Experience Makes Skillful: Enabling Generalizable Medical Agent Reasoning via Self-Evolving Skill Memory

• Static, single-turn benchmarks do not capture clinical decision making, which requires multistep interaction, evidence gathering, revision, and action adjustment under uncertainty.Source: §1 · p. 1
• Raw case memories are often redundant, noisy, and instance-specific, while existing skill systems may require parameter updates and lack long-term usefulness assessment.Source: §1 · p. 2
• SkeMex converts informative interaction trajectories into structured, reusable skills organized across general reasoning, task-specific knowledge, and action-level operations.Source: §1 · p. 2
• Context-dependent utility is estimated from clinical feedback and guides both value-aware skill retrieval and repository governance as a non-parametric reinforcement process.Source: §1 · p. 2
• The closed-loop Read–Write–Assess–Govern lifecycle writes and updates skills, promotes high-utility memories, merges redundancies, and removes low-quality or potentially harmful entries.Source: §1 · p. 2
• SkeMex consistently improves performance across diverse clinical tasks, generalizes across model backbones, and supports transferable skill memory across heterogeneous task settings.Source: §1 · p. 2

LLM-based Medical Agents

• Medical agents combine retrieval, tool use, multimodal specialization, and coordinated workflows, but often process cases independently without reusable cross-task experience.Prior systems include i-MedRAG, MedRAG [90], EHRAgent [51], MMedAgent [28], MedAgents [59], MDAgents [22], MAM [94], and MedAgent-Pro.Source: §2 · p. 3

Self-Evolving Memory

• SkeMex decouples memory evolution from fixed workflows and model training by distilling trajectory skills, estimating utility from environment feedback, and selectively governing repository contents.This avoids reliance on heuristic workflow-specific organization or parameter updates that may risk catastrophic forgetting and weakened clinical behaviors.Source: §2 · p. 3
• Self-evolving memory has progressed from static buffers and raw trajectories toward structured procedural routines, modular stores, and hierarchical skill libraries.Examples include Agent Workflow Memory [70], Dynamic Cheatsheet [57], GSEM [17], HealthFlow [95], Trace2Skill [37], SkillClaw [34], and EvoSkills [87].Source: §2 · p. 3

3.1 Problem Formulation

• SkeMex formulates agent–memory interaction as a Memory-based Markov Decision Process where skills are retrieved before actions, updated from feedback, and improve behavior without parameter updates.Each skill stores a retrieval key, reusable content, and utility statistic; memory evolution replaces backbone-model updates as the adaptation mechanism.Source: §3.1 · pp. 3–4

3.2 Overview of SkeMex

• SkeMex organizes distilled experience in a continuously evolving skills repository governed by a closed-loop Read–Write–Assess–Govern lifecycle.The repository separates general reasoning strategies, task-specific clinical patterns, and action-level tool-use knowledge into distinct branches.Source: §3.2 · pp. 4–5

3.3 Value-aware Skill Retrieval

• SkeMex retrieves skills once at episode onset using clinical-category routing, multi-channel screening, semantic similarity, historical utility, temporal strength, and branch-aware top-K selection.Episode-level retrieval stabilizes context across long-horizon tasks, reduces context fragmentation, and uses category-conditioned utility for general-branch skills.Source: §3.3 · p. 5

3.4 Trajectory-to-Skill Distillation

• SkeMex distills informative trajectories through gated retention and two-pass writing, creating or patching branch-specific skills subject to novelty and quality gates.The buffer filters infrastructure errors, repetitions, and trivial successes while preserving meaningful multi-step reasoning and informative failures.Source: §3.4 · pp. 5–6

3.5 Utility-driven Skill Valuation

• SkeMex values skills using window-level relative advantage, adoption-event credit, risk-sensitive penalties, and clipped utility updates that support category-aware retrieval and stable maintenance.Window aggregation reduces noise from per-sample outcomes, while category baselines account for differences in clinical-task difficulty.Source: §3.5 · p. 6

3.6 Closed-loop Self-evolution Memory

• Every N windows, governance merges redundant skills, deprecates low-utility entries, promotes consistently effective skills, and removes excess entries to keep repository branches compact.The cycle retrieves current skills, buffers evaluated trajectories, writes or updates skills, revises utilities, applies governance, clears the buffer, and continues without updating the backbone model.Source: §3.6 · p. 6

4.1 Experiment Settings

• Experiments cover nine medical benchmarks spanning clinical interaction, knowledge-intensive reasoning, multimodal cases, and held-out in-domain and out-of-domain evaluation.SkeMex uses DeepSeek-V3.2 as the main backbone, tests transfer to Qwen3.6-Plus, and is evaluated in both fixed-repository offline and streaming online modes.Source: §4.1 · pp. 6–7

4.2 Main Results

• SkeMex achieves the strongest offline results, improving ReAct by 7.88–10.59 points in-domain and 13.78 points out-of-domain, while remaining above ReAct across all out-of-domain benchmarks.In-domain, gains are from 48.20% to 56.08% with DeepSeek-V3.2 and 48.63% to 59.22% with Qwen3.6-Plus; out-of-domain, SkeMex gains 13.78 points over ReAct versus about 8.24 points for the strongest competing memory baseline.Source: §4.2 · pp. 7–8

4.3 Ablation Studies and Analyses

• Full SkeMex reaches 53.22%, while removing buffer gating lowers performance to 47.56%, showing that selective trajectory filtering and reviewed encoding protect repository quality.Single-Prompt Encoding and removing draft review reduce the average to 50.97% and 48.82%, respectively.Source: §4.3 · p. 9, Table 3 · p. 9
• Utility valuation requires baseline correction, adaptive learning rates, harm clamping, and conditional utility, while all three memory branches are complementary.Removing baseline correction causes drops up to 7.00% on LiveMedBench and 6.09% on MedXpertQA-MM; the full branch design outperforms General + Action by 4.57 points.Source: §4.3 · p. 9, Table 4 · p. 9

5 Conclusion

• SkeMex combines skill distillation, utility-driven valuation, and closed-loop governance to improve medical agents consistently over ReAct and representative memory-based agents across backbones and task settings.The framework supports scalable experience accumulation without changing model weights.Source: §5 · p. 10

A Agent Architecture

• The agent uses a bounded ReAct loop, while runtime skill injection guides decisions without changing task-state management, tool invocation, output enforcement, or termination semantics.This separation lets learned experience influence decisions while preserving stable execution behavior.Source: A Agent Architecture · p. 18

A.1 Runtime Components

• Each task loop combines model prompting, dataset-specific tools, local conversation memory, dynamically injected task categories and skills, and context guarding for long or repetitive trajectories.Injected evolution context is rendered per step but excluded from conversation memory to avoid duplication.Source: §A.1 · p. 18, Table 5 · p. 18
• The batch runner resets conversation and step histories per sample, sends images only on the first vision-model call, and updates tools with sample-specific context before execution.These controls prevent cross-sample leakage and reduce repeated visual input in later prompts.Source: §A.1 · p. 18

A.2 Stepwise Execution Protocol

• Every execution step requires reasoning followed by exactly one tool call or final response, creating a deterministic finite action interface that presents retrieved skills with applicability conditions.The prompt combines instructions, task history, the original request, and evolution context, with skill identifiers available for reference.Source: §A.2 · pp. 18–19
• The protocol records trajectories, executes valid tool calls, recovers from malformed outputs, and forces convergence to a final answer at the step limit.Fallback parsing ensures bounded tasks do not end without an output.Source: §A.2 · p. 19

A.3 Integration of Skill Memory into the Agent Loop

• Retrieved skills enter through runtime context as conditional references that the agent may adopt, ignore, or override, while recorded trajectories reveal which skills were useful, ignored, or harmful for later evolution.This creates a feedback loop in which retrieval guides execution and execution supplies behavioral evidence for memory updates.Source: §A.3 · p. 19

A.4 Intra-Task Context Control

• Loop breaking warns the agent after exact repeated tool actions, encouraging strategy changes while allowing reuse when inputs differ meaningfully.The mechanism is conservative because it matches repeated action signatures rather than merely repeated tools.Source: §A.4 · p. 20
• Confirmed-finding pinning summarizes verified early observations once and appends them to later prompts, while trimming compresses older history when token usage exceeds a configured budget fraction.Recent steps remain fuller, and pinned findings plus notices help the model interpret compressed history.Source: §A.4 · p. 20
• The context guard stabilizes intra-task reasoning through loop breaking, confirmed-finding pinning, and budget-aware history trimming while preserving the complete original trajectory.These mechanisms alter only the rendered prompt, not the stored execution record.Source: §A.4 · pp. 19–20, Table 6 · p. 20

B Tool Suite

• The modular tool suite spans evidence retrieval, structured medical knowledge, quantitative computation, multimodal perception, reasoning control, and benchmark-specific interaction through a common structured-call interface.This uniform protocol keeps the action space flexible across heterogeneous medical datasets; detailed LLM tool prompts appear in Appendix L.Source: B Tool Suite · pp. 20–21
• Specialized tools support current web information via Tavily Search API, medical evidence via MedRAG-style retrieval, medication safety via DrugBank-style resources, biomedical concepts via UMLS, and interactive history, examinations, and images under AgentClinic constraints [48].Additional tools safely perform numerical calculations, unit conversion, clinical scoring, image analysis, OCR, reflection, and final-answer verification.Source: B Tool Suite · pp. 21–22

C Additional Details of SkeMex

• SkeMex retrieves fixed skills at episode onset, preserves the backbone policy, and improves post-deployment by evolving general-, task-, and action-level memory branches.Each skill contains a retrieval key, procedural content, and utility statistic.Source: C Additional Details of SkeMex · p. 23
• Gated trajectories feed writing, valuation, and governance operators that create or patch skills, update utilities, merge redundancies, promote stable entries, and deprecate harmful ones.The operators act at the end of each learning window under branch-capacity constraints.Source: C Additional Details of SkeMex · p. 23

D Dataset Details

• The evaluation spans nine medical benchmarks covering interactive clinical decisions, patient-centered reasoning, medical question answering, and multimodal understanding across heterogeneous tasks and modalities.This design tests skill reuse in both text-only and multimodal clinical settings.Source: D Dataset Details · p. 23
• AgentClinic [48], MedJourney, LiveMedBench [81], HealthBench [3], and MediQ [30] assess interactive information gathering, patient workflows, rubric-based reasoning, health support, and uncertainty-aware questioning.These benchmarks collectively evaluate sequential decisions and clinically complete, patient-specific responses rather than only exact-match accuracy.Source: D Dataset Details · pp. 23–25
• LiveClin [67] provides evolving, contamination-resistant clinical cases from contemporary peer-reviewed reports, including text-only and multimodal questions.It targets recent, clinically grounded, and often complex patient cases.Source: D Dataset Details · p. 23
• MedXpertQA [96], MMMU [83], and MMMU-Pro [84] test advanced specialty reasoning and multimodal interpretation using medical images, diagrams, charts, tables, and stronger visual-textual integration.MMMU and MMMU-Pro use their Health & Medicine portions, while MedXpertQA includes text and multimodal subsets.Source: D Dataset Details · p. 25
• The unified preprocessing pipeline prioritizes difficult or multi-step cases, removes modality-inconsistent or unevaluable examples, normalizes inputs, and deduplicates accumulation and test splits.Open-ended examples are retained only when structured rubrics or other reliable evaluation criteria are available, preserving reusable experience while controlling cost and leakage.Source: D Dataset Details · p. 25
• After preprocessing, the evaluation pool contains 3,278 samples across 12 dataset configurations, with modality-separated variants and Health & Medicine restrictions for MMMU and MMMU-Pro.Table 8 reports dataset sizes and category distributions using task abbreviations such as differential diagnosis, treatment planning, imaging interpretation, and medication safety.Source: D Dataset Details · p. 25, Table 8 · p. 26

E.1 Baselines

• SkeMex is evaluated against specialist models, a no-memory tool-using agent, reflective methods, and multiple memory-agent families under matched tool-access and training-split conditions.The baseline suite spans medical specialist references, Vanilla ReAct, reflection-based methods, experience distillation, skill/workflow memory, persistent memory, and graph-structured clinical memory.Source: §E.1 · pp. 27–28
• Vanilla ReAct [82] removes skill memory, retrieval, utility estimation, and repository governance while preserving SkeMex’s prompts and tools, isolating the contribution of evolving skill memory.This comparison separates memory evolution from tool access alone.Source: §E.1 · p. 27
• The evaluated baselines reuse stored reflections or critiques from feedback, while Voyager [64] accumulates reusable skills from experience and retrieves procedures derived from successful medical trajectories for similar cases.Source: §E.1 · p. 27
• SkillWeaver [91], AgentWorkflowMemory [70], AgentKB [58], Evolver [75], DynamicCheatsheet [57], MobileE [68], CerebraFusionMemory [86], and GSEM [17] test skill, workflow, knowledge-base, persistent, fused, and graph-structured memory designs.SkeMex differs by combining structured skills with value-aware retrieval, category-normalized utility estimation, and periodic memory governance.Source: §E.1 · p. 28

E.2 Metrics

• Close-ended benchmarks use accuracy over successfully evaluated samples, with multiple-choice predictions and references normalized to uppercase option letters and unparseable responses counted as incorrect.For single-answer questions, correctness is exact option matching after normalizing formats such as A, (A), [A], A., and A :.Source: §E.2 · pp. 28–29
• Open-ended AgentClinic Text and MedJourney datasets also report accuracy, but correctness comes from a semantic-equivalence judge comparing each prediction with its reference answer.The judge returns a binary verdict used as the per-sample correctness indicator.Source: §E.2 · p. 29
• HealthBench and LiveMedBench use weighted rubric-based scoring, including penalties for unsafe or clinically inappropriate content, with mean sample or case scores reported respectively.Gemini-3-Flash evaluates all rubric criteria for a sample in one call; scores are clamped to [0, 1], and samples without positive-weight criteria receive zero.Source: §E.2 · p. 29
• The same scoring functions convert outcomes into memory rewards, supplying binary rewards for accuracy tasks and continuous rewards si ∈[0, 1] for rubric tasks during Assess-stage utility updates.This aligns reported benchmark metrics with the memory-evolution pipeline’s utility updates.Source: §E.2 · p. 29

• In-domain data are split into roughly equal accumulation and held-out test partitions, while out-of-domain benchmarks are reserved entirely for testing skill transfer.Table 9 summarizes the experimental train/test splits.Source: F Implementation Details · pp. 29–30, Table 9 · p. 30
• DeepSeek-V3.2 [32] is the primary backbone, while Qwen3.6-Plus reuses its repository without rewriting to test transferable skills; indexing uses text-embedding-3-large.Model calls run through APIs, while repositories, indices, utilities, and traces are maintained locally.Source: F Implementation Details · p. 30
• Each episode uses a maximum-seven-step ReAct trajectory, retrieves skills once at the start with default K = 6, and combines category, semantic-similarity, utility, maturity, and branch-aware signals.Trajectories exceeding the step budget are excluded from skill writing; retrieval uses a 0.2 similarity threshold, a 10% mature-skill bonus, and branch-specific pre-screening when candidates exceed 5.Source: F Implementation Details · pp. 30–31
• Utility valuation uses category-normalized rewards and adoption-aware credit assignment, with skill utilities clipped to [0, 1].The adoption-count-dependent learning rate uses warmup and decay, with base rate 0.05, maximum rate 0.20, 5 warmup steps, and 20 decay steps.Source: F Implementation Details · p. 31
• Every two learning windows, governance merges similar skills, deprecates low-utility entries, promotes mature skills, and enforces compact branch capacities.The merge threshold is 0.8; deprecation applies below 0.3, maturity requires utility at least 0.75 and usage count at least 15, and capacities are 12 general, 8 task-level per category, and 5 action-level per tool.Source: F Implementation Details · p. 31
• Context guarding preserves recent reasoning, retrieved skills, and pinned findings under a 16,384-token budget, while external and local tools communicate through interfaces that keep controller-side execution lightweight.The setup includes compressed older observations, modular medical retrieval and vision services, and CPU-side repository maintenance with separate GPU-based image analysis.Source: F Implementation Details · p. 31

Retrieval budget

• 42.52 at K = 3 rises to 42.93 at K = 9, while K = 12 reaches 42.70, indicating limited sensitivity to retrieval budget around the moderate default K = 6.Retrieving too few skills may provide insufficient guidance, whereas retrieving too many may introduce redundant or weakly relevant information.Source: G Sensitivity Analysis · p. 32, Figure 5 · p. 32

Retrieval weights

• 42.57 under memory-heavy retrieval is below 42.80 for the default weights, whereas semantic-similarity weighting reaches 42.85 and utility weighting 42.79.Semantic relevance and estimated utility support useful skill selection, while excessive emphasis on memory strength can reduce retrieval precision.Source: G Sensitivity Analysis · pp. 32–33, Figure 6 · p. 33

Learning window size

• 42.64 at L = 60 is the lowest tested average, while the default L = 30 scores 42.80 in a stable region near the best value of 42.82.Smaller windows adapt faster but may yield noisy utility estimates, whereas larger windows improve stability but can slow adaptation under distribution shifts or rare categories.Source: G Sensitivity Analysis · p. 33, Figure 7 · p. 33

Category baseline update coefficient

• 42.86 at α = 0.10 is the best average score, compared with 42.80 at the default α = 0.20, while larger coefficients degrade performance slightly.Overly large α values may track short-term reward fluctuations too closely, making moderate smoothing preferable.Source: G Sensitivity Analysis · p. 33, Figure 8 · p. 34

Utility update step

• 42.83 at ηmax = 0.30 remains close to the default 42.80, but increasing the step to 0.40 reduces the average to 42.54.Excessively large utility updates can amplify short-term feedback noise and destabilize skill valuation.Source: G Sensitivity Analysis · p. 33, Figure 9 · p. 34

Overall sensitivity

• Across sensitivity tests, SkeMex remains stable, with the largest but still modest variations arising from small retrieval sets, memory-heavy retrieval, or aggressive utility updates.The default configuration provides a balanced trade-off among retrieval coverage, retrieval precision, adaptation speed, and utility-estimation stability.Source: G Sensitivity Analysis · p. 34

H Ablation Study

• The full value-aware retrieval model reaches 53.22%, outperforming removal of memory strength at 50.76%, LLM ranking alone at 50.93%, and prescreening removal at 50.69%.These results indicate that retrieval benefits from jointly using semantic relevance, empirical utility, memory strength, and early candidate filtering.Source: Table 11 · p. 34, H Ablation Study · pp. 34–35
• Similarity Only reaches 47.64% and Utility Only 48.30%, respectively 5.58 and 4.92 points below the full model, showing that the signals capture complementary aspects of skill usefulness.Similarity can miss ineffective skills, whereas utility alone can favor overly general procedures that mismatch the current clinical context.Source: H Ablation Study · pp. 34–35
• The full closed-loop lifecycle achieves 53.22%, while removing maturation causes the largest decline to 47.60%, a 5.62-point drop.Maturation distinguishes repeatedly validated skills from unstable drafts and prevents insufficiently supported skills from dominating retrieval.Source: Table 12 · p. 35, H Ablation Study · p. 35
• Removing deprecation lowers the average to 49.27%, and disabling memory merging lowers it to 49.67%, showing that pruning harmful entries and consolidating redundancy preserve retrieval quality.Without deprecation, obsolete or low-utility skills add noise; without merging, similar skills crowd the candidate pool and reduce diversity.Source: H Ablation Study · p. 35
• Without capacity limits, the average falls to 51.05%, indicating that branch-wise repository control complements creation, validation, consolidation, and pruning.Unconstrained growth makes the repository harder to search and more vulnerable to low-value candidates.Source: H Ablation Study · p. 35

I.1 Offline OOD Generalization

• +13.78 points with DeepSeek-V3.2 and +13.91 points with Qwen3.6-Plus show SkeMex’s offline OOD gains over ReAct, while leading the strongest memory baselines by 5.44 and 2.97 points.SkeMex reaches 75.79% versus 62.01% with DeepSeek-V3.2 and 78.18% versus 64.27% with Qwen3.6-Plus.Source: Table 13 · p. 36, §I.1 · pp. 36–37
• +34.11 and +12.50 points over ReAct on AgentClinic-Text and MMMU-Pro under DeepSeek-V3.2, while SkeMex leads all five Qwen3.6-Plus benchmarks.Under Qwen3.6-Plus, gains over ReAct include +31.31 on AgentClinic-Text, +10.72 on MMMU-Pro, and +9.17 on AgentClinic-MM; GSEM gains only +1.87 and +2.83 on average and falls below ReAct on AgentClinic-Text.Source: §I.1 · p. 37

I.2 Cross Backbone Generalization

• On Qwen3.6-Max-Preview and Kimi-2.6, SkeMex achieves the best score on every benchmark, improving average performance over ReAct by 11.64 and 11.33 points, respectively.It also exceeds the strongest competing memory baseline by 3.93 points on Qwen3.6-Max-Preview.Source: §I.2 · pp. 37–38, Figure 10 · p. 37, Figure 11 · p. 38
• On GLM-5.1, SkeMex maintains nonnegative gains on all six datasets, ranks first on five, ties on one, and improves the ReAct average by 7.77 points.Competing memory baselines sometimes fall below ReAct, whereas SkeMex remains consistently beneficial.Source: §I.2 · pp. 38–39, Figure 12 · p. 38
• Across eighteen backbone–dataset pairs, SkeMex is best in seventeen and tied once, raising pooled average performance from 48.88% to 59.13% over ReAct.The pooled gain is 10.25 points, with positive average gains across all six evaluated datasets.Source: §I.2 · p. 39

I.3 Cross Backbone Skill Transfer

• A fixed skill repository learned from DeepSeek-V3.2 transfers to Claude Sonnet-4.6 and Qwen3.6-35B-A3B, testing portability across model families, deployment regimes, and architectures.Unlike rebuilding repositories for each target, this setting changes only the consuming backbone.Source: §I.3 · p. 39
• With the fixed DeepSeek-V3.2 repository, SkeMex improves Claude Sonnet-4.6’s average from 49.59% to 60.27%, leads all seven benchmarks, and beats CFM by 3.31 points.Gains span text-only, rubric-based, and multimodal medical tasks.Source: §I.3 · p. 39, Table 14 · p. 39
• On Qwen3.6-35B-A3B, transferred skills raise the ReAct average from 47.71% to 58.23%, with positive gains on every benchmark and the strongest overall average.SkeMex ranks first on five datasets despite small deficits to CFM and GM on two individual datasets.Source: §I.3 · p. 40

I.4 Execution Cost and Interaction Depth

• SkeMex averages 4.77 interaction steps versus 3.17 for ReAct, increasing MMMU-Pro performance from 35.71% to 48.21% as retrieved skills promote decomposition, verification, and tool use.The added interaction depth reflects more structured clinical reasoning rather than merely redundant actions.Source: §I.4 · p. 40, Table 15 · p. 41
• SkeMex averages 116.06 seconds per task versus 54.48 seconds for ReAct, although HealthBench runtime falls to 52.36 seconds versus ReAct’s 64.19 seconds.The general overhead comes from longer skill-rendered contexts and embedding-based retrieval, while relevant skills can streamline execution.Source: §I.4 · pp. 40–41, Table 16 · p. 42

I.5 Impact of Training Data Order

• The experiment evaluates Random, Category-based, Easy-to-Hard, and Hard-to-Easy ordering on held-out HealthBench and LiveMedBench tests after repository evolution on corresponding training splits.Difficulty is estimated from DeepSeek-V3.2 ReAct-style scores, with higher scores indicating easier cases; Category-based ordering processes larger task categories first.Source: §I.5 · p. 43
• Random ordering achieves the best overall performance at 45.03%, supporting it as SkeMex’s default for balanced and generalizable offline skill evolution.Random shuffling balances categories and difficulty levels across learning windows, reducing the risk that early repository updates are dominated by narrow or unusually difficult samples.Source: Table 17 · p. 43, §I.5 · p. 43
• Easy-to-Hard is the strongest structured curriculum, averaging 44.20% with a 30.89% HealthBench score, whereas Hard-to-Easy performs worst at 41.41%.Starting with easier cases may stabilize reusable skill extraction, while difficult-first processing can expose early skill writing to noisy or overly specific error patterns that affect later retrieval and governance.Source: §I.5 · p. 43

Evaluation scope

• SkeMex is evaluated on benchmarks that cannot fully represent the diversity and lack of standardization of real clinical environments.Real settings may involve more varied patient histories, institutional workflows, and decision constraints.Source: J Limitations · p. 43
• The evaluation covers several mainstream backbone models but not all available foundation models because of experimental cost.Source: J Limitations · p. 43

Efficiency

• Retrieved skills increase prompt length and reasoning time, causing higher API usage and longer wall-clock time than memory-free agents.This efficiency trade-off is common in skill-based and memory-augmented agents that seek more structured and reliable reasoning.Source: J Limitations · pp. 43–44

Safety and societal risks

• SkeMex may reinforce incorrect patterns or be misused in high-stakes settings without sufficient human oversight.The framework should therefore be treated as a decision-support and research tool, not a substitute for professional medical judgment.Source: J Limitations · p. 44

K Case Study

• SkeMex successfully guides differential diagnosis, administrative tool selection, diagnostic-boundary verification, and multimodal procedure selection across representative clinical cases.The cases demonstrate avoiding uninformative search loops to diagnose focal seizure, routing billing queries to authoritative coding search, verifying non-cirrhosis against Brunt criteria, and selecting at least six bite-on-bite jumbo forceps biopsies.Source: K Case Study · p. 44, Figure 13 · p. 45, Figure 14 · p. 46
• Early tool-format errors consumed the interaction budget, leaving insufficient workup and causing premature convergence on “Lymphoma” instead of the ground-truth Diffuse Large B-Cell Lymphoma.The failure case lacked confirmatory LDH, biopsy, and CT investigations and incorrectly favored Hodgkin lymphoma, producing a categorically correct but clinically insufficient diagnosis.Source: Figure 17 · p. 49, K Case Study · p. 49

L Prompts

• The prompts section provides the full text of prompts governing core agent interaction, memory evolution, and automatic evaluation.These prompts operationalize SkeMex’s planning, skill injection, retrieval, evolution, and grading procedures.Source: L Prompts · p. 50

L.1 Core Agent Interaction Prompts

• Core interaction prompts enforce planning, reasoning, tool, and response formats, concatenate history with the request and injected skills, force convergence at the step limit, and format skills by repository branch.The relevant prompts define strict behavior constraints, context assembly, final-answer convergence, and grouping into general, task-level, and action-level experience.Source: Figure 18 · p. 50, Figure 19 · p. 51, Figure 20 · p. 51

L.2 Memory Evolution Pipeline Prompts

• Memory-evolution prompts classify tasks into action-oriented categories and pre-screen semantically relevant skills before top-ranked candidates enter the agent context.The classifier supports methodology-level clinical-skill retrieval, while pre-screening selects candidates for downstream scoring and injection.Source: §L.2 · pp. 52–53, Figure 22 · p. 52, Figure 23 · p. 53

ANALYSIS_PROMPT_BINARY

• The binary trajectory-analysis prompt evaluates skill adoption and extracts decisive success or failure patterns from binary-outcome datasets.It assigns trajectory attribution to a medical AI knowledge engineer.Source: ANALYSIS_PROMPT_BINARY · p. 54, Figure 24 · p. 54

ANALYSIS_PROMPT_RUBRIC

• Rubric-based analysis attributes score deductions to agent actions, then mutation, review, and merge prompts create or patch skills, govern novelty and anti-fragmentation, and consolidate overlapping skills.The workflow supports concrete skill drafts or modifications, governance review, and concise merging of semantically similar representations.Source: ANALYSIS_PROMPT_RUBRIC · pp. 55, 58, Figure 25 · p. 55, Figure 26 · p. 56

L.3 Automatic Evaluation Prompts

• Automatic-evaluation prompts define rubric-based grading for HealthBench and interactive clinical evaluation for LiveMedBench using the query, prediction, and rubric criteria.Separate grader system prompts are tailored to HealthBench and LiveMedBench, with a shared user template supplying evaluation inputs.Source: Figure 29 · p. 59, Figure 30 · p. 59, Figure 31 · p. 59