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

This survey presents graph-based personalized memory as a lifecycle for long-term LLM agents, including personal assistants such as OpenClaw and Hermes and specialized systems such as LinkedIn’s hiring agent. Its central idea is to replace flat dialogue logs, summaries, or vector stores with explicit user models whose nodes represent evidence, facts, abstractions, and domain entities, while typed edges encode temporal order, provenance, association, and task-specific relations. The taxonomy covers five graph patterns—flat, hierarchical, hypergraph, hybrid, and multiple disjoint graphs—and five evolution operations: admission, integration, conflict resolution, consolidation, and removal. It distinguishes temporal supersession, user correction, and contextual coexistence rather than applying a simplistic newest-wins policy. Retrieval is organized as a pipeline combining similarity-based candidate generation, structure-based expansion across relations and abstraction levels, and adaptive or agentic control for query interpretation, pruning, routing, and compression; examples include GAM, MAGMA, HingeMem, MRAgent, PRISM, and APEX-MEM. The survey emphasizes that graph structure alone does not guarantee better personalization: performance depends on meaningful relations, valid updates, and controlled context expansion. Evaluation remains dominated by downstream answer quality, using benchmarks such as LoCoMo, LongMemEval, PersonaMem-v2, RealMem, Engram-aBench, EvoMemBench, EvoArena, ActMemEval, ATM-Bench, and StructMemEval. Open problems include scalable lifelong memory, multimodal evidence, causal and counterfactual user modeling, provenance, user control, and direct evaluation of graph correctness and lifecycle reliability.