@inproceedings{inger-etal-2026-longmem,
title = "{L}ong{M}em-{RAT}: Long-Term Memory for Cross-Session Red Teaming",
author = "Inger, Nurit Cohen and
Haziz, Barak and
Cohen, Seffi and
Diga, Yonatan and
Chizi, Barak and
Rokach, Lior and
Shapira, Bracha",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.939/",
doi = "10.18653/2026.findings-emnlp.939",
pages = "19798--19815",
ISBN = "979-8-89176-462-0",
abstract = "Red-teaming has become a standard tool for probing alignment and safety weaknesses in large language models, but most existing methods treat each attack session in isolation and discard what was learned once the dialogue ends. We introduce **LongMem-RAT**, a multi-agent red-teaming framework equipped with long-term memory that accumulates, consolidates, and reuses insights from past attacks across sessions. This design enables cross-session generalization, reduces redundant exploration, and focuses new attacks on fragile alignment behaviors. Across multiple proprietary and open-weight models, LongMem-RAT improves Attack Success Rate (ASR) over a fully reproduced X-Teaming baseline by **21.73 percentage points on HarmBench and 18.40 points on AdvBench** (20.06 points on average). Memory ablations show that long-term memory improves ASR by 13.23 and 6.07 points and reduces Average Turns to Success (ATtS) by 0.37 and 1.92 turns on HarmBench and AdvBench, respectively, averaging a 9.65-point ASR gain and a 1.14-turn ATtS reduction."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="inger-etal-2026-longmem">
<titleInfo>
<title>LongMem-RAT: Long-Term Memory for Cross-Session Red Teaming</title>
</titleInfo>
<name type="personal">
<namePart type="given">Nurit</namePart>
<namePart type="given">Cohen</namePart>
<namePart type="family">Inger</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Barak</namePart>
<namePart type="family">Haziz</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Seffi</namePart>
<namePart type="family">Cohen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yonatan</namePart>
<namePart type="family">Diga</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Barak</namePart>
<namePart type="family">Chizi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lior</namePart>
<namePart type="family">Rokach</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bracha</namePart>
<namePart type="family">Shapira</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Red-teaming has become a standard tool for probing alignment and safety weaknesses in large language models, but most existing methods treat each attack session in isolation and discard what was learned once the dialogue ends. We introduce **LongMem-RAT**, a multi-agent red-teaming framework equipped with long-term memory that accumulates, consolidates, and reuses insights from past attacks across sessions. This design enables cross-session generalization, reduces redundant exploration, and focuses new attacks on fragile alignment behaviors. Across multiple proprietary and open-weight models, LongMem-RAT improves Attack Success Rate (ASR) over a fully reproduced X-Teaming baseline by **21.73 percentage points on HarmBench and 18.40 points on AdvBench** (20.06 points on average). Memory ablations show that long-term memory improves ASR by 13.23 and 6.07 points and reduces Average Turns to Success (ATtS) by 0.37 and 1.92 turns on HarmBench and AdvBench, respectively, averaging a 9.65-point ASR gain and a 1.14-turn ATtS reduction.</abstract>
<identifier type="citekey">inger-etal-2026-longmem</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.939</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.939/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>19798</start>
<end>19815</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T LongMem-RAT: Long-Term Memory for Cross-Session Red Teaming
%A Inger, Nurit Cohen
%A Haziz, Barak
%A Cohen, Seffi
%A Diga, Yonatan
%A Chizi, Barak
%A Rokach, Lior
%A Shapira, Bracha
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F inger-etal-2026-longmem
%X Red-teaming has become a standard tool for probing alignment and safety weaknesses in large language models, but most existing methods treat each attack session in isolation and discard what was learned once the dialogue ends. We introduce **LongMem-RAT**, a multi-agent red-teaming framework equipped with long-term memory that accumulates, consolidates, and reuses insights from past attacks across sessions. This design enables cross-session generalization, reduces redundant exploration, and focuses new attacks on fragile alignment behaviors. Across multiple proprietary and open-weight models, LongMem-RAT improves Attack Success Rate (ASR) over a fully reproduced X-Teaming baseline by **21.73 percentage points on HarmBench and 18.40 points on AdvBench** (20.06 points on average). Memory ablations show that long-term memory improves ASR by 13.23 and 6.07 points and reduces Average Turns to Success (ATtS) by 0.37 and 1.92 turns on HarmBench and AdvBench, respectively, averaging a 9.65-point ASR gain and a 1.14-turn ATtS reduction.
%R 10.18653/2026.findings-emnlp.939
%U https://aclanthology.org/2026.findings-emnlp.939/
%U https://doi.org/10.18653/2026.findings-emnlp.939
%P 19798-19815
Markdown (Informal)
[LongMem-RAT: Long-Term Memory for Cross-Session Red Teaming](https://aclanthology.org/2026.findings-emnlp.939/) (Inger et al., Findings 2026)
ACL
- Nurit Cohen Inger, Barak Haziz, Seffi Cohen, Yonatan Diga, Barak Chizi, Lior Rokach, and Bracha Shapira. 2026. LongMem-RAT: Long-Term Memory for Cross-Session Red Teaming. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 19798–19815, Budapest, Hungary. Association for Computational Linguistics.