A personal-health agent that remembers you: HealthJev screens bad lessons in 86 ms and lifts six tasks by 13.3 points on average

A personal-health assistant now treats each encounter as an event that can change later advice, rather than an isolated question. HealthJev keeps shared behavioural safety rules and medical knowledge in layers L0 to L1 and splits user-specific state into L2 profile facts, L3 reusable procedures and L4 episodic traces, then decides after each interaction what to write, revise or withhold. Earlier medical language models and symptom-triage systems mainly handled question answering, diagnostic reasoning and triage within a single encounter, while general agent memory was evaluated on open-domain dialogue and tasks; neither addressed the persistence and disclosure decisions personal health data demand. On a longitudinal benchmark of 90 simulated daily interactions for 20 users built from real MIMIC-IV trajectories, HealthJev scored a mean 0.668 across all 1,800 interaction days, versus 0.634 for MemoryBank, 0.593 for Mem0, 0.510 for ReAct and 0.502 for Memp; all four paired comparisons were significant, with the largest p about 1.1×10⁻⁴, and 15 of 20 patients had a higher mean daily score than under the strongest baseline, MemoryBank.

Accumulated experience can be reused away from the model and task that produced it. Experience libraries built with Doubao-Seed-2.0-pro were frozen and mounted directly on six further language-model backbones, and the six-task macro-average improved on all six. The six tasks span electronic records, physiological signals, medical imaging, genomics and clinical text; HealthJev beat the base condition and every memory-augmented baseline on each task's primary metric, with a mean absolute gain of 13.3 percentage points over the base condition. The frozen library brought macro-average gains of 6.8 to 15.1 percentage points on DeepSeek-V4.1-Flash, GPT-5.4, GPT-4.1, Gemini-3.8-Flash, Qwen3.5-Plus and Qwen3.8-27B; across eight donor-to-target task pairs, cross-task transfer yielded gains of 0.3 to 29.5 percentage points, with the GeneTuring and ICU SOFA libraries lifting diabetes readmission by 29.5 and 27.5 points respectively, both with McNemar exact p < 0.05.

A lesson must pass a lightweight review before entering memory, and that review is treated as a design problem in its own right. JevGate builds on the Kev-9B decision model and uses a pointer head to score rejection and admission options directly instead of generating a verdict as text. Delegating many routine judgements to a general-purpose language model adds latency and cost, while fixed rules handle complex semantics poorly, so an admission controller is needed that is both cheap and able to make semantic judgements. On a balanced held-out set of 600 admission cases across six tasks, zero-shot Kev-9B reached a pooled agreement of 0.532 and fine-tuned JevGate reached 0.735, with pooled AUC rising from 0.733 to 0.810; across the 2,983 admission decisions in the six-task evaluation, median latency was 86 ms and 90th-percentile latency 184 ms, each decision consumed about 1,000 input and 22 output tokens on average, and 81.8% of candidate lessons were admitted.

The demonstrated scope is offline evaluation: longitudinal interactions come from simulated users built on de-identified MIMIC-IV records and are graded by the automated judge GLM-5.2 against a six-dimension rubric, with no human ratings; the six biomedical tasks use public or controlled public datasets, and at test time the experience library is frozen and writeback is disabled. The intended readers are researchers and engineers working on agent memory governance and medical AI evaluation.

Simulation cannot fully reproduce the ambiguity, missingness and behavioural feedback of prospective use, and deployment-level privacy still depends on access control, security testing and data governance. The ICU SOFA gain was 6.0 percentage points but did not reach statistical significance (p = 0.13), and MIMIC-ICD's mean set F1 improved by only 1.2 percentage points.