@inproceedings{banerjee-etal-2026-wk,
title = "{WK}-Drift: Benchmarking Procedural Memory Maintenance under World Knowledge Drift",
author = "Banerjee, Pratyay and
Moshtaghi, Masud and
Chadha, Ankit",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {IJCNLP}-{AACL} 2026",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-aacl.140/",
pages = "2719--2736",
ISBN = "979-8-89176-444-6",
abstract = "LLM agents that accumulate procedural memory can reuse successful plans, but what happens when the world changes? Deprecated APIs, revised clinical thresholds, and reorganized knowledge graphs silently invalidate stored procedures. No existing benchmark evaluates whether a memory system can \textit{maintain} its procedures under such drift. We introduce \textbf{WK-Drift}, a benchmark constructed from three public temporal datasets (Time-Sensitive QA, GitChameleon, DriftMed). WK-Drift formalizes drift as typed events with a fact-change/procedure-change taxonomy, defines a four-phase protocol (learn, drift, evaluate, recover), and introduces entity-plan linkage for targeted maintenance. We evaluate ten configurations spanning no-memory, unstructured (including Mem0 and MemRL), tagged, and graph-based approaches. Graph-based surgical updates achieve 90{\%} post-drift accuracy on knowledge QA (+6pp over stale baselines), 64{\%} on clinical (+26.5pp), and 57{\%} on code (+7.4pp), reducing staleness from 77{--}86{\%} to 41{--}42{\%}. The advantage concentrates on procedure-change tasks (+17.4pp) and scales with propagation depth (+50pp at depth{~}3). Surgical maintenance achieves 59{\%} from the first post-drift epoch versus 26{\%} for stale baselines, eliminating multi-epoch recovery latency."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="banerjee-etal-2026-wk">
<titleInfo>
<title>WK-Drift: Benchmarking Procedural Memory Maintenance under World Knowledge Drift</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pratyay</namePart>
<namePart type="family">Banerjee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Masud</namePart>
<namePart type="family">Moshtaghi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ankit</namePart>
<namePart type="family">Chadha</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-444-6</identifier>
</relatedItem>
<abstract>LLM agents that accumulate procedural memory can reuse successful plans, but what happens when the world changes? Deprecated APIs, revised clinical thresholds, and reorganized knowledge graphs silently invalidate stored procedures. No existing benchmark evaluates whether a memory system can maintain its procedures under such drift. We introduce WK-Drift, a benchmark constructed from three public temporal datasets (Time-Sensitive QA, GitChameleon, DriftMed). WK-Drift formalizes drift as typed events with a fact-change/procedure-change taxonomy, defines a four-phase protocol (learn, drift, evaluate, recover), and introduces entity-plan linkage for targeted maintenance. We evaluate ten configurations spanning no-memory, unstructured (including Mem0 and MemRL), tagged, and graph-based approaches. Graph-based surgical updates achieve 90% post-drift accuracy on knowledge QA (+6pp over stale baselines), 64% on clinical (+26.5pp), and 57% on code (+7.4pp), reducing staleness from 77–86% to 41–42%. The advantage concentrates on procedure-change tasks (+17.4pp) and scales with propagation depth (+50pp at depth 3). Surgical maintenance achieves 59% from the first post-drift epoch versus 26% for stale baselines, eliminating multi-epoch recovery latency.</abstract>
<identifier type="citekey">banerjee-etal-2026-wk</identifier>
<location>
<url>https://aclanthology.org/2026.findings-aacl.140/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>2719</start>
<end>2736</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T WK-Drift: Benchmarking Procedural Memory Maintenance under World Knowledge Drift
%A Banerjee, Pratyay
%A Moshtaghi, Masud
%A Chadha, Ankit
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-444-6
%F banerjee-etal-2026-wk
%X LLM agents that accumulate procedural memory can reuse successful plans, but what happens when the world changes? Deprecated APIs, revised clinical thresholds, and reorganized knowledge graphs silently invalidate stored procedures. No existing benchmark evaluates whether a memory system can maintain its procedures under such drift. We introduce WK-Drift, a benchmark constructed from three public temporal datasets (Time-Sensitive QA, GitChameleon, DriftMed). WK-Drift formalizes drift as typed events with a fact-change/procedure-change taxonomy, defines a four-phase protocol (learn, drift, evaluate, recover), and introduces entity-plan linkage for targeted maintenance. We evaluate ten configurations spanning no-memory, unstructured (including Mem0 and MemRL), tagged, and graph-based approaches. Graph-based surgical updates achieve 90% post-drift accuracy on knowledge QA (+6pp over stale baselines), 64% on clinical (+26.5pp), and 57% on code (+7.4pp), reducing staleness from 77–86% to 41–42%. The advantage concentrates on procedure-change tasks (+17.4pp) and scales with propagation depth (+50pp at depth 3). Surgical maintenance achieves 59% from the first post-drift epoch versus 26% for stale baselines, eliminating multi-epoch recovery latency.
%U https://aclanthology.org/2026.findings-aacl.140/
%P 2719-2736
Markdown (Informal)
[WK-Drift: Benchmarking Procedural Memory Maintenance under World Knowledge Drift](https://aclanthology.org/2026.findings-aacl.140/) (Banerjee et al., Findings 2026)
ACL