RAG / 检索增强生成

Retrieval-augmented generation (RAG) systems depend critically on how documents are chunked and searched. Fine-grained chunks can improve retrieval precision but expand the search space, increasing latency and cost; larger chunks reduce the number of candidates but make dense similarity less reliable, as the representation for each chunk mixes multiple topics and introduces more semantic noise. This trade-off becomes especially limiting in deep research tasks, where retrieval must be both fast and precise across large, heterogeneous corpora. We introduce MCompassRAG, a metadata-guided retrieval framework that uses topic-level signals as a semantic compass for selecting relevant evidence. Instead of relying only on cosine similarity between queries and noisy chunk embeddings, MCompassRAG enriches chunk representations with topic metadata in the same embedding space and trains a lightweight retriever through LLM-teacher distillation. At inference time, MCompassRAG performs topic-aware retrieval without additional LLM calls, improving both efficiency and evidence quality. Across six complex retrieval benchmarks, MCompassRAG improves information efficiency (IE) by 8.24% on average with over 5 times lower latency than the strongest efficient RAG baselines. Code is available on https://github.com/AmirAbaskohi/MCompassRAG.

Dense retrieval ranks one query vector against one document vector. On long documents, this interface can fail when a short but decisive span is weakened during document encoding before ranking. We study this failure mode as document-side early compression and introduce the Evidence Dilution Index (EDI) to measure how far a document-level representation falls below the strongest chunk-level evidence within the same gold document. Guided by this view, we propose DICE (Document Inference via Chunk Evidence), a training-free document-side strategy that splits documents into chunks, encodes them independently with a frozen model, and aggregates them back into a single vector while preserving the standard one-query-one-document interface. On LongEmbed, DICE improves retrieval across four backbones, with the largest gains on slices beyond 4k tokens: for Dream, Passkey >4k rises from 30.0 to 90.0 and Needle >4k from 23.3 to 74.0. Across 12,779 filtered samples, DICE yields lower EDI than the single-vector baseline in 92.8% of cases. These results establish document-level encoding as a practical and underexplored lever for long-document retrieval.

Deployable multilingual rerankers must generalize across languages, domains, and target ranking tasks while remaining efficient enough for second-stage reranking. However, adapting them to new target distributions typically requires extensive task-specific relevance annotations, which are costly to obtain. We present Querit-Reranker, a family of multilingual cross-encoder rerankers trained with a data-centric pipeline for label-efficient adaptation. We instantiate it as Querit-Reranker-A0.4B, initialized from an in-house MoE backbone with 0.4B activated parameters, and Querit-Reranker-4B, initialized from Qwen3-Embedding-4B. Our pipeline first learns general relevance modeling from large-scale ranking-oriented data, then adapts to target distributions through synthetic-query mining with teacher scores as continuous soft labels. To consolidate complementary task-adapted strengths, we further merge checkpoints via spherical linear interpolation, obtaining a single deployable model without runtime ensembling overhead. Using Qwen3-Embedding-0.6B as the shared first-stage retriever, Querit-Reranker-A0.4B improves average nDCG@10 from 54.11 to 59.28 on BEIR and from 59.87 to 67.70 on MIRACL. On MTEB Multilingual v2 Reranking, it also substantially outperforms larger embedding-based baselines, while Querit-Reranker-4B further achieves state-of-the-art performance among publicly available models. We release both models on Hugging Face.

Production LLM agents increasingly depend on real-time search, yet native search grounding bundles retrieval policy, provider choice, evidence injection, cost, latency, and generation behavior behind a single model-provider boundary. This coupling makes grounding hard to inspect, tune, reuse, or port, and can trigger Search-Induced Verbosity that breaks strict output contracts. We present Decoupled Search Grounding (DSG), a vendor-agnostic boundary that moves grounding outside the reasoning model through an MCP-compatible gateway, exposing provider routing, source-aware context rendering, configured fallback, retrieval-depth control, and exact plus semantic caching as first-class controls. Across five frontier models on SimpleQA, FreshQA, and HotpotQA, native search leads on recency-sensitive FreshQA, but DSG exposes a stronger frontier when control matters: on SimpleQA it nearly matches native accuracy (86.1% vs. 87.7%) at 91% lower search cost, preserves concise answer contracts, and reaches a 99.4% warm-cache hit rate with 68% lower latency. Deployed as a shared production grounding layer for large-scale agentic workloads with interchangeable models, DSG matches or slightly exceeds native-search accuracy on an e-commerce query-understanding (QIU) workload while cutting search cost by over 98%. Real-time grounding is best treated as an optimizable interface boundary, not a fixed model feature.

Learned sparse retrieval (LSR) models such as SPLADE have traditionally used BERT-style masked language models as backbone encoders. A natural expectation is that replacing BERT with stronger pretrained encoders should improve retrieval effectiveness. However, we find that under standard SPLADE training recipes, backbones with large MLM-head L2 norms can suffer performance degradation and even training collapse under standard SPLADE training recipes. We identify this failure as a scale mismatch in the MLM head: SPLADE directly uses MLM-head outputs to construct sparse lexical representations, and query-document relevance is computed by an unnormalized dot product over these representations. As a result, an inflated MLM-head scale can amplify sparse activations, distort matching scores, and destabilize contrastive training under common training settings. To address this issue, we introduce a simple initialization-time correction that rescales the MLM-head projection by a constant factor before SPLADE training. This zero-cost adjustment improves training stability without modifying the model architecture or training objective. Across both in-domain and out-of-domain retrieval benchmarks, this simple correction substantially improves large-norm backbones such as ModernBERT and Ettin, turning unstable training runs into competitive sparse retrievers. In several settings, the corrected models further match or surpass the classic BERT-SPLADE baseline. These findings suggest that the bottleneck in adapting pretrained encoders to LSR is not encoder capacity alone, but the calibration of the MLM-head scale used to construct sparse lexical representations.

Personalized dialogue agents require continuous long-term memory to maintain coherent interactions across multiple sessions. However, deploying these capabilities on consumer-grade hardware (e.g., 8 GB VRAM edge devices) introduces severe memory and compute bottlenecks. Existing systems typically rely on isotropic cosine similarity for retrieval and heuristic rules for context compression. These approaches lack a unified theoretical foundation, frequently suffering from the hubness problem in high-dimensional retrieval and syntactic fragmentation during compression. To overcome these limitations, we propose CoreMem, a resource-efficient edge-cloud memory architecture fundamentally unified by information geometry. First, Riemannian retrieval replaces cosine matching with a locally adaptive Fisher-Rao metric, effectively penalizing hub memories via Mahalanobis distance with O(Ndr) Woodbury acceleration for real-time search. Second, Fisher-guided discrete token distillation (FDTD) introduces a hierarchical sentence-to-token compression mechanism. It derives sensitivity scores from Fisher information traces, providing a principled compression-KL tradeoff augmented with explicit structural syntax protection. Evaluated on the LOCOMO and LongMemEval-S benchmarks, CoreMem achieves strong accuracy improvements, yielding substantial gains in Open-domain (+4.51 pp) and Temporal (+4.17 pp) reasoning. Extensive profiling confirms that CoreMem operates seamlessly within a strict 8 GB VRAM budget, successfully bridging the gap between resource-constrained edge devices and the demand for theoretically grounded, lifelong memory agents.

Injecting malicious knowledge into retrieval-augmented generation (RAG) systems can manipulate retrieved evidence and mislead downstream generation, posing a serious security threat for AI applications. Existing RAG injection attacks mainly rely on manipulating external knowledge bases, such as crafting malicious corpus. However, the synthetic text crafted by such data-centric methods could be detectable, leading to the failure of attacks. Beyond corpus manipulation, open-source retrievers are increasingly exposing RAG systems to model-centric attacks. In this paper, we propose conflict-aware retriever editing, i.e., CAREATTACK, a model-centric retriever attack framework for malicious knowledge injection in RAG. Specifically, CAREATTACK consists two stages of conflict-aware retriever editing and attack-preserving anchor repair. Conflict-aware retriever editing adapts efficient closed-form parameter editing to the dense retrieval model, promoting malicious knowledge above benign competing passages and resolving potential parameter conflicts through graph-based conflict detection and parameter editing projection. Then, attack-preserving anchor repair performs lightweight calibration on the edited retriever to further eliminate the impact on non-target prompts while preserving the attack effectiveness for target prompts. We instantiate CAREATTACK on Qwen3-Embedding-0.6B and BGE-M3, and conduct evaluation on three benchmark datasets. Experimental results demonstrate our method substantially promote malicious passages into the retrieved knowledge of RAG systems and can perform attacks for batches of target prompts and passages, given the access of retrieval model parameters. Since most RAG systems are built upon open-source retrieval models, this work reveals a practical attack surface in RAG systems. Codes are public accessible at https://anonymous.4open.science/r/CareAttack-3F1C.

Deep research agents are increasingly evaluated on their ability to search for evidence, reason over retrieved sources, and produce grounded answers. Existing browsing benchmarks, however, largely assume that the user's query and the supporting evidence are written in the same language, leaving open whether agentic search systems can operate when relevant evidence appears in another language. We introduce XBCP (Cross-lingual BrowseComp-Plus), a controlled benchmark that preserves the English question-and-answer space of BrowseComp-Plus but varies the languages of the supporting documents. XBCP instantiates two complementary settings: in the cross-lingual setting, each query is paired with evidence in a single assigned language. In the multilingual setting, the full evidence corpus is distributed equally and randomly across 12 languages spanning high-resource and low-resource regimes. We evaluate four deep research agents using sparse and dense multilingual retrievers, measuring answer accuracy, evidence recall, search behavior, calibration, citation fidelity, and oracle retrieval. Results reveal substantial degradation when evidence is translated. Even strong, dense retrievers lose evidence recall, and agents become less calibrated and cite evidence less reliably. Notably, accuracy remains lower even when all gold evidence is supplied directly. These findings suggest that cross-lingual deep research exposes both retrieval failures and an independent, agent-side difficulty in integrating language-mismatched evidence.

With the rapid expansion of massive multilingual corpora, Multilingual Information Retrieval (MLIR) has emerged as a critical technology for global information access. MLIR enables users to retrieve semantically relevant documents from multilingual text collections using a single-language query. However, recent multilingual dense retrieval models often exhibit a strong preference for documents in the same language as the query. This leads to severe language bias, where top-ranked results are dominated by documents of specific languages, even when documents in other languages contain more semantically relevant information. To address this issue, we propose SHIFT, a training-free method applicable in the indexing stage. Specifically, SHIFT utilizes parallel translation pairs to estimate a relative language vector for each target language with respect to a source language. Subsequently, SHIFT corrects the language-specific offset by subtracting this relative language vector from document embeddings during indexing. Our comprehensive evaluation across four MLIR benchmarks and diverse dense retrieval models confirms that SHIFT can effectively mitigate language bias and enhance MLIR performance.

Search agent benchmarks exemplified by BrowseComp have rapidly saturated over the past year, with the strongest models surpassing 90% accuracy. Since these benchmarks are predominantly human-authored, annotators lack a global perspective on entity statistics and cannot systematically maximize search space size and structural complexity. This creates a difficulty ceiling that is hard to break. To address this, we introduce LoHoSearch (Long-Horizon Search Agents), a challenging benchmark comprising 544 human-verified questions across 11 domains. LoHoSearch is constructed via an automated pipeline built upon a knowledge graph covering over 7 million Wikipedia entities, which selects relations with large search spaces and assembles them into structurally complex questions with KG-verified unique answers. Our evaluation demonstrates that even the strongest model achieves only 34.74% accuracy, and existing context management strategies (best +6.8%) yield far smaller gains than on prior benchmarks. LoHoSearch provides a more demanding standard for evaluating long-horizon reasoning and context management in search agents.

Recommender systems have a wide area of application, e.g. in fields like video streaming, social media, or digital marketplaces. But, for a recommender-system, finding the right algorithm with the right hyperparameters is a reoccurring challenge. There is no one-fits-all solution, since the performance of one algorithm can vary immensely on different data sets. Due to the challenges of finding the right algorithm and the broad use of recommender-systems, it is of interest to create an Automated Recommender System (AutoRecSys) that takes on the task of finding the right algorithm-hyperparameter-combination for a given data set. In this work, we present the enhancement of LensKit-Auto, a framework introduced by Vente et al., that solves exactly this task of finding a fitting algorithm-hyperparameter-combination. LensKit-Auto's biggest strength lies in its ease of use, where it operates as a black-box, into which the user can feed their data set and receive the information of which algorithm and hyperparameters work best on this data set. In this work, we bring LensKit-Auto up to date, so that it works with the new version of its underlying framework, LensKit. We also implement further functionalities, such as the Tree Parzen Estimator as an additional optimization method, the ability to reuse the found algorithm, updated documentation, and the ability to visualize the optimization process. We also adapt an existing meta-learning framework to generate a suitable meta-dataset for LensKit-Auto, which could enable the integration of meta-learning into LensKit-Auto in the future. The presented changes bring LensKit-Auto up to date and enhance its usability, so that even non-experts in the field can find the right algorithm for their use case.

For strings $u, D \in Σ^*$, a subsequence embedding of $u$ in $D$ is a function $e \colon \{1, 2, \ldots, |u|\} \to \{1, 2, \ldots, |D|\}$ with $e(i) < e(i+1)$ for every $i \in \{1, 2, \ldots, |u|-1\}$ and the $i$-th symbol of $u$ equals the $e(i)$-th symbol of $D$. A gap-constraint for $u$ is a triple $(i, j, L)$ with $1 \leq i < j \leq |u|$ and $L$ is a regular language over $Σ$. An embedding $e$ satisfies a gap-constraint $(i, j, L)$ if the factor of $D$ strictly between positions $e(i)$ and $e(j)$ is a word from $L$. We investigate the subsequence matching problem with gap-constraints, which is relevant in the context of complex event recognition (CER): given $u, D \in Σ^*$ and a set $C$ of gap-constraints, find an embedding of $u$ in $D$ that satisfies all gap-constraints from $C$. In general, subsequence matching is NP-complete and the only known tractable variants restrict the interval structure of the gap-constraints. In this work, we show that we can solve subsequence matching with gap-constraints with an arbitrary interval structure rather efficiently (in fact, optimally under SETH) in time $O(|D| (|u| + |C|))$ if the gap-constraint languages satisfy a property which we dub left-convexity: whenever $u v w \in L$ and $v \in L$, then also $uv \in L$. Left-convex languages are sufficiently expressive to model interesting real-world scenarios considered in CER, e.g., length constraints $L = \{w \mid a \leq |w| \leq b\}$ for $a, b \in \mathbb{N}$. We also show how our algorithm can be used in order to efficiently enumerate all satisfying embeddings, which is particularly relevant for possible applications in CER. Finally, we show how non-left-convex languages can lead to intractability, i.e., if in addition to length constraints we allow $\{aa, ε\}$ as the only non-left-convex constraint language, then the problem is NP-complete again.

Domain-camouflaged injection attacks embed malicious instructions in retrieved content using domain-appropriate vocabulary, evading standard detectors that rely on syntactic injection markers. When detection fails, practitioners need to know which defense architectures reduce attack success. We evaluate five prompting-based defenses (spotlighting, paraphrasing, prompt sandwiching, and two combinations) against domain-camouflaged injection across three model families (Claude Haiku, Llama 3.1 8B, Gemini 2.0 Flash) and three deployment domains (financial, legal, general) using 3,510 trials. Paraphrasing retrieved content before agent processing is the most consistently effective defense in this benchmark, reducing camouflage attack success rate by 55-84\% depending on model, and achieves lower attack success rates than our Llama Guard 4 configuration on every model tested. Defense effectiveness is strongly model-dependent: spotlighting halves attack success on Claude Haiku but provides no benefit on Llama 3.1 8B. Financial domain deployments face the highest residual risk at 26-33\% baseline attack success rate, with no prompting-based defense fully eliminating the threat on weaker models. These results provide the first systematic evaluation of prompting-based defenses specifically against camouflage-class injection attacks and establish benchmark-based recommendations for practitioners. All tasks use synthetically constructed professional documents; whether these benchmark rankings generalize to real enterprise documents remains an open question.

Retrieval-Augmented Generation (RAG) systems are emerging as a key approach for grounding Large Language Models (LLMs) in external knowledge, addressing limitations in factual accuracy and contextual relevance. However, there is a lack of empirical studies that report on the development of RAG-based implementations grounded in real-world use cases, evaluated through general user involvement, and accompanied by systematic documentation of lessons learned. This paper presents five domain-specific RAG applications developed for real-world scenarios across governance, cybersecurity, agriculture, industrial research, and medical diagnostics. Each system incorporates multilingual OCR, semantic retrieval via vector embeddings, and domain-adapted LLMs, deployed through local servers or cloud APIs to meet distinct user needs. A web-based evaluation involving a total of 100 participants assessed the systems across six dimensions: (i) Ease of Use, (ii) Relevance, (iii) Transparency, (iv) Responsiveness, (v) Accuracy, and (vi) Likelihood of Recommendation. Based on user feedback and our development experience, we documented twelve key lessons learned, highlighting technical, operational, and ethical challenges affecting the reliability and usability of RAG systems in practice.

With the rise of large language models (LLMs), they have become instrumental in applications such as Retrieval-Augmented Generation (RAG). Yet evaluating these systems remains bottlenecked by the time and cost of building specialized assessment datasets. We introduce KNIGHT, an LLM-based, knowledge-graph-driven framework for generating multiple-choice question (MCQ) datasets from external sources. KNIGHT constructs a topic-specific knowledge graph, a structured and parsimonious summary of entities and relations, that can be reused to generate instructor-controlled difficulty levels, including multi-hop questions, without repeatedly re-feeding the full source text. This knowledge graph acts as a compressed, reusable state, making question generation a cheap read over the graph. We instantiate KNIGHT on Wikipedia/Wikidata while keeping the framework domain- and ontology-agnostic. As a case study, KNIGHT produces six MCQ datasets in History, Biology, and Mathematics. We evaluate quality on five criteria: fluency, unambiguity (single correct answer), topic relevance, option uniqueness, and answerability given the provided sources (as a proxy for hallucination). Results show that KNIGHT enables token- and cost-efficient generation from a reusable graph representation, achieves high quality across these criteria, and yields model rankings aligned with MMLU-style benchmarks, while supporting topic-specific and difficulty-controlled evaluation.

Vision-Language Models (VLMs) remain prone to hallucinations, producing fluent but visually unfaithful outputs. Existing chain-of-thought and retrieval-augmented methods only partially address this, as they neither enforce step-level citation grounding nor route verification failures back to retrieval for correction. We present CaVe-VLM-CoT, a modular reflection-based agentic-RAG framework that enforces evidence-grounded reasoning through a five-stage closed-loop pipeline: Extractor, Retriever, Solver, Citation Injector, and Verifier, in which detected ungrounded claims trigger structured feedback to the Extractor for targeted re-retrieval. Since no existing framework jointly measures retrieval quality, step-wise citation faithfulness, and cross-modal grounding, we propose a suite of 23 component-wise metrics across all stages, anchored by CaVeScore, a composite metric weighting accuracy, citation precision and recall, attribution, and evidence grounding. Without any architectural or prompt modifications, CaVe-VLM-CoT achieves 87.1\% accuracy and 56.6\% CaVeScore on ScienceQA , and 55.2\% accuracy and 35.7\% CaVeScore on MMMU (30 subjects).

Abstractive summarization plays a crucial role in enabling efficient understanding of scientific literature, yet it inherently demands both linguistic fluency and factual faithfulness. Existing approaches often fail to reconcile these two requirements. Extractive methods rely on rigid sentence splicing that disrupts macro-level logical coherence, while large language model (LLM)-based generative approaches, despite mastering linguistic fluency, exhibit limited factual consistency. In this work, we propose ScholarSum, a hierarchical reflective graph-based framework that emulates a student-teacher writing process for fluent and faithful scientific summarization. ScholarSum first organizes the document into a hierarchical knowledge graph by segmenting it into semantically coherent units, whose multi-layered community structure captures global logic and macro-level themes. Guided by this global structure, the student generates an initial draft, which is subsequently refined through fine-grained evidence retrieval. To ensure factual consistency, a teacher-like reviewer then iteratively examines the draft, identifies unsupported content, and prompts targeted re-retrieval and rewriting until the summary meets rigorous quality standards. Extensive experiments demonstrate that ScholarSum significantly outperforms previous baselines in terms of both completeness and faithfulness. Our code is available at https://github.com/Xiaoyu-Tao/ScholarSum.

Retrieval-augmented generation (RAG) systems must balance retrieval granularity with contextual coherence, a challenge that existing methods address through LLM-guided chunking, single-level context expansion, or hierarchical summarization. These approaches variously depend on costly LLM calls during indexing or retrieval, limit context aggregation to a single granularity level, or introduce information loss through summarization. We present SproutRAG, an attention-guided hierarchical RAG framework that addresses this trade-off by organizing sentence-level chunks into progressively larger but semantically coherent units, using learned inter-sentence attention to construct a binary chunking tree. Unlike prior approaches that rely on external LLMs, fixed context expansion, or lossy summarization, SproutRAG learns which attention heads and layers best capture semantic document structure, enabling multi-granularity retrieval without additional LLM calls or compressed summaries. At retrieval time, SproutRAG uses hierarchical beam search to retrieve candidates at multiple granularities, capturing multi-sentence relevance beyond flat retrieval. The framework is trained end-to-end with a joint objective that improves both embeddings and tree structure. Experiments across four benchmarks spanning scientific, legal, and open-domain settings demonstrate that SproutRAG improves information efficiency (IE) by 6.1% on average over the strongest baseline. Code is available on https://github.com/AmirAbaskohi/SproutRAG.

Retrieval-augmented generation (RAG) allows large language models to access external and private corpora for factual, domain-specific responses. Modern RAG pipelines use hierarchical navigable small world (HNSW) vector databases for efficient similarity search. When a user requests data deletion, the systems typically only mark the record as deleted, leaving the embedding on disk physically unchanged. This soft-delete operation raises compliance concerns under data-erasure and retention requirements such as GDPR Article 17 and HIPAA. Analysis on three HNSW implementations confirms that deleted vectors remain physically recoverable by accessing the raw index files at the storage layer, bypassing API access. Using the Vec2Text inversion model without domain-specific fine-tuning, we show this vulnerability on multiple real-world datasets and data modalities. On Wikipedia biographical living persons dataset (BLP), we successfully recover 25.5% of exact person names and 46.4% of geographic locations (ROUGE-L 0.185). Recovery reaches 100% for both patient age and gender markers (ROUGE-L 0.290) on highly structured, sensitive data (NIH Synthea dataset). On soft-deleted image embeddings, we show 100% tissue classification on histopathology patches (p=1.02e-07) and top-1 identity recovery reaches 99% on facial embeddings (p<0.01). This work introduces Epoch Key Rotation, which encrypts vectors and discards the key upon deletion. Epoch key rotation reduces observed PII recovery to 0% and completes in 2.5 ms for 500 deleted vectors (approximately 0.005 ms/record). Additionally, it generates an ECDSA-signed cryptographic proof as an auditable record of the deletion event.

Intent-based recommender systems have gained significant attention for improving accuracy and interpretability by modeling the underlying motivations behind user behaviors. Most existing models derive intents directly from user sequences via clustering or prototype learning. However, they are sensitive to sequence quality, require presetting the number of intents, and lack explicit semantic grounding. These issues lead to an incomplete and coarse intent set and limit the effectiveness of recommendation. In this paper, we propose the Sparse Autoencoder for intent-based recommendation (SAERec), a novel recommender that automatically constructs a fine-grained and interpretable intent space from a textual corpus to guide recommendation. Rather than treating texts as side signals, SAERec leverages them as high information density evidence for intent construction. Specifically, we first extract a comprehensive set of fine-grained interpretable intents from the latent space of large language models (LLMs) by using a sparse autoencoder (SAE) to disentangle and interpret text embeddings, which isolates intent-related semantics from textual noise. Then, for each user, we retrieve relevant intents from this set as priors to guide recommendation. It contains personal intents matching a user's current interests and public intents capturing general item patterns shared across users (e.g., quality, price). Finally, to integrate retrieved intents into sequence modeling, we propose a multi-branch attention mechanism that captures temporal dependencies and injects both personal and public intent signals, followed by an adaptive fusion layer to construct the final user representation for recommendation. Extensive experiments on public datasets demonstrate the superiority of SAERec, consistently outperforming state-of-the-art baselines while providing human-understandable explanations.

Large language model (LLM) agents have achieved strong performance on a wide range of benchmarks, yet most evaluations assume static environments. In contrast, real-world deployment is inherently dynamic, requiring agents to continually align their knowledge, skills, and behavior with changing environments and updated task conditions. To address this gap, we introduce EvoArena, a benchmark suite that models environment changes as sequences of progressive updates across terminal, software, and social domains. We further propose EvoMem, a patch-based memory paradigm that records memory evolution as structured update histories, enabling agents to reason about environmental evolution through changes in their memory. Experiments show that current agents struggle on EvoArena, achieving an average accuracy of 39.6% across evolving terminal, software, and social-preference domains. EvoMem consistently improves performance, yielding an average gain of 1.5% on EvoArena and also improving standard benchmarks such as GAIA and LoCoMo by 6.1% and 4.8%. Beyond individual tasks, EvoMem further improves chain-level accuracy by 3.7% on EvoArena, where success requires completing a consecutive sequence of related evolutionary subtasks. Mechanistic analysis shows that EvoMem improves evidence capture in the memory, indicating better preservation of complete evolving environment states. Our results highlight the importance of modeling evolution in both evaluation and memory for reliable agent deployment.