TSUBASA: Improving Long-Horizon Personalization via Evolving Memory and Self-Learning with Context Distillation

Xinliang Frederick Zhang; Lu Wang

cs.CLApr 9, 2026

TSUBASA: Improving Long-Horizon Personalization via Evolving Memory and Self-Learning with Context Distillation

Authors: Xinliang Frederick Zhang, Lu Wang

Organizations: Computer Science and Engineering, University of Michigan, Ann Arbor, MI

Abstract

Personalized large language models (PLLMs) have garnered significant attention for their ability to align outputs with individual's needs and preferences. However, they still struggle with long-horizon tasks, such as tracking a user's extensive history of conversations or activities. Existing memory mechanisms often fail to capture evolving behaviors, and RAG paradigms are trapped by a quality-efficiency tradeoff. Meanwhile, parametric adaptation is bottlenecked by train-inference gap due to the scarcity of labeled data. To enhance the long-horizon capabilities of PLLMs, we introduce TSUBASA, a two-pronged approach designed to improve memory writing via dynamic memory evolution, and memory reading via self-learning with a context distillation objective to internalize user experiences. Extensive evaluations on long-horizon benchmarks using the Qwen-3 model family (4B to 32B) validate the effectiveness of TSUBASA, surpassing competitive memory-augmented systems that rely primarily on memory writing, such as Mem0 and Memory-R1. Our analyses further confirms that TSUBASA breaks the quality-efficiency barrier to achieve Pareto improvements, delivering robust, high-fidelity personalization with a reduced token budget.

Figures & tables

Figure 1: Overview of TSUBASA framework, built on two synergistic wings. Dynamic memory writing applies structured algorithmic evolution based on high-density observations distilled from raw utterances ( Section 4.1 ). Internalized memory reading adopts the self-learning pipeline and applies context distillation objective on synthetic data ( Section 4.2 ). During Inference , observations are retrieved for factual grounding.

Method	Qwen3 -8B			Qwen3 -14B			Input Len. (tokens)
Method	F1	BLEU-1	Rouge-L	F1	BLEU-1	Rouge-L	Input Len. (tokens)
RAG (k=1)	21.26	17.06	20.82	24.29	20.00	23.78	29.4
RAG (k=3)	28.10	22.53	27.30	31.38	26.02	30.59	75.4
Train	6.07	4.32	5.73	6.59	4.96	6.24	5.6
Session	41.22	33.64	39.61	44.26	36.40	42.87	2,158.5

Table 1: Preliminary study reveals train-inference gap and quality-efficiency tradeoff as discussed in Section 3 .

Method	Qwen3-4B			Qwen3-8B			Qwen3-14B			Qwen3-32B
Method	F1	BLEU-1	ROUGE-L	F1	BLEU-1	ROUGE-L	F1	BLEU-1	ROUGE-L	F1	BLEU-1	ROUGE-L
RAG (utterance)	28.53	23.20	27.66	28.10	22.53	27.30	31.38	26.02	30.59	28.64	23.19	27.83
RAG (observation)	31.17	25.50	30.31	31.77	25.91	30.63	34.83	28.79	33.43	34.00	28.00	32.56
[5pt/4pt] Vanilla Training	5.76	4.02	5.53	6.07	4.32	5.73	6.59	4.96	6.24	8.28	6.22	7.91
TSUBASA (no grounding)	12.26	9.50	11.84	11.99	9.33	11.67	14.71	11.90	14.41	14.81	11.27	14.37
[5pt/4pt] TSUBASA (static accumulation)	32.82	27.21	31.96	34.50	29.25	33.40	37.07	31.25	35.65	32.75	26.18	31.46

Table 2: Main results of TSUBASA and baselines on LoCoMo dataset (average of 3 runs), isolating module contributions. For each base model, the best results are in bold , while global optimal performances are u nderlined. The results demonstrate information density of observations, usefulness of internalized memory reading, necessity for factual grounding, and efficacy of evolving memory. Complete results are shown in Table A2 including efficiency.

Method	Single Hop			Multi-Hop			Open Domain			Temporal			Overall
Method	F1	BLEU-1	J	F1	BLEU-1	J	F1	BLEU-1	J	F1	BLEU-1	J	F1	BLEU-1	J
LoCoMo ( Maharana et al. )	9.57	7.00	15.06	11.84	10.02	19.28	8.67	6.52	12.79	8.35	8.74	5.43	8.97	7.27	12.17
A-Mem ( Xu et al. )	18.96	12.86	40.78	14.73	12.66	31.32	30.58	26.14	46.90	23.67	20.67	28.68	26.08	21.78	40.78
Mem0 ( Chhikara et al. )	24.96	18.05	61.92	20.31	15.82	48.19	32.74	25.27	65.20	33.16	26.28	38.76	30.61	23.55	53.30
Memory-R1-PPO ( Yan et al. )	34.22	23.61	57.74	32.87	29.48	53.01	44.78	38.72	66.99	42.88	30.30	42.25	41.72	33.70	59.53
Memory-R1-GRPO ( Yan et al. )	33.64	26.06	62.34	23.55	20.71	40.96	46.86	40.92	67.81	47.75	38.49	49.61	43.14	36.44	61.51

Table 3: Comparison of TSUBASA -PRO (average of 3 runs) with prior SOTA baselines. Results of non- TSUBASA baselines are taken from Yan et al. (2025) . Best numbers are in bold . Overall, TSUBASA -PRO consistently outperforms all baselines, including the highly competitive Memory-R1. The sole exception is the open-domain sub-task, where our performance is lower as the current pipeline is primarily optimized for factoid QA reasoning.

Figure 2: Quality-Efficiency tradeoff between input length and F1 metric on baselines and TSUBASA . Best-performing configuration is highlighted in circle with its gap to the establish ceiling ( Session ) indicated. Importantly, our TSUBASA achieves Pareto improvement : attaining higher peak performance while utilizing only a fraction of the context budget. Tradeoffs w.r.t. ROUGE-L and BLEU-1 are presented in Figure A1 and Figure A2 .

Figure 3: Impact of top- $d$ truncation for approximating KL divergence on distillation quality (F1) with TSUBASA variants. Best-performing config is highlighted in circle for each model size. Importantly, we find $d=1$ inadequate, while increasing $d$ beyond $5$ leads to performance degradation due to noises in long-tail distribution, establishing $d=5$ at the sweet spot. See Figure A3 and Figure A4 for impacts on ROUGE-L and BLEU-1.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Method	F1	BLEU-1	J
LoCoMo ( Maharana et al. )	18.27	14.57	22.20
A-Mem ( Xu et al. )	41.55	36.58	54.80
Mem0 ( Chhikara et al. )	38.44	34.53	46.80
Memory-R1-PPO ( Yan et al. )	40.30	35.50	47.40
Memory-R1-GRPO ( Yan et al. )	46.70	41.10	57.80
[5pt/4pt] TSUBASA -PRO (static acc.)	45.75	42.10	57.40

Appendix

Table A1: Comparison of TSUBASA -PRO (single run) with prior SOTA baselines on LongMemEval ( Wu et al., 2025a ) . Results of non- TSUBASA baselines are taken from Yan et al. (2025) . Best numbers are in bold . Overall, TSUBASA -PRO consistently outperforms all baselines, including the highly competitive Memory-R1.

Figure A1: Quality-Efficiency tradeoff between input length and ROUGE-L metric on baselines and TSUBASA . Best-peroforming configuration is highlighted in circle with its gap to the establish ceiling ( Session ) indicated. Importantly, our TSUBASA achieves Pareto improvement : attaining higher peak performance while utilizing only a fraction of the context budget.

Figure A2: Quality-Efficiency tradeoff between input length and BLEU-1 metric on baselines and TSUBASA . Best-peroforming configuration is highlighted in circle with its gap to the establish ceiling ( Session ) indicated. Importantly, our TSUBASA achieves Pareto improvement : attaining higher peak performance while utilizing only a fraction of the context budget.

Figure A3: Impact of top- $d$ truncation for approximating KL divergence on distillation quality (ROUGE-L) with TSUBASA variants. Best-performing configuration is highlighted in circle for each model size. Importantly, we find $d=1$ inadequate, while increasing $d$ beyond $5$ leads to performance degradation due to noises in long-tail distribution, establishing $d=5$ at the sweet spot.

Figure A4: Impact of top- $d$ truncation for approximating KL divergence on distillation quality (BLEU-1) with TSUBASA variants. Best-performing configuration is highlighted in circle for each model size. Importantly, we find $d=1$ inadequate, while increasing $d$ beyond $5$ leads to performance degradation due to noises in long-tail distribution, establishing $d=5$ at the sweet spot.

Model	Method	P	R	F1	BLEU-1	ROUGE-L	Input Len.
Qwen3-4B	RAG (utterance)	24.58	49.87	28.53	23.20	27.66	75.4
	RAG (observation)	27.25	51.20	31.17	25.50	30.31	57.1
	Vanilla Training	4.34	13.44	5.76	4.02	5.53	5.6
	TSUBASA (no grounding)	10.43	21.54	12.26	9.50	11.84	5.6
	TSUBASA (static accumulation)	29.14	50.42	32.82	27.21	31.96	57.1
	TSUBASA (evolving memory)	27.95	47.72	31.38	26.23	30.40	51.7

Appendix

Table A2: Complete results of TSUBASA and baselines on LoCoMo dataset (average of three runs), allowing for a holistic understanding of each module’s contribution. For each base model, the best results are highlighted in bold , while global optimal performances are u nderlined. The results demonstrate superior information density of observations, usefulness of internalized memory reading, necessity for factual grounding, and efficacy of evolving memory.

Panel of Figure 13 — Figure A5: Prompt used for extracting salient observations from a conversation session.

Panel of Figure 14 — Figure A6: Prompt used by our memory manager to autonomously evolve the episodic memory.

Panel of Figure 15 — Figure A7: A sample example illustrating the TSUBASA memory writing process. Raw utterances are first compressed into dense observations, which are subsequently merged by the memory manager (RECONCILE action in this particular example) to form a coherent, temporally grounded narrative. The colored text indicates the flow. For example, there are two observations that can be extracted from the truncated raw conversations, one in red and another in green . The sampled entry (index: 22) from our evolving memory is derived from observations in session 6 and session 28 . Note, this evolving memory piece is temporally grounded with absolute time markers.

Personalized Language Models

Personalized language models adapt responses to individual users’ preferences, histories, writing styles, and changing needs.

See all 89 papers

Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-horizon Agents
May 25, 2026Yeonjun In, Wonjoong Kim, Sangwu Park +2Memorization in Language Models Persistent Memory for Language Models
KAIST
AdaMem: Learning What to Remember with Adaptive Memory Policies for Personalized Agents
Jun 19, 2026Xingyu Chen, Rui Wang, Zhaopeng Tu +1LLM Agent Memory Persistent Memory for Language Models
Shanghai Jiao Tong University, Tencent · Shanghai Jiao Tong University · Tencent
AdaMem: Adaptive User-Centric Memory for Long-Horizon Dialogue Agents
Mar 17, 2026Shannan Yan, Jingchen Ni, Leqi Zheng +6LLM Agent Memory Long-Term Conversational Memory
Tsinghua University · WeChat Vision, Tencent Inc.

May 25, 2026J/K move · Enter open · S save

Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-horizon Agents

Yeonjun In, Wonjoong Kim, Sangwu Park, Kanghoon Yoon+1

KAIST

Memorization in Language Models Persistent Memory for Language Models Memory Agent Benchmarks Long-Horizon LLM Agents Agent Memory Management

Open paper AskPDF arXiv

Existing large language model (LLM) based memory systems apply universal, static policies that overlook a fundamental reality: the contexts that are worth storing in memory are different across users. This misalignment wastes limited memory budget on transient interactions while failing to preserve critical context for long horizon tasks. To address this gap, we investigate an underexplored question: can LLM based memory systems learn personalized memory policies? We introduce PerMemBench, the first benchmark for evaluating personalized memory systems, featuring multi year, multi domain interaction histories across diverse user personas. We further present the first empirical study of memory personalization, proposing session level storage gating, a lightweight framework that selectively bypasses memory operations for transient sessions. Our study confirms that personalization yields substantial retention gains under perfect gating, yet reveals that accurate gating remains an open and critical challenge.

Abstract

Personalized Language Models

Explore similar work

Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-horizon Agents

AdaMem: Learning What to Remember with Adaptive Memory Policies for Personalized Agents

AdaMem: Adaptive User-Centric Memory for Long-Horizon Dialogue Agents