Mem++: Non-Destructive Memory for Long-Term Organizational LLM Agents

Ahmad Yehia; Aly O. Abdelkareem; Islam Ahmed; Hesham Omran; Khaled Alashmouny; Christian Claudel; Abduallah Mohamed

cs.CLOct 1, 2026

Mem++: Non-Destructive Memory for Long-Term Organizational LLM Agents

Authors: Ahmad Yehia, Aly O. Abdelkareem, Islam Ahmed, Hesham Omran, Khaled Alashmouny, Christian Claudel, Abduallah Mohamed

Organizations: The University of Texas at Austin · AIDAChip Inc.

Abstract

Large Language Model (LLM) agents now take part in organizational work, where many authors record decisions across documents over months. Because a revised decision arrives as a new document rather than an edit, answering a question requires knowing which version held at a given time. However, most memory systems compress the record at write time. By distilling each document into facts, notes or graph edges, these methods fix what can be answered before any question is asked. To address this, we propose Mem++, a non-destructive memory framework shifting from write-time distillation to read-time selection. Mem++ stores every document whole with its date and author, and it calls no generative model at write time. At read time, it retrieves only documents dated up to the time a question asks about and fuses lexical and semantic rankings. Unlike systems that overwrite older versions, Mem++ keeps them and leaves the choice to the answering model. Evaluations on the organizational benchmark OrgMemBench demonstrate that Mem++ surpasses the strongest memory system baseline by 8.0 to 13.1 points across two answering models. With gpt-4.1-mini, it also achieves the best overall score, 2.6 points above RAG. In addition, Mem++ achieves the best average LLM-judge score on LoCoMo and ranks second on LongMemEval-S, behind only its entity-graph variant. Code for benchmark evaluation is available at https://github.com/AIDAChip-Inc/mem-plus-plus.

Figures & tables

Figure 1: Conversational vs. organizational memory. Top: a single narrator restates their own facts, and LLM extraction at ingest keeps only the surviving fact. Bottom: different authors in an organization write the same fact, and an earlier decision may still hold.

Figure 2: Overview of Mem++ framework. (3.2) Each document is stored whole with its author, date and sentence embedding, and it is indexed three ways. (3.3) Records valid at the as-of date $\theta$ are retrieved from the three indexes and fused with weighted RRF, and three of the $k$ slots are reserved for the latest-dated matches. (3.4) The top- $k$ records are passed with their dates and authors to a fixed answering model. Mem g ++ graph triples carry no date.

Figure 3: The six capabilities of OrgMemBench ( Gardner, 2026 ) , with questions and answers examples. The tier we evaluated holds 443 artifacts across 157 threads for 18 months, with 73 questions.

	Method	Supersession	Decision Provenance	Bi-temporal	Audit Replay	Justification Chain	Contradiction	Overall
gpt-4.1-mini	Full Context	0.0	27.5	0.0	33.0	22.7	8.3	17.8 $\pm$ 0.1
	RAG	66.1	79.7	66.7	40.9	19.6	75.4	55.0 $\pm$ 0.3
	Zep	31.4	42.5	23.0	76.8	20.7	43.3	41.0 $\pm$ 0.2
	Mem0	40.8	55.0	50.0	25.1	14.4	52.8	36.9 $\pm$ 0.0
	A-Mem	58.3	63.3	43.6	36.5	20.7	43.5	44.5 $\pm$ 0.3
	gbrain	54.8	60.7	30.3	48.5	22.7	22.9	43.2 $\pm$ 0.1

Table 1: Performance on OrgMemBench ( Gardner, 2026 ) categorized by question type. Overall is weighted by the number of questions in each category and reported with its standard deviation across runs. Bold indicates the best performance. Underline indicates the second best performance.

	Method	Temporal Reasoning			Open Domain			Multi-Hop			Single-Hop			Average
	Method	$\uparrow$ LLM	$\uparrow$ F1	$\uparrow$ BLEU	$\uparrow$ LLM	$\uparrow$ F1	$\uparrow$ BLEU	$\uparrow$ LLM	$\uparrow$ F1	$\uparrow$ BLEU	$\uparrow$ LLM	$\uparrow$ F1	$\uparrow$ BLEU	$\uparrow$ LLM	$\uparrow$ F1	$\uparrow$ BLEU
gpt-4.1-mini	Full Context	74.2	47.5	40.0	56.6	28.4	22.2	77.2	44.2	33.7	86.9	61.4	53.4	80.6	53.3	45.0
	RAG-2048	66.8	50.3	40.1	48.6	28.3	21.7	67.4	39.1	28.9	82.8	57.7	48.1	74.5	50.9	41.3
	RAG-4096	27.4	22.3	19.1	28.8	17.9	13.9	31.7	20.1	12.8	35.9	25.8	22.0	32.9	23.5	19.2
	Zep	60.2	23.9	20.0	43.8	24.2	19.3	53.7	30.5	20.4	66.9	45.5	40.0	61.6	36.9	30.9
	Mem0	56.9	39.2	33.2	47.9	23.7	17.7	68.2	40.1	30.3	71.4	48.6	42.0	66.3	43.5	36.5

Table 2: Performance on LoCoMo ( Maharana et al., 2024 ) categorized by type. Bold, underline indicates best and second best performance. The baseline performance comes from Nan et al. (2025) .

	Question Type	Full Context	Zep	Nemori	Mem++	Mem g ++
gpt-4o-mini	Single-session preference	6.7	20.0	46.7	46.7	50.0
	Single-session assistant	89.3	80.4	83.9	94.6	96.4
	Temporal reasoning	42.1	62.4	61.7	56.7	56.7
	Multi-session	38.3	57.9	51.1	63.4	62.0
	Knowledge update	78.2	83.3	61.5	84.3	85.2
	Single-session user	78.6	92.9	88.6	98.4	98.4

Table 3: Performance on LongMemEval

{}_{\text{S}}

( Wu et al., 2024 ) by question type. LLM-judge accuracy is reported. Bold indicates the best performance. Underline indicates the second best performance.

Variant	OrgMemBench	LoCoMo	LongMemEval ${}_{\text{S}}$
Mem++	57.0	85.7	87.9
w/o vector leg	21.3 ( $-$ 35.7)	76.7 ( $-$ 9.1)	11.9 ( $-$ 76.0)
w/o lexical leg	57.2 ( $+$ 0.2)	85.2 ( $-$ 0.5)	88.7 ( $+$ 0.7)
w/o tag leg	55.0 ( $-$ 2.0)	–	–
w/ consolidation	55.6 ( $-$ 1.3)	85.4 ( $-$ 0.3)	86.9 ( $-$ 1.1)
w/ fact index	55.2 ( $-$ 1.7)	83.1 ( $-$ 2.6)	–

Table 4: Ablation with claude-sonnet-4-6 as the answerer. Parentheses give the change from Mem++.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

OrgMemBench (medium)		LoCoMo		LongMemEval S
Category	Count	Category	Count	Category	Count
C1 Supersession	15	Single-hop	841	Knowledge update	72
C2 Provenance	15	Multi-hop	282	Multi-session	121
C3 Bi-temporal	7	Temporal	321	Temporal reasoning	127
C4 Audit replay	15	Open domain	96	Single-session user	64
C5 Justification	15	–	–	Single-session assistant	56

Appendix

Table 5: Category distributions of the OrgMemBench ( Gardner, 2026 ) , LoCoMo ( Maharana et al., 2024 ) and LongMemEval S ( Wu et al., 2024 ) datasets.

Setting	OrgMemBench	LoCoMo	LongMemEval ${}_{\text{S}}$
Embedding model	all-MiniLM-L6-v2	text-embedding-3-small	text-embedding-3-small
Embedding size	384	1536	1536
Weights $(w_{\text{lex}},w_{\text{tag}},w_{\text{vec}})$	(1, 1, 4)	(1, 1, 2)	(1, 1, 4)
Candidates per index $C$	50	200	50
Lexical matching	conjunctive	disjunctive	conjunctive

Appendix

Table 6: Retrieval settings that differ across benchmarks.

Figure 4: OrgMemBench score of Mem++ as the number of retrieved rows $k$ varies. The dashed line is Full Context, and the dotted line marks $k=50$ used in the main experiments.

$k$	gpt-4o-mini	gpt-4.1-mini	Retrieved tokens
5	36.57	43.99	0.20M
10	45.10	50.56	0.41M
15	47.16	54.44	0.61M
30	45.94	56.49	1.22M
50	44.23	57.00	2.01M
70	48.67	56.70	2.77M

Appendix

Table 7: OrgMemBench score of Mem++ for each

k

. Retrieved tokens are summed over all questions and counted before the 40,000-character cut.

Explore similar work

Infini Memory: Maintainable Topic Documents for Long-Term LLM Agent Memory
Jun 9, 2026Suozhao Ji, Baodong Wu, Zehao Wang +8Long-Term Agent Memory Long-Term Memory
1Infinigence AI · 2Tsinghua University · 3Shanghai Jiaotong University
MEME: Multi-entity & Evolving Memory Evaluation
May 12, 2026Seokwon Jung, Alexander Rubinstein, Arnas Uselis +2Peak Memory Memes
1KAIST AI · 2Tübingen AI Center, University of Tübingen · 3NAVER AI Lab
Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability
Jul 29, 2026Sizhe Zhou, Sheldon Yu, Hui Wei +8Large Language Model Agents Peak Memory
University of Illinois Urbana-Champaign · University of California, San Diego · University of California, Merced +2

Jun 9, 2026, cs.AIJ/K move · Enter open · S save

Infini Memory: Maintainable Topic Documents for Long-Term LLM Agent Memory

Suozhao Ji, Baodong Wu, Zehao Wang, Lei Xia+7

1Infinigence AI · 2Tsinghua University · 3Shanghai Jiaotong University

Long-Term Agent Memory Long-Term Memory

Open paper AskPDF arXiv

Long-term LLM agents need persistent memory that can track changing facts and provide relevant evidence across sessions. Existing memory systems often store observations as isolated records, summaries, or indexed fragments, which makes evidence aggregation, fact revision, and memory maintenance difficult. We propose Infini Memory, a maintainable text-based persistent memory architecture that treats agent memory as topic-structured documents. Each topic document serves as a semantic unit for collecting related evidence, preserving metadata, and revising facts over time. New observations are first staged in a buffer and periodically consolidated into coherent textual contexts. At inference time, an agentic retrieval procedure lets the LLM read memory through iterative tool calls rather than a single retrieval step. On MemoryAgentBench, Infini Memory achieves 64.7% overall score. Ablations show that topic-structured maintenance and iterative evidence inspection improve complementary aspects of long-term memory use.

Abstract

Explore similar work

Infini Memory: Maintainable Topic Documents for Long-Term LLM Agent Memory

MEME: Multi-entity & Evolving Memory Evaluation

Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability