MemoRAG: Moving towards Next-Gen RAG Via Memory-Inspired Knowledge Discovery

SUP-A02 原表

来源：作者原文。2026-09-07抓取，原始行列，不是统一复跑成绩。

Table 1. Main experiment results. Best results are in bold, second-best ones are underlined , and “ † \dagger ” indicates performance surpasses all baselines in a t-test at p < 0.05 p<0.05 . Evaluation metrics for all datasets are in Appendix B .

Dataset	nar	qas	mul	mus	2wiki	hot	news	gov	en.sum	en.qa	fin	legal	misc	ave.
	LongBench								InfBench		UltraDomain
Full	21.4	39.4	51.5	28.2	38.1	48.1	24.9	32.6	13.0	15.2	47.8	46.5	48.7	35.0
Mnference	20.7	39.0	50.8	27.4	35.9	46.2	24.8	32.2	13.3	12.1	44.7	39.8	46.3	33.3
SelfExtend	19.6	37.8	47.4	22.7	37.2	42.0	21.4	29.1	11.1	9.3	41.2	37.9	34.1	30.1
BGE-M3	20.3	33.0	44.3	21.1	35.4	42.1	17.7	19.8	9.6	16.3	41.7	41.2	43.7	29.7
Stella-v5	13.7	32.4	43.5	21.0	35.6	40.6	20.3	18.2	10.0	19.5	42.8	35.1	43.9	29.0
Jina-emb-v3	15.9	34.7	42.8	17.8	33.1	41.8	21.9	25.2	11.3	18.7	41.8	37.1	43.8	29.7
GraphRAG	16.2	36.3	45.4	19.3	37.5	38.0	18.4	25.6	10.8	13.5	39.9	39.6	41.7	29.4
RQ-RAG	19.6	34.1	46.5	21.9	36.1	41.7	20.1	18.6	10.4	16.1	41.8	40.9	43.2	30.1
HyDE	18.7	36.0	47.5	20.5	36.8	42.7	-	-	-	19.6	43.1	41.6	44.2	-
MemoRAG	27.5^†	43.9^†	52.2^†	33.9^†	54.1^†	54.8^†	26.3^†	32.9^†	15.7^†	22.9^†	51.5^†	51.0^†	55.6^†	40.2

Table 2. Case study on the Legal dataset. Predicted answers that overlap with the ground-truth answers are marked in teal .

Query: What is the significance of the Outside Date mentioned in the agreement? Context: A Legal Contract (56.4K tokens)

Ground-truth target: The Outside Date is the deadline by which the Plan must become effective, or else the Agreement will terminate automatically. It is set as October 5, 2020, at 11:59 p.m. Eastern Time.

Standard RAG: The Outside Date is significant as it is a date where both parties have agreed in advance that if the merger or acquisition has not yet completed either side. It is set as October 5, 2020. (F1-Score: 0.36)

Clues #1: Definition of the “Outside Date” in the agreement Clues #2: “Outside Date” means October 5, 2020 at 11:59 p.m. Eastern Time.

MemoRAG: The Outside Date mentioned in the agreement is October 5, 2020, at 11:59 p.m. Eastern Time. It is a significant date in the context of the agreement because it is the deadline for the Plan to become effective. If the Plan has not become effective by this date, certain parties may have the right to terminate the agreement. (F1-Score: 0.83)

Table 3. Statistical information of the datasets utilized in this paper.

Dataset	Narrative	Qasper	MultiField	Hotpot	MuSiQue	2Wiki
Num of Samples	200	200	150	200	200	200
Ave. Length	18,409	3,619	4,559	9,151	11,214	4,887
Metric	F1	F1	F1	F1	F1	F1
Dataset	GovReport	MultiNews	En.Sum	En.QA	Fin	Legal
Num of Samples	200	200	103	351	345	438
Ave. Length	8,734	2,113	171,500	192,600	40,625	51,413
Metric	Rouge-L	Rouge-L	F1	Rouge-L	F1	F1

Table 4. Statistical information of the out-of-domain evaluation datasets utilized in this paper.

Dataset	Num	$\max(\|{\mathcal{C}}\|)$	$\min(\|{\mathcal{C}}\|)$	$\text{ave}(\|{\mathcal{C}}\|)$	$\text{ave}(\|{\mathcal{Q}}\|)$	$\text{ave}(\|{\mathcal{A}}\|)$
Technology	240	306,073	44,549	144029.7	14.4	40.2
Biology	220	257,644	39,218	125284.9	16.8	49.1
Religion	220	1,071,342	34,257	131424.8	17.4	54.2
Fiction	220	564,980	44,057	137689.7	16.2	43.6
Psychology	200	571,725	37,988	150119.5	16.7	46.5
Music	200	381,043	51,517	168672.9	17.5	49.7
Art	200	305,001	32,793	128961.2	17.8	52.2
Philosophy	200	678,553	38,729	135682.7	17.2	51.0
Health	180	289,258	50,600	135902.0	16.2	48.2
History	180	688,074	53,277	195265.0	17.9	51.0
Literature	180	534,836	33,043	129363.7	16.9	47.0
Biography	180	408,969	45,052	163522.3	18.0	52.0
Politics	180	387,157	49,853	139624.3	17.9	54.9
Mathematics	160	726,144	60,936	197924.6	16.7	47.6
Physics	160	226,811	36,717	105805.6	14.8	54.2
Cooking	120	466,885	58,360	156139.2	16.5	46.6
Agriculture	100	385,915	76,581	150969.6	15.6	45.9
Computer	100	437,070	51,704	215929.5	14.3	39.8
Total	3,240	1,071,342	32,793	150684.0	16.6	48.5