memory and assigns functional types or granularity.Examples retain persona-relevant content, facts, preferences, or roles such as facts, procedural instructions, and constraints.

B. Memory Integration

Memory integration and later evolution attach, reconcile, consolidate, forget, or delete records while preserving evidence where possible. Removing source evidence also raises dependency-propagation requirements for derived user representations.

  • Memory Integration: Integration attaches new evidence to existing graph structure, preserves provenance, and can reduce redundancy through node reuse, edge invalidation, or buffered consolidation.GAM buffers utterances before consolidating semantically complete units to reduce semantic drift in the main graph.
  • Conflict Resolution: Conflict resolution distinguishes temporal supersession, user correction, and contextual coexistence rather than applying a single newest-wins rule.Systems may overwrite values, mark them inactive while retaining evidence, or preserve successive queryable versions.
  • Consolidation: Consolidation abstracts, deduplicates, restructures, and archives accumulated memory after multiple writes.GAM promotes completed event buffers into topic structures, while All-Mem applies gated topology edits while preserving immutable evidence.
  • Memory Removal: Forgetting reduces exposure, resolution, weight, or retrieval likelihood while retaining some representation, whereas deletion changes what remains stored, active, or recoverable.ScrapMem progressively reduces older detail, while Kumiho retains immutable revisions and excludes deprecated memories from normal retrieval.
  • Memory Removal: Removing source evidence may require revising derived facts, summaries, personas, and abstractions to prevent internal inconsistency.The reliability and generality of dependency propagation remain open evaluation questions.

F. Discussion

Graph-based memory systems coordinate evidence across timescales and retrieval stages, but the survey presents these designs as a fragmented space without one universal preference.

  • F. Discussion: Memory evolution links incoming evidence, structural reorganization, and active-state selection across different timescales.GAM buffers events before consolidation, All-Mem edits topology while retaining immutable evidence, and AriadneMem and Kumiho represent temporal revisions differently.
  • F. Discussion: Retrieval combines similarity-based candidate generation, structure-based evidence expansion, and adaptive control over interpretation, exploration, or compression.These mechanisms map requests to bounded subsets of persistent memory and can be combined within one retrieval pipeline.
  • F. Discussion: Similarity matching commonly anchors queries to graph candidates, but flat top-k selection can return redundant fragments or miss connected evidence.EMG-RAG applies similarity selection to an editable graph of smartphone memories for personalized tasks.

B. Structure-Based Retrieval

Structure-based retrieval expands initial candidates through relations, abstraction levels, and temporal or relational views, while adaptive control limits exploration and context cost. The survey cautions that graph structure helps only when representations, evolution, and retrieval preserve relevant evidence.

  • B. Structure-Based Retrieval: Structural retrieval connects requests to distributed user evidence across relations, abstraction levels, or temporal views.GAM retrieves topic nodes before accessing archived event graphs, while MAGMA selects among semantic, temporal, causal, and entity views.
  • B. Structure-Based Retrieval: Associative retrieval combines relevance, importance, and temporal alignment, while complementary-group expansion can recover source segments when more evidence is needed.AssoMem anchors utterances to extracted clues, and xMemory expands selected components to source messages.
  • B. Structure-Based Retrieval: Adaptive retrieval changes query interpretation, graph exploration, stopping, routing, or compression according to the request and intermediate evidence.MRAgent prunes retrieval paths, PRISM compresses evidence under a context budget, HingeMem controls route depth, and APEX-MEM resolves evolving information at retrieval time.
  • B. Structure-Based Retrieval: Similarity anchoring, structural expansion, and adaptive control can form a pipeline that selects and compresses relevant graph evidence under a context window.The mechanisms respectively identify entry points, recover distributed evidence, and regulate the final result.
  • B. Structure-Based Retrieval: Controlled comparisons show that graph structure does not automatically improve retrieval because outcomes can depend on foundational system settings.Useful graph retrieval requires meaningful user relations, valid evolution, and relevant evidence selection.

VI. EVALUATION

Personalized-memory evaluation mainly tests downstream answer quality, evidence retrieval, user-state modeling, evolution, organization, and efficiency across benchmarks. Long-horizon conversational recall remains dominant, while newer benchmarks target dynamic personas, evolving knowledge, and graph-specific structure.

  • VI. EVALUATION: Long-horizon conversational recall remains the dominant benchmark family, with LoCoMo and LongMemEval testing multi-session histories, updates, temporal reasoning, and abstention.These benchmarks force retrieval under multi-session noise but primarily assess final-answer correctness rather than graph correctness.
  • VI. EVALUATION: Personalization benchmarks test dynamic profiles, preference tracking, implicit personas, longer contexts, and evolving project goals across sessions.PersonaMem, PersonaMem-v2, and RealMem shift evaluation from factual recall toward user-state modeling.
  • VI. EVALUATION: Graph-focused benchmarks test structured retrieval, self-evolving memory, persistent environmental change, and preference-evolution chains.EngramaBench, EvoMemBench, and EvoArena isolate structure and evolution requirements that general recall benchmarks may not expose.
  • VI. EVALUATION: Reported metrics span answer quality, evidence retrieval, user-state modeling, memory evolution and structure, and efficiency.Examples include F1 and task success, Recall@k and evidence-location accuracy, persona consistency, update accuracy, latency, and token usage.

VII. CHALLENGES AND FUTURE DIRECTIONS

Future graph-based personalized memory must scale across years of interaction while preserving historical context, adapting to preference drift, and maintaining accurate, efficient retrieval. The survey also calls for multimodal, causal, and counterfactual user modeling.

  • VII. CHALLENGES AND FUTURE DIRECTIONS: Lifelong graph memories should grow with users, consolidate recurring evidence, preserve historical context, adapt to preference drift, and retain accurate, efficient retrieval.The challenge is to support years of interaction rather than isolated sessions or short benchmark histories.
  • VII. CHALLENGES AND FUTURE DIRECTIONS: Future multimodal memory graphs should connect speech, images, documents, locations, actions, and interaction traces while preserving modality, reliability, and temporal context.Current systems mostly construct memory from text.
  • VII. CHALLENGES AND FUTURE DIRECTIONS: Causal and counterfactual user modeling should distinguish stable traits from temporary constraints or accidental observations and represent why preferences change.Most existing memories are associative records of what users said or did.
Loading 2609.08599v1…