CAMR: Compatibility-Aware Memory Replay for Continual Learning in Large Language Models

Jingpu Yang; Tianhao Li; Fengxian Ji; Yufeng Wang; Shuxin Zheng; Xiaoqing Zhang

cs.CLOct 27, 2026

CAMR: Compatibility-Aware Memory Replay for Continual Learning in Large Language Models

Authors: Jingpu Yang, Tianhao Li, Fengxian Ji, Yufeng Wang, Shuxin Zheng, Xiaoqing Zhang

Organizations: Beijing Zhongguancun Academy · Northeastern University · Beihang University · Renmin University of China

Abstract

Continual learning requires large language models to acquire new knowledge and capabilities while retaining what they have learned. Fine-tuning and replay support acquisition and retention, but deciding whether incoming examples need supervision and which historical examples to revisit remains a challenge. We propose Compatibility-Aware Memory Replay (CAMR), an adaptive framework that jointly controls current-example supervision and historical replay. CAMR combines compatibility checks with a learned memory gate to guide an online policy over three actions: learning incoming examples with general replay, jointly reinforcing compatible related examples, and skipping direct supervision of verified mastered duplicates. The policy learns from acquisition, retention, and training-cost feedback under a bounded per-arrival replay budget. We evaluate CAMR with Qwen3-8B-Base on continual factual updating and cross-task and cross-domain learning. On WikiBigEdit, CAMR consistently leads evaluated methods in Original Match from 3K to 100K facts, reaching 93.66% after just 3K facts and surpassing Full-history Replay (LoRA) after 100K facts. It also achieves the highest final scores on all four selected TRACE tasks and maintains or improves earlier-task performance relative to post-acquisition levels in a sequence spanning financial question answering, mathematical reasoning, code generation, and summarization.

Figures & tables

Figure 1: Overview of CAMR and its empirical evaluation. (a) Schematic illustration of CAMR . It shows how Add , Merge , and Ignore handle incoming data during continual learning. (b) Example comparison of CAMR with Random Replay in the multi-task setting. (c) Results for single-task and cross-task settings, together with Replay-Count Scaling experiments.

Figure 2: Overview of CAMR . The writer retrieves and filters compatible history, then selects Add , Merge , or Ignore to form the training batch; feedback from the updated language model trains the writer.

Unmodified base models
Method	3K				10K				50K				100K (change vs. 3K)
Method	Original	Rephrase	Persona	Locality	Original	Rephrase	Persona	Locality	Original	Rephrase	Persona	Locality	Original	Rephrase	Persona	Locality
Qwen3-8B-Base	14.67	13.71	11.20	5.88	13.16	12.51	9.33	9.62	12.04	11.90	8.76	10.55	12.88 (-1.79)	12.33 (-1.38)	9.08 (-2.12)	13.09 (+7.21)
Qwen3-14B-Base	21.63	20.20	11.53	29.41	20.42	19.69	9.99	21.15	17.26	16.55	8.72	17.58	17.40 (-4.23)	16.76 (-3.44)	8.74 (-2.79)	18.75 (-10.66)
Fine-tuning and parameter-space methods
Fine-Tuning (LoRA)	51.40	47.83	38.57	21.97	56.83	52.34	44.61	21.95	48.27	43.10	34.70	24.62	41.08 (-10.32)	37.40 (-10.43)	28.23 (-10.34)	23.21 (+1.24)

Table 1: Long-stream match accuracy (%) and checkpoint change. All checkpoints belong to the same 100K continual stream. Parentheses at 100K report the change from 3K in percentage points. The light peach shading marks the highest value in each column, these values are also bold. The additional CAMR LoRA run uses

K=8

and a learning rate of

2\times 10^{-4}

through 50K updates and

10^{-4}

thereafter, continuing the same model and optimizer.

Figure 3: WikiFactDiff replacement results over the 36K-update stream, evaluated on the current stage every 3K updates. ES/GS report success rates, while ED/GD report new-minus-obsolete answer-probability gaps.

Fine-tuning and parameter-space methods
Method	T1	T2		T3			T4 (change from post-acquisition)
	Acc.	Acc.	ROUGE-L	Acc.	ROUGE-L	Edit	Acc.	ROUGE-L	Edit	Acc.
	C-STANCE	C-STANCE	MeetingBank	C-STANCE	MeetingBank	Py150	C-STANCE	MeetingBank	Py150	ScienceQA
Fine-Tuning (LoRA)	50.55	50.10	58.91	53.70	56.06	67.88	53.85 (+3.30)	56.30 (-2.61)	67.30 (-0.58)	94.25
Sequential Fine-Tuning (LoRA)	56.60	56.75	65.30	56.80	58.66	67.69	55.00 (-1.60)	53.85 (-11.45)	66.25 (-1.44)	91.95
Sequential Fine-Tuning (Full FT)	49.30	46.05	69.20	46.75	68.14	63.39	47.70 (-1.60)	68.97 (-0.23)	62.76 (-0.63)	92.60

Table 2: Continual learning performance (%) and final checkpoint change on the four-task TRACE stream. T4 parentheses report changes from the stage when each task was first learned. The light peach shading marks the highest value in each column, these values are also bold.

Method	T1	T2		T3			T4 (change from post-acquisition)
	Acc.	Acc.	Acc.	Acc.	Acc.	pass@1	Acc.	Acc.	pass@1	ROUGE-L
	FinQA	FinQA	GSM8K	FinQA	GSM8K	APPS	FinQA	GSM8K	APPS	SAMSum
Fine-tuning and parameter-space methods
Fine-Tuning (LoRA)	30.51	22.23	82.41	26.94	80.89	30.40	22.23 (-8.28)	70.05 (-12.36)	30.60 (+0.20)	43.79
Sequential Fine-Tuning (Full FT)	23.37	24.24	75.06	24.85	73.62	20.80	23.54 (+0.17)	70.66 (-4.40)	22.80 (+2.00)	42.25
O-LoRA	31.56	31.39	82.94	30.95	82.64	28.80	30.43 (-1.13)	83.55 (+0.61)	29.10 (+0.30)	44.59

Table 3: Continual learning performance (%) and final checkpoint change on the four-task cross-domain stream. T4 parentheses report changes from the stage when each task was first learned. The light peach shading marks the highest value in each column, these values are also bold.

$K$	Method	Original	Rephrase	Persona	Locality
4	CAMR (ours)	93.98%	86.56%	69.98%	18.34%
	Random Replay	93.62%	86.08%	70.74%	19.58%
	Semantic Replay	80.40%	73.36%	63.21%	11.91%
8	CAMR (ours)	96.70%	89.26%	72.62%	19.86%
	Random Replay	96.62%	89.08%	70.92%	19.30%
	Semantic Replay	82.62%	74.73%	65.70%	13.67%

Table 4: Replay strategies under different per-arrival replay caps

K

. All methods use full-parameter fine-tuning. Best scores within each

K

group are bold.

Variant	FOMC	NumGLUE-cm	NumGLUE-ds	Py150
CAMR (ours)	72.38	76.54	77.54	67.75
w/o Action Selection	71.25	74.07	74.57	66.89
w/o Online Feedback	71.25	76.54	75.69	67.58
w/o Memory Gate	70.63	74.07	72.85	66.65
w/o Ignore	71.88	74.07	76.63	66.63
w/o Replay on Ignore	71.25	72.84	73.54	66.49

Table 5: Writer ablation results under full-parameter fine-tuning. The light peach shading marks the highest score in each column, including ties.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Original	Rephrase	Persona	Locality
Full-history Replay (LoRA)	93.75	78.71	45.31	22.66
CAMR (Full FT, ours)	97.53	87.77	67.59	30.08

Appendix

Table 6: WikiBigEdit Match accuracy (%) after 150,000 continual updates. The best result in each column is shown in bold.

Method	3K	10K	50K	100K
Fact presentations (current + replay)
CAMR (Full FT)	14,801	49,516	246,364	494,424
Random Replay (Full FT)	14,984	49,984	249,984	499,984
Full-history Replay (LoRA)	13,500	141,625	3,440,625	13,682,125

Appendix

Table 7: Cumulative WikiBigEdit fact presentations, including current facts and historical replay.

Figure 4: Age- and eligibility-corrected semantic replay allocation across the WikiBigEdit stream. Rows show CAMR , count-matched Random Replay from CAMR ’s logged eligible history pools, count-matched cosine Top-8 Semantic Replay from the same pools, and the theoretical 375-record Full-history Replay at 3K, 10K, 50K, 100K, and 150K. Each panel reports local cumulative replay intensity divided by the exact event-wise Random Replay expectation after identical fixed-grid smoothing. Neutral gray denotes $1\!\times$ expectation, red denotes over-allocation, and blue denotes under-allocation on the shared log 2 scale. All checkpoints use the same frozen representation and PCA projection.

Stream	Size	Add	Merge	Ignore
WikiBigEdit	3K	67.83	32.17	0.00
	10K	59.55	40.45	0.00
	50K	71.75	28.25	0.00
	100K	69.73	30.27	0.00
APPS	5K	58.94	21.92	19.14

Appendix

Table 8: Cumulative writer-action shares (%) under the Ignore gate.

Figure 5: Unnormalized cumulative replay counts in the same semantic space. Rows and checkpoints follow Figure 4 . Each point is a replayed fact at its actual PCA coordinate, and color gives its cumulative replay count on a shared logarithmic scale. The shared axes, frozen representation, and PCA projection make semantic coverage and per-fact repetition directly comparable across methods and stream prefixes.

Hyperparameter	Value
Training and optimization
Base language model	Qwen3-8B-Base
Language-model update	Full-parameter SFT
Language-model optimizer	AdamW
Language-model learning rate	$2\times 10^{-6}$
Writer optimizer	Adam

Appendix

Table 9: Complete set of CAMR hyperparameters governing memory, replay, the writer architecture, and the learning objectives.

Method	3K	10K	50K	100K
CAMR	25.54	20.93	9.74	7.30
Random Replay	19.49	13.56	5.94	3.02

Appendix

Table 10: Share (%) of replay slots allocated to the fixed high-frequency core-3K set as the WikiBigEdit history grows.

Continual Learning

Continual Learning studies how models acquire knowledge from changing data streams while retaining or selectively revising what they learned before.

See all 307 papers

When to Review: Spaced Repetition for Continual Pre-Training of Language Models
Aug 18, 2026Alankar Atreya, Devesh Batra, Yoages Kumar Mantri +3Continual Learning for LLMs Continued Pretraining
NatWest AI Research · University College London
Self-Preservation in Language Models
Oct 8, 2026Dongkyu Cho, Rumi Chunara, Ali BehrouzContinual Learning for LLMs Replay-Based Continual Learning
New York University · Google Research · Cornell University
Towards Understanding Continual Factual Knowledge Acquisition of Language Models: From Theory to Algorithm
May 11, 2026Haoyu Wang, Yifan Shang, Zhongxiang Sun +3Continual Learning for LLMs Generative Replay
Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China · School of Artificial Intelligence and Data Science, University of International Business and Economics, Beijing, China

Aug 18, 2026J/K move · Enter open · S save

When to Review: Spaced Repetition for Continual Pre-Training of Language Models

Alankar Atreya, Devesh Batra, Yoages Kumar Mantri, Geremy Bantug+2

NatWest AI Research · University College London

Continual Learning for LLMs Continued Pretraining Replay-Based Continual Learning Catastrophic Forgetting

Open paper AskPDF arXiv

Continual pre-training of large language models must acquire new information without erasing old knowledge. Existing replay methods often choose a global old/new mixture and sample uniformly, ignoring that examples differ in how quickly they are forgotten. We formulate continual pre-training as adaptive review scheduling: the training loop should decide not only how much history to replay, but which examples should return at each step. We introduce Spaced Repetition Training (SRT), a continual learning framework inspired by cognitive science, which schedules sample-rehearsal using the SuperMemo-2 (SM-2) algorithm. SRT maintains per-example review state, maps per-example perplexity to a recall-quality signal, and schedules historical examples for retention and new examples for consolidation while leaving the model, objective, and optimizer unchanged. On temporally separated Wikipedia and code corpora, SRT improves the stability-plasticity trade-off, recovering 5 to 37 percentage points of old-knowledge accuracy lost by naive continual pre-training across model scales while preserving or improving new-knowledge acquisition. At larger scale, SRT preserves broad benchmark performance that naive continual pre-training and uniform replay substantially degrade. Experiments with vision and tabular data further suggest that the scheduling principle extends beyond language when paired with an appropriate recall signal.

Abstract

Continual Learning

Explore similar work

When to Review: Spaced Repetition for Continual Pre-Training of Language Models

Self-Preservation in Language Models

Towards Understanding Continual Factual Knowledge Acquisition of Language Models: From Theory to Algorithm