TEM: Temporal State Tracking Memory for Large Language Model Agents

Pengfeng Song; Yunkai Zhuang; Liye Shi

cs.CLOct 15, 2026

TEM: Temporal State Tracking Memory for Large Language Model Agents

Authors: Pengfeng Song, Yunkai Zhuang, Liye Shi

Organizations: School of Information and Intelligent Engineering,Zhejiang Wanli University,Ningbo,315100,China.

Abstract

An agent's memory must remain useful as information changes across conversations. A new fact may replace an earlier value without making that value irrelevant to a later question. We present Temporal State Tracking Memory (TEM), a training-free framework for tracking these changes. TEM stores facts as versioned records linked to their sources. It retrieves relevant versions, places them in source order and labels historical records before passing the evidence to the answer model. Event memory and original dialogue turns provide details missing from structured facts. We evaluated TEM on 420 questions drawn from 600 sessions involving 60 personas. Across four open-weight models, key-entity hit rates reached 75.71-97.62% overall and 96.67-100.00% on the historical-residence subset. Adding historical labels to otherwise identical, source-ordered facts increased overall hit rates by 17.14-55.95 percentage points. The evidence was compact: 94-118 tokens per answer, or 48.75-56.73% less than session-aware retrieval-augmented generation (RAG). A further evaluation on 1,540 LoCoMo questions tested hybrid memory under a fixed evidence budget. TEM exceeded session-aware RAG with both backbones; conservative score differences were 7.92 and 10.58 percentage points. The results support preserving fact versions and exposing their temporal status to the generator when answering questions about repeatedly updated information.

Figures & tables

Figure 1: TEM framework. Fact versions are linked to source sessions. Retrieved facts appear in source order, with historical labels on superseded versions. Dashed connections supply event and original-text evidence.

Figure 2: Fact-version updates and evidence construction. a, Appending a new version updates the status of retained records. b, Retrieval selects relevant facts, which are then ordered by source and labelled if historical. Symbols v and s denote a fact version and its source session.

Backbone	Unmarked (%)	TEM (%)	Gain (pp)
Llama-3.1-8B	19.76	75.71	+55.95
Qwen3.5-4B	48.10	97.62	+49.52
Qwen3.5-9B	74.52	91.67	+17.14
gpt-oss-20b	55.24	96.67	+41.43

Table 1: Effect of historical labels on overall key-entity hit rate

Method	Llama-3.1-8B	Qwen3.5-4B	Qwen3.5-9B	gpt-oss-20b
Naive RAG	14.52	42.38	14.05	46.43
Session-aware RAG	56.67	57.62	79.76	95.24
Full context	49.76	95.48	87.38	66.43
Overwrite	43.33	29.76	41.43	34.76
Retain, unmarked	19.76	48.10	74.52	55.24
TEM-Status	75.71	97.62	91.67	96.67

Table 2: Overall key-entity hit rates across four backbones and six configurations (%)

Figure 3: Effects of version retention and historical labels. a, Historical-residence hit rates (60 questions). b, Overall gains over Unmarked (420 questions). pp denotes percentage points.

Method	Llama-3.1-8B	Qwen3.5-4B	Qwen3.5-9B	gpt-oss-20b
Naive RAG	0.00	0.00	0.00	5.00
Session-aware RAG	10.00	88.33	20.00	86.67
Full context	58.33	96.67	73.33	91.67
Overwrite	0.00	8.33	0.00	0.00
Retain, unmarked	41.67	65.00	81.67	76.67
TEM-Status	98.33	100.00	100.00	96.67

Table 3: Historical-residence retrieval after repeated updates (key-entity hit rate, %)

Backbone	Unmarked hits	TEM hits	Net difference
Llama-3.1-8B	83	318	+235
Qwen3.5-4B	202	410	+208
Qwen3.5-9B	313	385	+72
gpt-oss-20b	232	406	+174

Table 4: Overall hit counts with and without historical labels

Figure 4: Selected successful corrections across backbones. Arrows show source-session order; bold states give the reference values. Check marks and crosses indicate correct and incorrect answers. The control configuration is identified in each panel. Panel c marks omitted text with […].

Backbone	Other queries Unmarked / TEM	Gain (pp)	Pre-move residence Unmarked / TEM	Gain (pp)
Llama-3.1-8B	16.11 / 71.94	+55.83	41.67 / 98.33	+56.67
Qwen3.5-4B	45.28 / 97.22	+51.94	65.00 / 100.00	+35.00
Qwen3.5-9B	73.33 / 90.28	+16.94	81.67 / 100.00	+18.33
gpt-oss-20b	51.67 / 96.67	+45.00	76.67 / 96.67	+20.00

Table 5: Key-entity hit rates and label gains by query group

Figure 5: Answer-evidence length distributions, with 420 queries per configuration. Boxes show the interquartile range, centre lines the median, and whiskers the minimum and maximum.

Backbone	Session-aware RAG	Full context	TEM-Status	Saving (%)
Llama-3.1-8B	214.72	180.13	95.93	55.32
Qwen3.5-4B	229.36	199.52	99.25	56.73
Qwen3.5-9B	229.36	199.52	117.55	48.75
gpt-oss-20b	206.55	171.95	94.05	54.47

Table 6: Mean evidence length during answering

Backbone	Memory	Overall	Temporal	Multi-hop	Single-hop	Open-domain
Llama-3.1-8B	TEM-Hybrid-State	27.53–29.22	14.02–17.76	17.73–18.09	37.57–38.88	13.54–15.62
Llama-3.1-8B	TEM-Hybrid-Unmarked	27.40–28.96	14.33–17.45	14.89–15.25	38.17–39.48	13.54–15.62
Llama-3.1-8B	Session-aware RAG	17.92–19.61	8.72–11.84	12.06–12.77	23.90–25.33	13.54–15.62
Llama-3.1-8B	Long-context ref.	37.66–39.55	14.64–17.13	24.11–24.82	53.51–55.65	15.62–16.67
Qwen3.5-9B	TEM-Hybrid-State	35.84–37.53	38.63–42.06	22.70–23.05	41.38–42.81	16.67–18.75
Qwen3.5-9B	TEM-Hybrid-Unmarked	35.97–37.66	38.32–41.43	25.18–25.89	40.78–42.21	17.71–19.79

Table 7: Strict semantic scores on LoCoMo for two backbones (%)

Method	Generator	Benchmark	Metric	Overall	Temporal	Single-hop
A-MEM Xu et al. (2025)	GPT-4o-mini	LoCoMo	F1	—	45.85	44.65
MemoryOS Kang et al. (2025)	GPT-4o-mini	LoCoMo	F1	—	20.02	35.27
Mem0 Chhikara et al. (2025)	GPT-4o-mini	LoCoMo	J	66.88 $\pm$ 0.15	55.51 $\pm$ 0.34	67.13 $\pm$ 0.65
Mem0-graph Chhikara et al. (2025)	GPT-4o-mini	LoCoMo	J	68.44 $\pm$ 0.17	58.13 $\pm$ 0.44	65.71 $\pm$ 0.45
Zep Rasmussen et al. (2025)	GPT-4o-mini	LongMemEval-s	GPT-4o judge	63.80	54.10	—
Zep Rasmussen et al. (2025)	GPT-4o	LongMemEval-s	GPT-4o judge	71.20	62.40	—

Table 8: Published results and evaluation settings of representative memory methods

LLM Agent Memory

Large language model agent memory lets agents retain, organize, retrieve, and update information across interactions and long-running tasks.

See all 373 papers

TransMem: Transforming Hidden States into Memory for Large Language Models
Jul 31, 2026Haodong Lei, Junming Liu, Yirong Chen +4Memory-Augmented Language Models LLM Agent Memory
Southeast University, Nanjing, Jiangsu, China · Shanghai Artificial Intelligence Laboratory, Shanghai, China
Remember by Asking: Retrieval-Induced Memory Evolution for LLM Agents
Sep 28, 2026Wanqi Zhou, Jiawei Lu, Yang Wang +4LLM Agent Memory Persistent Memory for Language Models
Chang’an University
Temporal Evidence Forests for Long-Horizon Multi-Party Conversational Memory
Oct 9, 2026Junyou Guo, Chongyang Tao, Yan He +2LLM Agent Memory Conversational Memory
School of Computer Science and Engineering, Beihang University, Beijing, China · Alibaba Group

Jul 31, 2026J/K move · Enter open · S save

TransMem: Transforming Hidden States into Memory for Large Language Models

Haodong Lei, Junming Liu, Yirong Chen, Pinlong Cai+3

Southeast University, Nanjing, Jiangsu, China · Shanghai Artificial Intelligence Laboratory, Shanghai, China

Memory-Augmented Language Models LLM Agent Memory Persistent Memory for Language Models Memory-Augmented Language Model Agents Memory Management for LLM Agents

Open paper AskPDF arXiv

Large language model (LLM) agents increasingly operate over long interaction histories, where effective reasoning requires identifying and exploiting task-relevant evidence distributed across past observations and actions. However, useful information encoded in previously computed representations is often underutilized during subsequent generation. We propose TransMem, a lightweight inference-time parametric memory module that transforms sparse historical hidden states from a frozen LLM backbone into reusable memory representations. TransMem uses a lightweight gating network to dynamically apply the latent intervention to the current hidden states, without repeatedly encoding the preceding context. To learn transferable memory utilization rather than task-specific knowledge, we introduce evidence-conditioned self-distillation. A memory-augmented student processes the full context and matches the predictive distribution of an evidence-only teacher that shares the same frozen backbone. Experiments on LoCoMo, HotpotQA, and MemoryAgentBench demonstrate consistent improvements across different model architectures and scales. TransMem yields gains of 11.58--29.25

F_1

on LoCoMo and 10.20--13.03

F_1

on HotpotQA, while improving the average MemoryAgentBench accuracy from 29.54% to 40.00%. These results establish sparse historical hidden states as an effective and efficient memory substrate for long-context LLM agents. Our code is available at https://github.com/Haodong-Lei-Ray/TransMem.

Abstract

LLM Agent Memory

Explore similar work

TransMem: Transforming Hidden States into Memory for Large Language Models

Remember by Asking: Retrieval-Induced Memory Evolution for LLM Agents

Temporal Evidence Forests for Long-Horizon Multi-Party Conversational Memory