💻 computer science

A Self-Evolving Agent for Longitudinal Personal Health Management

The paper introduces HealthJev, an open-source, self-evolving agent that utilizes a lightweight, self-trained filtering mechanism (JevGate) to safely accumulate and distill longitudinal memory from medical encounters, significantly outperforming existing frameworks across diverse clinical tasks in a 90-day simulated benchmark.

Original authors: Hongcheng Guo, Haoran Li, Jiebi Deng, Tong Jin, Jinghong Han, Yuxin Wang, Zexin Wang, Jinyu Sun, Qingyi Si, Weikang Gong, Xiahai Zhuang, Jia You, Wei Cheng, Jianfeng Feng

Published 2026-11-05

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Hongcheng Guo, Haoran Li, Jiebi Deng, Tong Jin, Jinghong Han, Yuxin Wang, Zexin Wang, Jinyu Sun, Qingyi Si, Weikang Gong, Xiahai Zhuang, Jia You, Wei Cheng, Jianfeng Feng

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a doctor who remembers everything you have ever told them about your health, not just the facts from your last visit, but the patterns of your life over months and years. They recall that you tend to skip meals when you are stressed, that your blood pressure rises on rainy days, and that a specific medication made you feel dizzy last winter. This kind of continuous, evolving understanding is the holy grail of personal health management. Yet, most current digital health tools operate like a series of isolated conversations. They are excellent at answering a single question or interpreting a snapshot of data, but they struggle to learn from one interaction to the next. They often forget the context that makes a piece of advice truly useful, treating every new request as if it were the very first time you ever spoke to them. This limitation means that even the smartest digital assistants cannot adapt to the slow, shifting rhythms of a real human life.

Researchers at Fudan University and several other institutions have built a new kind of digital health agent designed to solve this problem. They call it HealthJev. Unlike standard programs that simply store a long list of past conversations, HealthJev is designed to learn and change its own behavior over time. It separates general medical knowledge, which stays the same for everyone, from a private, personal memory that belongs only to you. This personal memory is not just a storage bin; it is a living system that decides what to keep, what to update, and what to discard after every interaction. The system is built to handle the complexity of long-term care, where a stable allergy might need to be remembered forever, while a temporary symptom might fade away, and a failed diet plan might need to be revised into a new, better rule.

The core innovation of HealthJev is a mechanism that acts as a gatekeeper for its own learning. After every conversation or task, the agent asks itself: "Did I learn something useful here that I should remember for next time?" It generates a potential lesson, such as "This user prefers dinner recommendations after 8 PM" or "This specific drug interaction caused a side effect." Before this lesson is saved into the permanent memory, it passes through a lightweight, specialized filter called JevGate. This filter is a small, fast decision-making model trained specifically to spot errors. It checks the lesson against the evidence to ensure it is correct before allowing it to enter the system. If the lesson is flawed, the gatekeeper rejects it, preventing the agent from learning the wrong thing. This process happens incredibly quickly, taking less than a tenth of a second for each decision, which allows the system to accumulate thousands of lessons without slowing down.

To test whether this approach actually works, the researchers created a rigorous simulation based on real medical records. They took the health data of twenty actual patients from a large hospital database and used it to create twenty digital personas. Each persona was given a unique personality, job, and set of health conditions, ranging from diabetes and heart disease to high blood pressure. These digital patients then interacted with the HealthJev agent every day for ninety days, simulating a full quarter of a year of continuous care. During this time, the users reported their daily habits, shared new symptoms, asked for advice, and sometimes made mistakes or changed their minds. The researchers also included specific tests to see if the agent remembered things from weeks ago and if it respected privacy boundaries, such as refusing to share sensitive details with unauthorized third parties.

The results showed that HealthJev significantly outperformed other existing systems. When compared to agents that had no memory at all, or to other systems that tried to remember the past but lacked a smart filtering mechanism, HealthJev consistently provided better, more accurate, and more personalized support. In the ninety-day simulation, it scored higher on every measure, from recalling details mentioned days earlier to correcting misreported health values. The system successfully learned to adapt its advice as the user's life changed, such as adjusting meal plans when a user started working late shifts or modifying exercise recommendations based on new pain levels. Crucially, the system did not just memorize facts; it learned how to apply them. It developed reusable procedures for common tasks, like how to track blood pressure or manage a specific diet, and refined these procedures over time based on what worked and what did not.

Beyond the long-term simulation, the researchers tested whether the lessons HealthJev learned could be transferred to other medical tasks. They took the library of lessons the agent had accumulated and applied them to six different biomedical challenges, including predicting hospital readmissions, interpreting medical images, and answering complex questions about genetics. In every single case, the agent that had access to this learned library performed better than the same agent without it. The improvement was substantial, with an average increase in accuracy of over thirteen percentage points across all tasks. Perhaps most surprisingly, the lessons learned from one type of medical problem helped the agent solve completely different types of problems. For instance, knowledge gained from analyzing genetic data helped improve predictions about diabetes readmission. This suggests that the agent is not just memorizing answers but is building a deeper, more flexible understanding of health that can be applied in many different contexts.

The study also demonstrated that this learned knowledge is portable. The researchers took the library of lessons created by one specific type of AI model and applied it to six other different models. Even though these other models had never seen the original training data, they all improved their performance when they were allowed to use the lessons. This indicates that the value of the system lies in the quality of the filtered lessons themselves, not just in the specific computer brain that generated them. The system proved capable of governing its own memory, deciding what was worth keeping and what was not, and doing so with a speed and efficiency that made it practical for real-world use.

While the results are promising, the researchers are careful to note that these findings come from simulations and offline tests. The ninety-day interactions were created by a computer program acting as a user, and the evaluations were performed by automated scoring systems. The study did not involve real patients in a live clinical setting, nor did it measure actual health outcomes like whether people got better or avoided hospitalization. The work serves as a proof of concept, showing that a self-evolving agent with a governed memory can handle the complexity of longitudinal health management better than current tools. It suggests a path forward where digital health assistants do not just answer questions but truly learn from the people they help, adapting to their lives over months and years. The code and the simulation framework used in the study are now available for other scientists to explore, offering a new foundation for building health tools that remember, learn, and evolve alongside the people they serve.

Technical Summary: A Self-Evolving Agent for Longitudinal Personal Health Management

Problem Statement

Current medical large language models (LLMs) and agents excel at single-encounter question answering and evidence interpretation but struggle to accumulate and govern experience across multiple interactions. Personal health management is inherently longitudinal; effective support requires adapting to delayed feedback, changing routines, and evolving health trajectories rather than treating every request as a fresh interaction.

Existing solutions face a critical design gap: they either lack state across visits (encounter-level AI) or rely on general-purpose memory mechanisms evaluated on open-domain dialogues rather than clinical trajectories. A specific challenge in personal health is selective memory updates. A single interaction may contain stable facts (e.g., allergies), transient symptoms, failed recommendations, and sensitive identifiers, each requiring different persistence and disclosure policies. Simply appending full history increases context costs and privacy risks, while withholding history breaks continuity. Furthermore, lessons distilled from single cases can be erroneous; if admitted to memory without scrutiny, they can degrade future reasoning. Existing systems lack a mechanism to explicitly govern what is retained, revised, or withheld, and to filter erroneous lessons before they influence future behavior.

Methodology: HealthJev

The authors present HealthJev, an open-source agent designed to separate shared medical knowledge and safety rules from a private, longitudinal memory that evolves through a closed-loop process.

Architecture and Memory Layers

HealthJev operates through a four-stage closed loop: Perception, Reasoning, Action, and Post-episode Induction.

Perception: Assembles current requests, health signals, and relevant prior context.
Reasoning: Combines task-relevant memory with shared medical knowledge to plan under user-specific constraints.
Action: Executes the task workflow or generates a response.
Induction: Determines how information from the completed episode should update the memory.

The memory is structured into five distinct layers to support selective retrieval and writeback:

L0 (Shared): Behavioral rules, safety boundaries, and escalation conditions.
L1 (Shared): Medical knowledge, tool indices, and task knowledge.
L2 (User-Specific): Durable profile facts (chronic conditions, allergies, stable preferences).
L3 (User-Specific): Reusable procedures (e.g., meal planning SOPs, risk routing) updated by distilled lessons.
L4 (User-Specific): Episodic traces (local context, transient symptoms, feedback).

The JevGate Admission Controller

To prevent erroneous lessons from corrupting the memory, HealthJev employs JevGate, a lightweight, self-trained "System-One" decision model.

Function: JevGate screens candidate lessons distilled from completed cases to decide whether they should be admitted to the reusable experience library (L3).
Training: Built on the Kev-9B decision model (frozen Qwen3.5-9B base with LoRA and a pointer head), it is fine-tuned on admission data derived from real interaction outcomes. It uses a pointer mechanism to score admission vs. rejection options rather than generating text.
Efficiency: Decisions are made locally with a median latency of 86 ms and consume approximately 1,000 input tokens, ensuring negligible runtime overhead for experience accumulation.

Evaluation Frameworks

The authors evaluated HealthJev using two primary frameworks:

Longitudinal Benchmark: Constructed from real MIMIC-IV trajectories of 20 patients, simulating 90 days of daily interactions (1,800 total days). The benchmark includes diverse goals (lifestyle, event reports, medication consults) and probes for longitudinal recall, temporal validity, and privacy governance.
Six Biomedical Tasks: Spanning electronic records (diabetes readmission), physiological signals (ICU SOFA), clinical text (MIMIC-ICD coding), medical calculation (MedCalc), imaging (ODIR5K fundus), and genomics (GeneTuring).

Key Results

Longitudinal Personal Health Management

In the 90-day simulated benchmark, HealthJev significantly outperformed a memory-free ReAct agent and three memory-augmented frameworks (MemoryBank, Mem0, Memp) across all paired comparisons.

Performance: HealthJev achieved a mean judge score of 0.668, compared to 0.634 for the strongest baseline (MemoryBank) and 0.510 for ReAct.
Significance: All differences were statistically significant ( $p < 0.001$ ) via paired Wilcoxon signed-rank tests. HealthJev outperformed the strongest baseline for 15 of the 20 simulated users.
Governance: The system successfully managed privacy boundaries, retaining key medical information while minimizing the storage of sensitive everyday details.

Performance Across Biomedical Tasks

HealthJev improved performance on all six heterogeneous biomedical tasks compared to the base condition (no memory/tools).

Mean Gain: A mean absolute gain of 13.3 percentage points over the base condition.
Task-Specific Gains:
- Diabetes Readmission: +31.0 pp ( $p < 0.001$ )
- GeneTuring: +16.5 pp ( $p < 0.001$ )
- ODIR5K (Fundus): +16.3 pp ( $p < 0.001$ )
- MedCalc: +9.0 pp ( $p = 0.022$ )
- MIMIC-ICD: Significant increase in non-zero scores ( $p = 0.008$ )
- ICU SOFA: +6.0 pp (not statistically significant, $p = 0.13$ )

Transferability and Efficiency

Cross-Backbone Transfer: When the frozen experience library (accumulated on Doubao-Seed-2.0-pro) was mounted on six other backbone models (including GPT-4.1, Gemini, and Qwen variants), it improved the macro-average metric on all backbones by 6.8 to 15.1 percentage points.
Cross-Task Transfer: Libraries accumulated on one task improved performance on other tasks. Notably, transferring GeneTuring and ICU SOFA libraries to diabetes readmission yielded gains of 29.5 and 27.5 percentage points, respectively.
JevGate Efficiency: The admission decision model achieved a pooled agreement of 0.735 (up from 0.532 in zero-shot) and an AUC of 0.810 on a held-out set. The median decision latency remained at 86 ms.

Significance and Claims

The paper claims that HealthJev demonstrates that governed, self-evolving memory is essential for longitudinal personal health management beyond single encounters. Key contributions include:

Explicit Memory Governance: The separation of shared knowledge from user-specific state and the implementation of a selective admission mechanism (JevGate) allow the agent to decide what to keep, revise, or withhold, addressing the specific persistence and disclosure needs of health data.
Reusable Experience: The study shows that experience accumulated by an agent is not only reusable within the same task but can be transferred across different tasks and different language model backbones, suggesting a form of generalizable medical reasoning.
Efficiency: The use of a lightweight, non-generative controller (JevGate) for memory admission decisions enables continuous experience accumulation with negligible runtime overhead, solving the cost-latency trade-off of using large models for routine memory checks.

The authors maintain a modest stance, noting that the evidence is limited to offline benchmarks and simulated interactions. They emphasize that HealthJev is positioned as an assistive system for information organization and management, and that deployment for diagnosis or autonomous intervention would require separate prospective clinical evaluation and regulatory review. The work is publicly available at https://github.com/HC-Guo/Self-Evolving_HealthJev.