GSEM: Graph-based Self-Evolving Memory for Experience Augmented Clinical Reasoning

1 Introduction

• 70.90% and 69.24% average accuracy with DeepSeek-V3.2 and Qwen3.5-35B, respectively, as GSEM consistently outperforms three categories of strong baselines.The introduction also reports especially pronounced gains on treatment planning.Source: §1 · p. 2
• GSEM organizes experiences as indications and contraindications in a dual-layer graph encoding each experience’s decision structure and relations across experiences.Experience Reliability Validation calibrates initial quality before deployment; edge weights quantify inter-experience association strength.Source: §1 · p. 2
• Applicability-aware retrieval combines hybrid seed recall with quality- and relation-guided multi-seed traversal to find boundary-compatible experiences with coherent joint-use value.This targets failures from ignoring critical boundary conditions and co-retrieving conflicting experiences without joint-use value.Source: §1 · pp. 1–2, Figure 1 · p. 1
• After each task, GSEM updates node quality and edge weights from outcome feedback, refining reliability and relational structure without changing experience content.The weight-based calibration is motivated as a way to adapt deployment outcomes without destabilizing previously validated knowledge.Source: §1 · p. 2

2.1 Agent Memory

• Most agent-memory approaches store independent entries, leaving relations across experiences largely unmodeled.Examples include systems for selective reinforcement, fact consolidation, cross-task insight extraction, and workflow-level memory without parameter updates.Source: §2.1 · p. 2

2.2 Graph-based Agent Memory

• Existing graph-memory systems provide graph representations or graph-enhanced retrieval, but offer limited modeling of experience applicability boundaries or inter-experience conflicts.The review also notes graph augmentation for temporally grounded and multi-hop queries and hierarchical tracing across strategic and interaction-level memory.Source: §2.2 · p. 2

2.3 Self-Evolving Memory

• Self-evolving memory systems refine memories through task feedback, yet most operate on flat banks without explicit relations across experiences; GSEM addresses this gap with graph organization, applicability-aware retrieval, and online structural evolution.The reviewed approaches include reflection, strategy distillation, non-parametric reinforcement learning, generative memory, and healthcare-agent planning.Source: §2.3 · pp. 2–3

3.1 Task Definition

• For each task, the agent observes clinical context x_t, produces decision y_t using retrieved relevant experiences R_t, and updates memory after receiving scalar feedback.The decision function is parameterized by Θ.Source: §3.1 · p. 3

3.2 Experience Definition

• Each experience stores reusable decision knowledge as an applicable context, a decision strategy, a polarity, and a current reliability score.Polarity marks an indication associated with successful outcomes or a contraindication associated with failures to be avoided.Source: §3.2 · p. 3

3.3 Dual-layer Memory Graph

• The dual-layer graph represents typed decision entities and their internal flow within experiences, alongside directed relations between experiences weighted by association strength.Entity roles capture condition, constraint, action, rationale, and expected outcome; cross-layer mapping links each experience to its entity nodes.Source: §3.3 · p. 3

4 Methodology

• GSEM uses a three-stage pipeline in which construction builds the memory graph, retrieval selects task-relevant experiences, and evolution updates the graph for future retrieval.The architecture links the three stages in a feedback-driven refinement process.Source: §4 · p. 3, Figure 2 · p. 4

4.1 Memory Construction

• Successful trajectories yield Indication experiences, while failed trajectories are compared with similar successful ones to distill Contraindications; near-identical experiences are deduplicated.ERV initializes experience quality from accuracy on independent held-out trials, with scores above the baseline receiving higher values; the graph links experiences when their relation score exceeds a threshold.Source: §4.1 · pp. 3–4

4.2 Memory Retrieval

• Retrieval combines sparse entity-based and dense embedding-based candidates, reranks their union, and traverses from multiple seeds using association strength and node reliability.The agent can collect an experience, explore forward, backtrack to an ancestor layer, or stop; traversal has a maximum step limit.Source: §4.2 · p. 4

4.3 Self-Evolving Memory

• Task feedback updates the quality of activated experiences and weights of their induced edges, while retaining prior similarity structure and strengthening or down-weighting relations based on outcomes.New experiences can also be added with initialized quality and construction-time relation scoring, subject to the same edge threshold.Source: §4.3 · p. 5

5.1 Experimental Setup

• Experiments use two complementary benchmarks, shared-backbone RAG, memory, and self-evolving baselines, and Acc/R on MedR-Bench versus Pass@1 on MedAgentsBench.MedR-Bench testing holds out 329 diagnosis and 148 treatment cases; MedAgentsBench evaluates five subsets. Backbones were DeepSeek-V3.2 and Qwen3.5-35B-A3B, with five sampled reasoning trajectories per training instance.Source: §5.1 · pp. 5–6

5.2 Main Results

• GSEM reached 94.59% treatment accuracy vs. A-Mem’s 92.57% on MedR-Bench, and led MedAgentsBench averages with 70.90% using DeepSeek-V3.2 and 69.24% using Qwen3.5-35B.The main-results table covers MedR-Bench and MedAgentsBench under two backbones.Source: Table 1 · p. 6, §5.2 · p. 6
• GSEM outperformed self-evolving baselines on both MedR-Bench tasks, with gains in experience-intensive reasoning settings.Source: §5.2 · p. 6

5.3 Self-evolving Analysis

• After 50 online updates, errors fell from 19 to 9 for diagnosis and 8 to 4 for treatment, with gains persisting across 50, 150, and 250 updates.The update-count comparison is reported in Table 2.Source: §5.3 · pp. 6–7, Table 2 · p. 7
• Evolution shifts node quality toward experiences associated with correct decisions and strengthens or weakens relations according to their association with successful or failed outcomes.This makes both memory content and graph relations more performance-informed for later retrieval.Source: §5.3 · p. 7

6.1 Retrieval Ablation Study

• Removing entity-based recall mainly hurts condition-intensive settings; embedding-based recall causes the largest drops, especially on Treatment and MedBullets; single-seed traversal consistently degrades performance.Source: Table 3 · p. 7, §6.1 · p. 7

6.2 Reasoning Ability Analysis

• GSEM achieves the best Diagnosis/R, while GraphRAG has the highest Treatment/R but substantially lower treatment accuracy; GSEM performs consistently well on reasoning quality and task accuracy.Source: §6.2 · p. 7, Table 4 · p. 8

6.3 Effect of Model Size Analysis

• Performance is largely determined by the generator, with smaller-generator accuracy drops especially on Treatment; Qwen+DeepSeek performs best, with little difference from using a large retriever.Source: Table 5 · p. 8, §6.3 · p. 8

• GSEM represents clinical experiences as indications and contraindications in a dual-layer graph encoding decision structure and inter-experience relations.Source: §7 · p. 8
• Applicability-aware retrieval uses condition-compatible seed recall and graph traversal, while outcome feedback calibrates memory reliability and relational structure without changing experience content.Source: §7 · p. 8
• Experiments show GSEM outperforms three baseline categories, with particularly strong gains on multi-step treatment planning; future work includes richer feedback and longer-horizon real-world evaluation.Source: §7 · p. 8

• Public benchmarks with automatic protocols do not fully reflect real clinical workflows, interactive decisions, longitudinal follow-up, or deployment constraints.Source: Limitations · p. 9
• Memory construction requires multiple sampled trajectories and reliability validation, while graph growth may add maintenance and traversal latency.Source: Limitations · p. 9
• LLM-guided discrete traversal can be prompt-sensitive and nondeterministic in path selection; the authors call for more stable policies and stronger safety-critical constraints.Source: Limitations · p. 9

Ethical Consideration

• The work is not intended for direct clinical use; accuracy-based online feedback may miss safety-critical nuances, so real-world use requires human oversight and additional safeguards.Experiments use public benchmarks and collect no identifiable patient information.Source: Ethical Consideration · p. 9

A Benchmark Details

• MedR-Bench assesses multi-step clinical decisions, while MedAgentsBench evaluates complex medical question answering in a zero-shot Pass@1 setting.The benchmarks cover clinical cases and exam-style reasoning tasks, respectively.Source: A Benchmark Details · p. 11

B Baselines Details

• Naïve RAG combines BM25 and dense retrieval, while GraphRAG retrieves graph-informed subgraphs and linearizes them as generation context.Mem0 maintains conversational memories through embedding-based retrieval and memory-management operations.Source: B Baselines Details · pp. 11–12
• Mem0g, A-Mem, ReMe, and FLEX respectively use graph memories, linked atomic notes, refined procedural experiences, and incrementally updated experience libraries.These descriptions distinguish memory representations and update or retrieval strategies.Source: B Baselines Details · pp. 12–13

C.1 Hyperparameters

• The four edge-weight similarity coefficients are each 0.25, and traversal exposes 5 neighbor candidates per step for at most 60 steps.DeepSeek-V3.2 performs semantic similarity scoring.Source: §C.1 · p. 13

C.2 Similarity Calculation

• Experience-layer edge weights in Gexp are initialized from four similarity signals; cited passages describe entity and structural similarity, plus semantic similarity measuring joint-use value for co-retrieval and co-application.Entity similarity uses TF-IDF-weighted vectors and cosine similarity; structural similarity uses role-edge paths of lengths 1–4, with levels weighted proportional to path length.Source: §C.2 · p. 13

D Additional Experimental Results

• Tables 6 and 7 report efficiency, factuality, completeness, and average reasoning scores for MedR-Bench diagnosis and treatment tasks, respectively.The tables cover 329 diagnosis cases and 148 treatment cases and report mean±half-CI under two-sided 95% z-based confidence intervals.Source: D Additional Experimental Results · p. 13

E Dual-layer Memory Graph

• The additional graph material provides experience examples and visualizations of both the entity and experience layers.It is presented to facilitate intuitive understanding of the dual-layer memory graph.Source: E Dual-layer Memory Graph · p. 13

E.3 Experience Examples

• Each experience is an atomic memory unit comprising a clinical condition, decision content, polarity, and quality score.Polarity marks an indication associated with successful outcomes or a contraindication associated with failures; quality is refined through online task feedback.Source: Figure 6 · p. 14, §E.3 · p. 14

E.1 Visualization of the Entity-layer Graph

• Entity-layer nodes represent role-typed decision entities, and directed edges show reasoning flow within and across experiences.Node colors indicate entity types, including Condition, Constraint, Action, Rationale, and Outcome.Source: §E.1 · pp. 13–14, Figure 7 · p. 14

E.2 Visualization of the Experience-layer Graph

• Experience-layer nodes represent experiences, with node size encoding quality and weighted directed edges encoding inter-experience associations.The associations derive from entity overlap, structural similarity, synergy, and task consistency.Source: Figure 8 · p. 14, §E.2 · p. 14

F Case Study

• In the boundary-failure case, ReMe treats spontaneous regression as diagnostic conflict and recommends unnecessary systemic therapy, whereas GSEM recognizes the abscopal effect and defers therapy pending confirmation.The case’s ground truth is close follow-up while regression continues, without systemic therapy.Source: F Case Study · pp. 14–15, Table 8 · p. 16
• In the collaboration-failure case, ReMe misses the emergency pancreaticoduodenectomy decision, while GSEM’s complementary experiences support on-site surgery for irreparable pancreatic head injury.The case ground truth specifies emergency pancreaticoduodenectomy along with other procedures.Source: F Case Study · p. 15, Table 9 · p. 16

G.1 Indication Experience Extraction

• Tables 10 and 11 contain system and human prompts for extracting indication experiences from successful trajectories.Source: §G.1 · p. 15

G.2 Contraindication Experience Extraction — Divergence Analysis

• Tables 12 and 13 contain prompts for identifying fatal divergence points between successful and failed trajectories.Source: §G.2 · p. 15

G.3 Contraindication Experience Extraction

• Tables 14 and 15 contain prompts for extracting contraindication experiences from divergence analysis.Source: §G.3 · p. 15

G.4 Entity Extraction

• Tables 16 and 17 contain prompts for extracting decision-structure entities from each experience.Source: §G.4 · p. 15

G.5 Role-Edge Structure Extraction

• Tables 18 and 19 contain prompts for extracting the role-edge reasoning skeleton from each experience.Source: §G.5 · p. 15

G.6 Semantic Similarity Evaluation

• Tables 20 and 21 contain prompts for evaluating semantic similarity between experience pairs for edge-weight computation.Source: §G.6 · p. 15