Personalized State-Transition-Aware Memory for Clinical Agents

Maryam Haghifam; Zahra Rajabi; Yizhou Sun; Carlos Morato

cs.CLSep 29, 2026

Personalized State-Transition-Aware Memory for Clinical Agents

Authors: Maryam Haghifam, Zahra Rajabi, Yizhou Sun, Carlos Morato

Organizations: Department of Computer Science, University of California, Los Angeles, CA, USA · Optum AI, UnitedHealth Group, Minneapolis, MN, USA

Abstract

Large language model (LLM) agents that reason over clinical records must track changes in a patient's state while preserving the history needed to understand them. Simply accumulating memories leaves it unclear which information still applies, whereas overwriting earlier memories can erase evidence needed to reconstruct treatment history and clinical trajectories. We introduce STAM, a state-transition-aware memory framework that records state changes as new clinical entries arrive. STAM combines semantic retrieval with typed clinical relations to identify affected memories, maintaining current information in Active and superseded or resolved information in History. At read time, a query-dependent gate selectively serves historical memory. Across four longitudinal clinical benchmarks, we evaluate STAM with downstream question answering, direct state-maintenance diagnostics, and comparisons at approximately matched context lengths.

Figures & tables

Figure 1: Illustrative example of write-time state maintenance. In this example, Full Context and flat retrieval mix earlier positive chest-pain evidence with the current state and answer incorrectly.

Figure 2: Overview of the proposed state-aware memory framework.

Method	MMB [-1pt]mAcc.	MLC [-1pt]mF1	MQA [-1pt]F1	CB [-1pt]Acc.
Graphiti (Zep)	26.8	31.9	23.3	47.5
LightMem	35.7	33.2	22.2	50.5
A-Mem	32.5	39.0	42.3	49.2
MemOS	32.0	27.5	19.9	46.8
Mem0	22.1	28.3	23.1	45.2
MemRL	21.6	22.7	22.2	47.0

Table 1: Answer quality (%; higher is better) across persistent memory systems with Qwen3-4B-Instruct and GPT-5 readers.

		Qwen		GPT-5
Data	Type	STAM	FC	STAM	FC
MMB	State update	42.0	34.0	68.0	66.0
MLC	Longitudinal	21.9	14.8	24.8	28.9
CB	Current state	56.0	50.0	66.0	62.0
	Sequence	47.5	20.0	50.0	45.0

Table 2: (a) State- and time-oriented question categories. (b) Overall comparison with Full Context (FC) and Flat RAG. For the Qwen3-4B-Instruct comparisons with Flat RAG, STAM uses the approximately context-matched configurations described in Appendix C.2 .

Dataset	A+H	ACTIVE only	$\Delta$	A+H	ACTIVE only	$\Delta$
	Qwen3-4B-Instruct			GPT-5
MMB	35.68	34.12	+1.57	51.27	46.27	+5.00
MLC	35.91	36.00	$-0.09$	36.73	38.28	$-1.56$
MQA	54.73	52.63	+2.10	85.26	53.54	+31.72
CB	50.50	49.75	+0.75	49.50	50.00	$-0.50$

Table 3: Historical-access ablation. Comparisons are paired within each dataset/model configuration; A+H serves both stores, whereas ACTIVE only omits History .

Method	MMB	MLC	MQA	CB
Append-only	34.16	35.40	55.99	49.25
STAM	37.40	36.20	57.45	50.20
$\Delta$	+3.24	+0.80	+1.46	+0.95

Table 4: Paired comparison with append-only memory under the Qwen3-4B-Instruct reader using the append-only ablation configuration.

\Delta

denotes STAM minus append-only in score points. MMB: macro accuracy; MLC: type-macro F1; MQA: Token-F1; CB: accuracy.

Variant	Pair recall $\uparrow$	False archival $\downarrow$
Full model	53.9	7.6
Semantic only, $B=48$	51.6	12.3
Rule: all non-episodic	28.6	17.6

Table 5: Write-time ablations: (a) MedMemoryBench and (b) MIMIC-QA.

B

is the impact-candidate budget.

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Persons	Records	Mean tokens	Median	Range
MedMemoryBench	5	101.0	718.9K	666.4K	644.5K–815.0K
MedLoCoMo	12	31.1	37.9K	37.1K	22.6K–61.8K
MIMIC-QA (test)	20	1.1	64.7K	44.8K	9.8K–211.1K
ClinicalBench	43	5.2	20.0K	15.0K	3.4K–73.0K

Appendix

Table 6: Per-person context-length statistics for the evaluated cohorts. “Records” denotes sessions for MedMemoryBench and admissions for the remaining datasets. Token counts cover each individual’s complete record.

Benchmark	Question types	Scoring
MedMemoryBench	Entity exact match; multiple choice	Deterministic
	Temporal localization; state update; inference generation; multi-hop clinical deduction	GPT-5 judge
MedLoCoMo	All six question types	Macro deterministic token-level F1
MIMIC-QA	All question types	Mean deterministic token-level F1
ClinicalBench	All nine question types	GPT-5 judge against physician-corrected references

Appendix

Table 7: Evaluation scoring by benchmark.

Benchmark	Reader	Evidence cap	Questions truncated
MedMemoryBench	GPT-5	275K	329 / 496 (66.3%)
MedLoCoMo	Qwen3-4B	100K	0 / 320 (0%)
MedLoCoMo	GPT-5	100K	0 / 320 (0%)
MIMIC-QA	Qwen3-4B	110K	179 / 400 (44.8%)
MIMIC-QA	GPT-5	280K	21 / 400 (5.2%)
ClinicalBench	Qwen3-4B	120K	0 / 400 (0%)

Appendix

Table 8: Raw-record truncation in the reported Full Context evaluations. Only final Full Context configurations reported in the main results are included.

Qwen3-4B-Instruct
Method	Selection	MMB	MedLoCoMo	MIMIC-QA	ClinicalBench
LightMem	20	1.1	0.8	1.3	1.1
A-Mem	5	23.6	8.0	18.0	8.1
MemOS	5	0.4	0.4	1.0	0.7
Mem0	5	0.1	0.1	0.1	0.1
MemRL	2 of 20	0.5	0.6	0.7	0.7

Appendix

Table 9: Baseline selection settings and median evidence tokens per question (thousands). MMB denotes MedMemoryBench. a 199 logged questions; b one persona; c Qwen3-4B memories read by GPT-5. NR: not recorded; NA: unavailable. CUPMem reports only evidence passing its premise check, so 0.0 does not imply an empty store.

Benchmark	Training	Validation	Evaluation
MedMemoryBench	Personas 6, 8–12	Persona 7	Personas 1–5
MedLoCoMo	5 patients	1 patient	12 patients; 320 questions
MIMIC-QA	9 admissions from 9 patients	1 admission from 1 patient	20 patients
ClinicalBench	30–31 patients per fold	4 patients per fold	8–9 patients per fold; 400 questions overall

Appendix

Table 10: Data partitions for write-module fine-tuning. MIMIC-QA adapter validation is an internal holdout from the benchmark’s training split. ClinicalBench combines predictions from five held-out folds.

	MedMemoryBench		MedLoCoMo		MIMIC-QA
Method	Tokens/q	Recall	Tokens/q	Recall	Tokens/q	Recall
Flat RAG	1.9K	0.536	1.3K	0.560	2.3K	0.608
LightMem	1.3K	0.727	0.8K	0.624	1.3K	0.515 ‡
MemOS	0.4K	0.657	0.4K	0.543	1.0K	0.338 ‡
Mem0	0.1K	0.618	0.1K	0.543	0.1K	0.500 ‡
Ours	4.2K	0.802	4.1K	0.822	2.3K	0.959

Appendix

Table 11: Retrieval coverage under the configurations used for downstream QA. Because methods provide different amounts of context to the answering model, these results characterize their evaluation configurations rather than matched-budget retrieval comparisons. Recall uses dataset-specific gold evidence annotations and is not comparable across datasets. ClinicalBench has no evidence-level gold annotations.

Dataset	Comparator (ctx.)	STAM ctx.	Ratio	STAM	Comparator	$\Delta$ ( $p$ )
MedMemoryBench	LightMem (1,312)	1,318	1.00	39.87	39.90	$-0.03$ (.95)
	Flat RAG (1,898)	1,892	1.00	37.36	39.10	$-1.74$ (.46)
	MemOS (449)	444	0.99	30.19	32.59	$-2.40$ (.30)
	MemRL (596)	564	0.95	29.30	22.72	$+6.58$ (.007)
	Mem0 (134)	157	1.17	24.75	22.62	$+2.13$ (.36)
	LightMem lifted (4,505)	4,652	1.03	39.35	41.57	$-2.22$ (.28)

Appendix

Table 13: Context-matched evaluation. Context values are median rendered reader tokens, and Ratio is STAM context divided by comparator context. All arms are answered by the same Qwen3-4B reader under the benchmark harness’s reader settings. MedMemoryBench reports macro accuracy, MedLoCoMo reports type-macro token-F1, MIMIC-QA reports mean Token-F1, and ClinicalBench reports pooled accuracy. Rows labeled “lifted” use expanded comparator context. † Per-question comparator outputs are unavailable, so no paired significance test is reported.

Dataset	Comparator	Ratio	STAM	Cmp.	$\Delta$ ( $p$ )
MMB	A-Mem	0.39	35.16	33.43	$+1.73$ (.51)
	CUPMem	0.42	20.72	5.37	$\mathbf{+15.35}$ ( $\mathbf{<.001}$ )
MedLoCoMo	A-Mem	0.49	34.20	36.60	$-2.40$ (.41)
	CUPMem	0.56	27.12	18.30	$\mathbf{+8.82}$ ( $\mathbf{<.001}$ )
MIMIC-QA	A-Mem	0.25	54.67	42.30	$\mathbf{+12.37}$ ( $\mathbf{<.001}$ )
	CUPMem	2.12	43.96	5.92	$\mathbf{+38.04}$ ( $\mathbf{<.001}$ )

Appendix

Table 14: Comparisons outside the context-matching range. Ratio denotes STAM context divided by comparator context. These rows are excluded from the matched-context significance summary.

Dataset	$\lambda$	HISTORY skip	Quality $\Delta$ ( $p$ )
MedMemoryBench	0.01	26.2%	$-0.67$ ( $p=.289$ )
MedLoCoMo	0.5	51.6%	$-0.11$ ( $p=.469$ )
MIMIC-QA	0.2	79.0%	$-0.70$ ( $p=.147$ )
ClinicalBench	0.5	1.0%	$0.00$ ( $p=1.000$ )

Appendix

Table 16: Selective HISTORY serving under the published gate configuration. The gate chooses between the original ACTIVE+HISTORY packet and the same packet with HISTORY removed. HISTORY skip is the fraction of queries served without HISTORY. Quality differences are paired against the corresponding BOTH condition.

System	Dataset	Active sufficient	History critical	History harmful	Neither correct
1.7B write / 4B reader	MMB	35.7%	4.8%	2.8%	59.5%
	MLC	43.4%	0.6%	0.3%	55.9%
	MQA	56.2%	4.2%	3.0%	39.5%
	CB	50.2%	5.2%	5.2%	44.5%
GPT-5 write / GPT-5 reader	MMB	47.4%	11.9%	7.5%	40.7%
	MLC	43.8%	4.1%	5.6%	52.2%

Appendix

Table 17: Question-level outcomes under A+H and refilled ACTIVE-only serving. A question is correct under the judge verdict for MedMemoryBench and ClinicalBench, or at token-F1

\geq 0.5

for MedLoCoMo and MIMIC-QA. History harmful is a subset of Active sufficient, so the four columns do not form a disjoint partition.

Dataset	Query type	$n$	1.7B write / 4B reader	GPT-5
MQA	first	37	2.7%	45.9%
MQA	minimum	15	0.0%	60.0%
MQA	maximum	11	0.0%	45.5%
MQA	at_timestamp	86	4.7%	45.3%
MQA	top_n	43	0.0%	0.0%
MMB	temporal localization	100	5.0%	15.0%

Appendix

Table 18: History-critical rates for selected query types under the refilled design: the fraction of questions answered correctly with A+H but not with refilled ACTIVE-only serving.

Dataset	Retrieval	Evidence statistic	Score
MedMemoryBench	Store-aware	History/pkt 44.8%	36.3
MedMemoryBench	Joint ranking	History/pkt 52.3%	35.1
MIMIC-QA	Store-aware	Gold served 97.4%	55.34
MIMIC-QA	Joint ranking	Gold served 98.7%	56.23

Appendix

Table 19: Retrieval-configuration comparisons using the same underlying memories and total retrieval budget. MedMemoryBench reports macro accuracy and MIMIC-QA reports mean Token-F1. The MIMIC-QA joint-ranking condition also uses neutral evidence rendering; Table 20 separates retrieval allocation from label rendering.

Comparison	From	To	$\Delta$	95% CI	$p$
Remove labels, store-aware	54.65	55.26	+0.61	$[-1.8,+3.0]$	.64
Add labels, joint ranking	56.80	55.54	-1.26	$[-3.9,+1.3]$	.35
Ranking change, both unlabelled	55.26	56.80	+1.54	$[-0.4,+3.5]$	.13
Ranking change, both labelled	54.65	55.54	+0.89	$[-1.3,+3.1]$	.43
Labelled Full $\rightarrow$ unlabelled merged	54.65	56.80	+2.15	$[-0.2,+4.5]$	.08

Appendix

Table 20: MIMIC-QA analysis separating retrieval allocation from store-label rendering. Scores are mean Token-F1 over 400 test questions.

\Delta

is the paired mean difference; 95% confidence intervals are obtained by paired bootstrap and

p

-values by a two-sided paired sign-flip permutation test.

Variant	Archived	Pair recall	False arch.	Impact ceiling
Full, semantic+graph, $B=48$	34.8	53.9	7.6	71.4
Semantic only, $B=48$	33.0	51.6	12.3	66.7
Semantic only, $B=24$	25.6	42.1	17.6	52.6
Rule: measurement/dose	6.8	16.4	8.6	N/A
Rule: all non-episodic concepts	20.6	28.6	17.6	N/A

Appendix

Table 21: Additional MMB write-path controls. Impact ceiling measures coverage of the learned impact-retrieval candidate set and is therefore not applicable to deterministic supersession rules.

Criterion	Store removed	Safe recall	Safe precision	NF loss
Strict protected loss	2.3%	16.1%	92.9%	0.3%
Never-forget only	14.7%	76.3%	69.4%	8.1%

Appendix

Table 22: Effect of the calibrated loss definition on deployment-store deletion. The never-forget-only criterion is shown as a diagnostic and is not used by the system.

Candidate source	Candidates	Safe recall	Safe precision	Store removed
Updater Delete proposals	535	10.2%	84.6%	0.05%
All archived memories	8,418	16.1%	92.9%	2.3%

Appendix

Table 23: Reported deletion diagnostics by candidate source. Both runs use the same deletion gate; the updater-proposal diagnostic covers 12 personas.

Store	Macro	Pooled	EEM	IG	MCD	MQ	SUA	TLA
Untouched	36.82	39.11	63.0	22.0	8.5	41.4	44.0	42.0
After deletion	37.78	40.32	63.0	24.0	4.3	41.4	48.0	46.0

Appendix

Table 24: Paired MedMemoryBench evaluation before and after strict deletion.

Dataset	Both		Active -only
	Tokens	TTFT	Tokens	TTFT
MedMemoryBench	4,482	228.7 ms	1,876	97.3 ms
MIMIC-QA	2,484	125.3 ms	1,044	59.9 ms
ClinicalBench	1,052	55.9 ms	653	41.9 ms
MedLoCoMo	1,680	90.9 ms	1,478	84.1 ms

Appendix

Table 25: Prompt length and TTFT. Values are means over the one-token prefill experiment.

Dataset	Completion tokens		Decode time
	Both	Active -only	Both	Active -only
MedMemoryBench	102.8	99.5	0.877 s	0.821 s
MIMIC-QA	53.2	52.8	0.432 s	0.419 s
ClinicalBench	189.5	146.5	1.550 s	1.187 s
MedLoCoMo	8.8	8.7	0.056 s	0.054 s

Appendix

Table 26: Decode behavior on the deterministic 100-question generation subset.

Dataset	Both	Active -only	Reduction
MedMemoryBench	38.5 TFLOP	14.7 TFLOP	61.9%
MIMIC-QA	19.9 TFLOP	7.9 TFLOP	60.2%
ClinicalBench	8.0 TFLOP	4.9 TFLOP	38.9%
MedLoCoMo	13.0 TFLOP	11.4 TFLOP	12.7%

Appendix

Table 27: Prefill FLOPs of the two serving arms. These values characterize the computational cost of presenting each context to the reader and are independent of a gate’s commit frequency.

Dataset	Both	Active -only	KV reduction	Prompt-only capacity
MedMemoryBench	630 MiB	264 MiB	58%	$41.6\rightarrow 99.3$
MIMIC-QA	349 MiB	147 MiB	58%	$75.0\rightarrow 178.4$
ClinicalBench	148 MiB	92 MiB	38%	$177.1\rightarrow 285.3$
MedLoCoMo	236 MiB	208 MiB	12%	$110.9\rightarrow 126.0$

Appendix

Table 28: KV-cache footprint and prompt-only capacity under the two serving actions. Capacity is computed from the measured 186,288-token cache and ignores generated tokens and 16-token block rounding; it is therefore an upper bound on simultaneous requests, not a throughput measurement.

Dataset	Record	Source tok.	Time
MedMemoryBench	shortest	647.6K	444 s
	median	668.1K	538 s
	longest	818.1K	595 s
ClinicalBench	shortest	3.4K	54 s
	median	15.1K	402 s
	longest	73.1K	2,128 s

Appendix

Table 29: Write-time construction cost. Panel (a) reports end-to-end wall-clock latency on representative records using Qwen3-1.7B on a single otherwise idle GPU. Panel (b) reports model compute for the published Qwen3-1.7B MedMemoryBench stores over the five evaluation personas, including atomic extraction, updating, and relation linking.

Explore similar work

PACMI: Provenance-Aware Cascading Memory Invalidation for Long-Term LLM Agents
Oct 5, 2026Yiqi Wang, Jiaqi Liu, Jiaqi Zhang +4
University of Southern Queensland · Southern University of Science and Technology · Jiangsu University +2
TransMem: Transforming Hidden States into Memory for Large Language Models
Jul 31, 2026Haodong Lei, Junming Liu, Yirong Chen +4Large Language Model Memory Large Language Model Agents
Southeast University, Nanjing, Jiangsu, China · Shanghai Artificial Intelligence Laboratory, Shanghai, China
Learning to Remember: End-to-End Training of Memory Agents for Long-Context Reasoning
Feb 13, 2026Kehao Zhang, Shangtong Gui, Sheng Yang +2Agentic Memory Long-Context Reasoning
1Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences (ICT/CAS) · University of Chinese Academy of Sciences, Beijing, China · 4Li Auto Inc. +1

Oct 5, 2026, cs.LGJ/K move · Enter open · S save

PACMI: Provenance-Aware Cascading Memory Invalidation for Long-Term LLM Agents

Yiqi Wang, Jiaqi Liu, Jiaqi Zhang, Zhangkai Wu+3

University of Southern Queensland · Southern University of Science and Technology · Jiangsu University+2

Open paper AskPDF arXiv

LLM agents rely on long-term memory to retain and reuse information when performing tasks over long horizons. Existing methods provide limited support for handling memories that become outdated as new observations or domain evidence arrive. Such outdated memories may remain semantically relevant, continue to affect dependent records, and retain value as historical evidence. This calls for two capabilities: dependency tracking to identify downstream effects and historical preservation to retain useful past records. We propose Provenance-Aware Cascading Memory Invalidation (PACMI), a framework that represents memories and new evidence in a provenance graph with typed dependency edges. PACMI assigns records to a four-state validity lattice, propagates validity changes to dependent memories, and uses the resulting states for retrieval and stale-premise detection. We also introduce a diagnostic benchmark with 100 cases and 300 queries across five domains. The evaluation separates node, context-, and answer-level performance. PACMI achieves the highest final-answer accuracy on this benchmark, and its paired difference from the strongest baseline is significant under an exact McNemar test. The premise checker achieves perfect precision, recall, and F 1 on the controlled query distribution. Cascading propagation primarily improves memorystate correctness: removing it increases final-answer errors from 3 to 11, but the paired difference does not reach the 0.05 significance threshold. Code and data will be made publicly available.

Abstract

Explore similar work

PACMI: Provenance-Aware Cascading Memory Invalidation for Long-Term LLM Agents

TransMem: Transforming Hidden States into Memory for Large Language Models

Learning to Remember: End-to-End Training of Memory Agents for Long-Context Reasoning