Large Language Model Memory

Large Language Models (LLMs) are increasingly expected to operate over long contexts, yet standard softmax attention incurs a KV cache that grows linearly with sequence length, quickly becoming the bottleneck for long context inference. A practical remedy is to evict less important KV entries; however, existing eviction policies are largely heuristic and struggle to capture the rich, input-dependent distribution of token importance. In this work, we introduce a learnable indexer that predicts KV importance, enabling more accurate retention of critical tokens. Meanwhile, naively evicting tokens permanently discards their information, leading to irreversible forgetting and degraded retrieval over long ranges. To address this, we propose a lightweight latent memory module that compresses evicted tokens into a compact, online-updated state and provides residual readouts to compensate for the attention contributions lost through KV eviction. Collectively, our method enables accurate long-context inference under a bounded KV budget, delivering consistent improvements on RULER (4K/16K) across Qwen, Mistral, and Llama models (up to 25 points under aggressive eviction), markedly more stable Needle-in-a-Haystack retrieval, and superior LongBench scores and compression curves compared to existing eviction policies.

Long-horizon LLM inference turns the key--value (KV) cache into the dominant GPU memory consumer and makes per-token attention increasingly expensive. Many common eviction policies use static recency windows or historical attention, leaving unused a signal computed on every decoding step: the model's current uncertainty. We introduce CONF-KV, a KV-cache manager that converts the next-token distribution into a scalar confidence score and uses it to choose the per-step cache budget, retaining more context when the model is uncertain and pruning aggressively when it is confident. Within each budget, tokens are ranked by a composite of accumulated attention mass and recency, while a protected recent window preserves local coherence. We combine the policy with blockwise online-softmax attention, mixed FP16/INT8 storage, and a pyramidal per-layer budget variant. Across four model families and generated lengths up to 4K, CONF-KV stays near the footprint of a fixed 512-token sliding window while remaining within 1.5--2.1 perplexity points of full KV. On Needle-in-a-Haystack up to 32K tokens, CONF-KV reaches 91.4% retrieval accuracy versus 53.8% for sliding windows and 80.6% for H2O; on 75 VisualWebArena tasks it retains 95.3% of full-KV success at 2.8 times lower peak memory.

Long-context memory systems often fail under fixed budgets, but end-to-end evaluation does not reveal whether evidence was discarded during compression or preserved but never retrieved. We introduce a four-condition diagnostic protocol that evaluates a fixed reader under truncated full context (TFC), oracle evidence (OE), complete stored memory (CSM), and retrieved memory (RM). Under this fixed-budget LongMemEval setup, write-side gaps exceed retrieval-side gaps for most tested baselines, with four of six baselines robustly write-dominant under our default diagnosis margin. Motivated by this diagnosis, we propose Expected Predictive Compression (EPC), which moves the key decision--what information to retain--to write time by using an LLM to anticipate likely future questions and preserve the minimal supporting evidence under the token budget, while leaving retrieval unchanged at question time. Across all 500 LongMemEval questions with three readers (GPT-5.2, Claude Sonnet 4, Gemini 2.5 Pro), EPC achieves the highest CSM scores among all systems (0.49 vs. 0.44 for Summary (LLM), the strongest baseline), reducing Delta_write to 0.04 while leaving Delta_retr comparable to other LLM-based systems. These results suggest that, on this benchmark and evaluation setup, improving what the write stage preserves is a key avenue for performance gains in the tested systems.

Existing language model applications struggle to meet the demand for emotionally oriented support, primarily due to their inability to maintain deep, persistent context across sessions. This report introduces Psych LM, an iOS application that validates the thesis that, for such applications, the surrounding architecture is paramount. Psych LM runs a local, on-device language model within a purpose-built, local-first runtime designed for behavioral and life-coaching applications. The system achieves the practical effect of a near-infinite context window through an automated, user-inspectable memory corpus that converts conversations into structured memory cards, including facts, goals, and events, and dynamically injects them into the prompt via semantic and vector search. As such, the system can be defined as an active-learning, retrieval-augmented generative, on-device architecture. This architecture delivers four primary contributions: a local-first design where privacy is a core property; a detailed description of the memory corpus for persistent context of key user information; a deterministic orchestration layer that provides a stable behavioral spine independent of the model's internal state; and a benchmark framework focused on evaluating the integrated system's reliability under realistic operating conditions. The R and D process confirms that complex, context-aware interaction can be reliably achieved under the strict constraints of a mobile environment by prioritizing architectural control and resource management over simple model size.

Citation counts remain the dominant metric for assessing research impact, yet they suffer from well-documented limitations: temporal lag, disciplinary bias, and Matthew effects. Here we propose LLM-Metrics, a research-impact assessment metric derived from the parametric memory of large language models (LLMs). The central hypothesis is that high-impact papers receive greater exposure in the academic community, that this exposure enters LLM training data in textual form, and that models consequently form stronger parametric memory of these papers. We designed four types of multiple-choice probes, covering title recognition, author recognition, method recognition, and venue recognition, and evaluated 549 computer science papers published in 2023-2024 across 17 LLMs spanning 0.5B to 72B parameters from six vendors. Of the 17 models, 15 produced positive predictions, 9 of which were significant at p less than 0.05, with an overall Spearman correlation of rho = 0.1495 and p = 0.0004 against citation counts. Three additional findings support the proposed mechanism. First, the predictive signal was stronger for 2024 papers, rho = 0.1880, whose citation counts were near zero at model-training time, reducing the plausibility of a simple reverse-causality explanation. Second, author-recognition probes showed the strongest discriminative power, consistent with an exposure-driven memory mechanism. Third, model scale and predictive power were non-monotonic: a 3B-parameter model, Llama-3.2-3B-Instruct, with rho = 0.1829, outperformed most larger models, supporting a selective-memory hypothesis in which the limited capacity of smaller models can serve as an effective information filter. LLM-Metrics offers a real-time, cross-disciplinary, citation-independent paradigm for research assessment.

As Large Language Models (LLMs) evolve into persistent scientific collaborators, context window saturation has emerged as a critical bottleneck. Scientific workflows involving iterative data analysis and hypothesis refinement rapidly saturate even extended contexts with dense technical content, while monolithic approaches suffer from quadratic cost scaling and cognitive degradation. We evaluate a Dual Process Memory Architecture that decouples immediate episodic needs (constant 10-message window) from long-term consolidated knowledge (growing at approximately 3 tokens/message). Unlike prior social agent memory systems, our domain-specific consolidation addresses contradictory parameter evolution, multi-hop reasoning across experimental phases, and precise technical fact retention. Through large-scale evaluation spanning 15,000 messages with cross-model validation across six LLMs from three families (OpenAI, Anthropic, Google), totaling 1,440 queries, we establish three key findings. First, while full-context models fail at 10,000 messages due to context overflow, our system maintains 70-85% accuracy with 1-2 second latency using 62% fewer tokens (45,434 vs 120,000+ limit). Second, cross-model validation reveals architecture-level trade-offs independent of specific LLMs: Dual Process excels at numeric/temporal queries (65-90% accuracy) while RAG excels at historical retrieval (60-85%), suggesting complementary deployment strategies. Third, we identify a "Sim-to-Real" gap where synthetic tests maintain constant memory but realistic workflows exhibit linear growth (about 3 tokens/message), with consolidation quality emerging as the primary scalability bottleneck. The architecture successfully manages profiles with 14,000+ scientific facts (125k tokens), demonstrating that domain-specific memory consolidation enables sustained operation beyond full-context limits.

Recent studies introduce conditional memory modules that decouple knowledge storage from neural computation, enabling more direct knowledge access. Compared to MoE, which relies on dynamic computation paths, explicit lookup provides a more efficient knowledge retrieval mechanism. However, these approaches still depend on learned memory embeddings, requiring additional training and limiting flexibility. To address this, we propose N-gram Memory (NGM), a training-free, plug-and-play module composed of a Causal N-Gram Encoder and a Cosine-Gated Memory Injector. The Causal N-Gram Encoder directly averages the pretrained token embeddings of the backbone model to construct N-gram representations, thereby eliminating the need to train separate N-gram embeddings from scratch. This design requires neither an additional memory table nor a retrieval pipeline. The Cosine-Gated Memory Injector then uses a non-parametric cosine gate with ReLU to modulate the retrieved embeddings into the contextual representations. We evaluate NGM on the Qwen3 series from 0.6B to 14B across eight benchmarks. NGM improves average performance by 0.5 to 1.2 points, with particularly clear gains on code generation and knowledge-intensive tasks (e.g., +3.0 on LiveCodeBench and +3.03 on GPQA for Qwen3-14B). Moreover, NGM also improves performance in multimodal benchmarks (e.g., MMStar +1.53 on Qwen3-VL-2B).

Memory plays a central role in enabling large language models (LLMs) to operate over sequential tasks by accumulating and reusing experience over time. However, existing evaluations of LLM memory mostly rely on aggregate metrics such as final hold-out accuracy or cumulative online performance, which can obscure critical failure modes such as forgetting and negative transfer. In this paper, we introduce SeqMem-Eval, a diagnostic evaluation framework for sequentially evolving LLM memory. Drawing inspiration from continual learning, it targets a test-time setting in which memory is external, prompt-mediated, and updated without modifying model parameters. Rather than focusing only on final performance, SeqMem-Eval evaluates how memory states evolve, generalize, consolidate experience, and retain useful information during sequential inference. Specifically, it measures online utility, hold-out generalization, backward transfer, and forgetting, providing a finer-grained view of memory quality. Through extensive experiments across diverse tasks and memory methods, we show that higher final or cumulative accuracy does not necessarily imply better memory quality: many methods exhibit strong performance gains while suffering from substantial forgetting or negative transfer. Moreover, different memory designs exhibit distinct trade-offs between adaptability and stability that remain invisible under standard evaluation metrics.

Large language models (LLMs) achieve strong performance across a wide range of tasks, but remain frozen after pretraining until subsequent updates. Many real-world applications require timely, domain-specific information, motivating the need for efficient mechanisms to incorporate new knowledge. In this paper, we introduce MeMo (Memory as a Model), a modular framework that encodes new knowledge into a dedicated memory model while keeping the LLM parameters unchanged. Compared to existing methods, MeMo offers several advantages: (a) it captures complex cross-document relationships, (b) it is robust to retrieval noise, (c) it avoids catastrophic forgetting in the LLM, (d) it does not require access to the LLM's weights or output logits, enabling plug-and-play integration with both open and proprietary closed-source LLMs, and (e) its retrieval cost is independent of corpus size at inference time. Our experimental results on three benchmarks, BrowseComp-Plus, NarrativeQA, and MuSiQue, show that MeMo achieves strong performance compared to existing methods across diverse settings.

Prefix caching is a key optimization in Large Language Model (LLM) serving, reusing attention Key-Value (KV) states across requests with shared prompt prefixes to reduce expensive prefill computation. However, its benefit depends critically on the eviction policy as GPU memory is scarce, and existing policies such as LRU largely treat cached blocks uniformly. This view ignores a fundamental property of LLM prompts: not all tokens are equally worth caching. We show that different token types within a prompt, including system prompts, user queries, tool outputs, model responses, and chain-of-thought reasoning, exhibit up to 756x variation in reuse rates, yet no existing eviction policy exploits this signal. In this paper, we present SAECache (Semantic-Adaptive Eviction for prefix caches), a semantic-adaptive prefix cache eviction policy that addresses this gap through three innovations: (1) a multi-queue architecture that routes KV blocks to task-specific queues with tailored priority metrics, capturing both session reuse in multi-turn requests and structural reuse in templated single-turn requests; (2) a semantic-aware token weighting mechanism that learns the reuse value of different token types online through eviction feedback; and (3) a fully adaptive online learning schema for all parameter updates, including log-normal timing parameters, position decay power, queue weights, and meta-parameters, which eliminates manual tuning and enables automatic adaptation to deployment-specific workload characteristics. Through extensive evaluation across heterogeneous workloads, we demonstrate that SAECache achieves 1.4x-2.7x TTFT improvement over production-style baselines, while fixed-parameter alternatives can degrade by up to 2.7x under workload mismatch -- a failure mode our adaptive approach avoids entirely.

Current LLM memory benchmarks evaluate answer quality rather than retrieval accuracy. Consequently, a system that dumps its entire belief store can achieve perfect recall and mask severe precision failures. We show this evaluation gap persists across multiple embedding models where similarity-based retrieval over domain-specific corpora inherently struggles to isolate target beliefs from semantically proximate ones. Furthermore, multi-turn topic drift compounds this retrieval noise while driving up latency and operational costs. To decouple retrieval quality from generative performance, we introduce PrecisionMemBench, an 89-case benchmark measuring precision, noise isolation, session latency, and belief mutability. We also present Tenure, a structured belief-store proxy that resolves scope and retrieval before inference and injects typed belief state as ambient instruction before the model sees the prompt, removing model-side discretion over whether memory is consulted. Evaluated across 13 providers, Tenure achieves perfect retrieval passes across all active, non-session, and session test cases. In contrast, the baseline configurations fail to reach even half of the active passes, with precision scores clustering at 0.22 and below. Our results demonstrate that while current memory systems successfully store information, they fail to retrieve it cleanly; a structural vulnerability that traditional answer-quality benchmarks conceal.

Continual post-training aims to extend large language models (LLMs) with new knowledge, skills, and behaviors, yet it remains unclear when sequential updates enable capability transfer and when they cause catastrophic forgetting. Existing methods mitigate forgetting through sequential fine-tuning, replay, regularization, or model merging, but offer limited criteria for determining when incorporating new updates is beneficial or harmful. In this work, we study LLM continual post-training through three questions: What drives forgetting? When do sequentially acquired capabilities transfer or interfere? How can compatibility be used to control update integration? We address these questions through task geometry: we represent each post-training task by its parameter update and study the covariance geometry induced by the update. Our central finding is that: forgetting can be considered as a state-relative update-integration failure, it arises when the covariance geometries induced by tasks misalign with the geometry of the evolving model state. Sequential updates transfer when they remain compatible with the model state shaped by previous updates, and interfere when state-relative geometry conflict becomes high. Motivated by this finding, we propose Geometry-Conflict Wasserstein Merging (GCWM), a data-free update-integration method that constructs a shared Wasserstein metric via Gaussian Wasserstein barycenters and uses geometry conflict to gate geometry-aware correction. Across Qwen3 0.6B--14B on domain-continual and capability-continual settings, GCWM consistently outperforms data-free baselines, improving retention and final performance without replay data. These results identify geometry conflict as both an explanatory signal for forgetting and a practical control signal for LLM continual post-training.

Large language models (LLMs) support long-context inference but suffer from substantial memory and runtime overhead due to Key-Value (KV) Cache growth. Existing KV Cache eviction methods primarily rely on local attention weights, neglecting the influence of value representations, output projection, and inter-head interactions. In this work, we reformulate KV Cache eviction from a conventional head-wise, weight-averaging approach into an output-aware, layer-wise matrix multiplication approximation problem. We introduce LaProx, a novel eviction strategy that explicitly models the multiplicative interaction between attention maps and projected value states to accurately quantify token contributions while accounting for inter-head dependencies. Building on this metric, we propose the first unified eviction strategy that assigns globally comparable importance scores to tokens, enabling model-wide selection instead of local, head-wise decisions. Experimental results across 19 datasets on long-context benchmarks LongBench and Needle-In-A-Haystack demonstrate that our approach maintains model performance with only 5% of the KV cache and consistently outperforms prior works across all configurations. Notably, our method achieves up to 2$\times$ accuracy loss reduction under extreme compression scenarios compared to existing state-of-the-art baselines with minimal overhead.

Continual incorporation of new knowledge is essential for the long-term evolution of large language models (LLMs). Existing approaches typically rely on parameter-update algorithms to mitigate catastrophic forgetting, yet they suffer from fundamental limitations: 1) forgetting is unavoidable as the amount of newly injected knowledge grows; and 2) model updates are often irreversible. As modern LLMs become increasingly expressive, it is natural to question whether large-scale weight updates are necessary for acquiring a small amount of new knowledge. In this work, we propose a principled framework that models autoregressive language generation as a Markov process over tokens, where model memory is represented by a Markov transition matrix. Under this formulation, incorporating new knowledge/tokens corresponds to extending the state space, and preserving existing transitions guarantees retention of previously learned knowledge. We then prove a sample complexity bound for incorporating new tokens via a token-to-dictionary mapping strategy. In particular, for learning the transition behavior of each new token, the required number of samples scales linearly with the number of existing tokens it is mapped to. To realize this mapping, we propose an embedding-tuning algorithm that requires minimal parameter updates and induces zero forgetting. Experimental results further demonstrate the effectiveness of our method and validate our theoretical findings.

Adapting a pretrained language model to a new task often hurts the general capabilities it already had, a problem known as catastrophic forgetting. Sparse Memory Finetuning (SMF) tries to avoid this by adding key-value memory layers to the model and, on each training step, updating only the small set of memory rows that the current batch reads most heavily. We re-implement SMF on Qwen-2.5-0.5B-Instruct and compare it with LoRA and full finetuning on MedMCQA, a 4-choice medical exam task, using WikiText perplexity and TriviaQA accuracy as forgetting probes. SMF improves MedMCQA by 2.5 percentage points while keeping both forgetting probes within roughly 1 point of the base model, whereas LoRA and full finetuning achieve larger gains but with clear drift on both. We also compare two row-selection rules (KL-divergence and TF-IDF), which balance the two forgetting metrics differently.

Key-value (KV) caching is essential for large language model inference, yet its memory overhead poses a critical bottleneck for long-context generation. Existing eviction policies predominantly rely on empirical heuristics, lacking a rigorous theoretical foundation. This work rethinks KV cache eviction through the lens of the Information Bottleneck principle. Under a linear-Gaussian surrogate of attention, we derive a closed-form mutual information objective that characterizes the effective information capacity of a retained KV cache subset. This formulation reveals that a wide range of existing eviction strategies can be interpreted as different approximations of the same capacity-maximization principle. Guided by this insight, we introduce CapKV, a capacity-aware eviction method that directly targets information preservation via a log-determinant approximation using statistical leverage scores. This approach replaces heuristic selection with a theoretically grounded mechanism that preserves the maximum predictive signal. Extensive experiments across multiple models and long-context benchmarks show that CapKV consistently outperforms prior methods, achieving a better trade-off between memory efficiency and generational fidelity.

Evolve pairs a small local language model with a persistent, teacher-compiled knowledge store -- refined through sleep consolidation and usage-driven refresh -- to deliver substantial accuracy gains over the model's parametric baseline while amortizing teacher costs through cross-query knowledge reuse. Rather than retrieving document fragments at query time, Evolve constructs a store of semantically coherent sections compiled by teacher models at natural conceptual boundaries; new sections are staged on acquisition, consolidated offline through teacher-mediated merging, and refreshed inline when expired. A 2B-parameter local model handles classification and generation; large teacher models are invoked only for knowledge operations. Across 750 benchmark queries spanning custom specialist questions, NaturalQuestions, and TriviaQA, the 2B model augmented by Evolve improves from 20-33% baseline accuracy to 60-84% (+40-52pp) while reducing teacher invocations by over 50% through reuse. Post-consolidation compresses the knowledge store by 31-33.5% across three independent benchmarks while preserving accuracy; section-based retrieval outperforms chunk-based retrieval by 5-9pp across every lifecycle condition. The architecture supports two generation modes over the same lifecycle -- suppress (strict section-only grounding, auditable) and augment (section-supplemented responses).

MemPalace is an open-source AI memory system that applies the ancient method of loci (memory palace) spatial metaphor to organize long-term memory for large language models; launched in April 2026, it accumulated over 47,000 GitHub stars in its first two weeks and claims state-of-the-art retrieval performance on the LongMemEval benchmark (96.6% Recall@5) without requiring any LLM inference at write time. Through independent codebase analysis, benchmark replication, and comparison with competing systems, we find that MemPalace's headline retrieval performance is attributable primarily to its verbatim storage philosophy combined with ChromaDB's default embedding model (all-MiniLM-L6-v2), rather than to its spatial organizational metaphor per se -- the palace hierarchy (Wings->Rooms->Closets->Drawers) operates as standard vector database metadata filtering, an effective but well-established technique. However, MemPalace makes several genuinely novel contributions: (1) a contrarian verbatim-first storage philosophy that challenges extraction-based competitors, (2) an extremely low wake-up cost (approximately 170 tokens) through its four-layer memory stack, (3) a fully deterministic, zero-LLM write path enabling offline operation at zero API cost, and (4) the first systematic application of spatial memory metaphors as an organizing principle for AI memory systems. We also note that the competitive landscape is evolving rapidly, with Mem0's April 2026 token-efficient algorithm raising their LongMemEval score from approximately 49% to 93.4%, narrowing the gap between extraction-based and verbatim approaches. Our analysis concludes that MemPalace represents significant architectural insight wrapped in overstated claims -- a pattern common in rapidly adopted open-source projects where marketing velocity exceeds scientific rigor.

We present SCM (Sleep-Consolidated Memory), a research preview of a memory architecture for large language models that draws on neuroscientific principles to address a fundamental limitation in current systems: the absence of persistent, structured, and biologically plausible memory. Existing approaches rely on truncating context windows, growing vector databases without bound, or tiered storage systems that lack consolidation and forgetting mechanisms. SCM implements five core components inspired by human memory: a limited-capacity working memory, multi-dimensional importance tagging, offline sleep-stage consolidation with distinct NREM and REM phases, intentional value-based forgetting, and a computational self-model enabling introspection. Across a standardized benchmark suite of eight tests, the prototype achieves perfect recall accuracy over ten-turn conversations while reducing memory noise by 90.9% through adaptive forgetting. Memory search latency remains below one millisecond even with hundreds of stored concepts. This work establishes the architectural foundations for memory systems that consolidate, prioritize, and forget, offering a testable platform for advancing LLM memory research.

Constructivist epistemology argues that knowledge is actively constructed rather than passively copied. Despite the generative nature of Large Language Models (LLMs), most existing agent memory systems are still based on dense retrieval. However, dense retrieval heavily relies on semantic overlap or entity matching within sentences. Consequently, embeddings often fail to distinguish instances that are semantically similar but contextually distinct, introducing substantial noise by retrieving context-mismatched entries. Conversely, directly employing open-ended generation for memory access risks "Structural Hallucination" where the model generates memory keys that do not exist in the memory, leading to lookup failures. Inspired by this epistemology, we posit that memory is fundamentally organized by cognitive schemas, and valid recall must be a generative process performed within these schematic structures. To realize this, we propose SCG-MEM, a schema-constrained generative memory architecture. SCG-MEM reformulates memory access as Schema-Constrained Generation. By maintaining a dynamic Cognitive Schema, we strictly constrain LLM decoding to generate only valid memory entry keys, providing a formal guarantee against structural hallucinations. To support long-term adaptation, we model memory updates via assimilation (grounding inputs into existing schemas) and accommodation (expanding schemas with novel concepts). Furthermore, we construct an Associative Graph to enable multi-hop reasoning through activation propagation. Experiments on the LoCoMo benchmark show that SCG-MEM substantially improves performance across all categories over retrieval-based baselines.

As Large Language Models (LLMs) become increasingly popular, caching responses so that they can be reused by users with semantically similar queries has become a vital strategy for reducing inference costs and latency. Existing caching frameworks have proposed to decide which query responses to cache by assuming a finite, known universe of discrete queries and learning their serving costs and arrival probabilities. As LLMs' pool of users and queries expands, however, such an assumption becomes increasingly untenable: real-world LLM queries reside in an infinite, continuous embedding space. In this paper, we establish the first rigorous theoretical framework for semantic LLM response caching in continuous query space under uncertainty. To bridge the gap between discrete optimization and continuous representation spaces, we introduce dynamic $ε$-net discretization coupled with Kernel Ridge Regression. This design enables the system to formally quantify estimation uncertainty and generalize partial feedback on LLM query costs across continuous semantic query neighborhoods. We develop both offline learning and online adaptive algorithms optimized to reduce switching costs incurred by changing the cached responses. We prove that our online algorithm achieves a sublinear regret bound against an optimal continuous oracle, which reduces to existing bounds for discrete query models. Extensive empirical evaluations demonstrate that our framework approximates the continuous optimal cache well while also reducing computational and switching overhead compared to existing methods.

Chain-of-thought reasoning has driven striking advances in language model capability, yet every reasoning step grows the KV cache, creating a bottleneck to scaling this paradigm further. Current approaches manage these constraints on the model's behalf using hand-designed criteria. A more scalable approach would let end-to-end learning subsume this design choice entirely, following a broader pattern in deep learning. After all, if a model can learn to reason, why can't it learn to forget? We introduce Neural Garbage Collection (NGC), in which a language model learns to forget while learning to reason, trained end-to-end from outcome-based task reward alone. As the model reasons, it periodically pauses, decides which KV cache entries to evict, and continues to reason conditioned on the remaining cache. By treating tokens in a chain-of-thought and cache-eviction decisions as discrete actions sampled from the language model, we can use reinforcement learning to jointly optimize how the model reasons and how it manages its own memory: what the model evicts shapes what it remembers, what it remembers shapes its reasoning, and the correctness of that reasoning determines its reward. Crucially, the model learns this behavior entirely from a single learning signal - the outcome-based task reward - without supervised fine-tuning or proxy objectives. On Countdown, AMC, and AIME tasks, NGC maintains strong accuracy relative to the full-cache upper bound at 2-3x peak KV cache size compression and substantially outperforms eviction baselines. Our results are a first step towards a broader vision where end-to-end optimization drives both capability and efficiency in language models.

While large language models have achieved remarkable performance in complex tasks, they still need a memory system to utilize historical experience in long-term interactions. Existing memory methods (e.g., A-Mem, Mem0) place excessive emphasis on organizing interactions by frequently rewriting them, however, this heavy reliance on summarization risks diluting essential contextual nuances and obscuring key retrieval features. To bridge this gap, we introduce AnchorMem, a novel memory framework inspired by the Proust Phenomenon in cognitive science, where a specific anchor triggers a holistic recollection. We propose a method that decouples the retrieval unit from the generation context. AnchorMem extracts atomic facts from interaction history to serve as retrieval anchors, while preserving the original context as the immutable context. To reveal implicit narrative cues, we construct an associative event graph that uses higher-order event links that bind sets of related facts into shared event representations, strengthening cross-memory integration without relying on generic entities as bridges. During retrieval, the system anchors queries to specific facts and events to locate relevant memories, but reconstructs the context using the associated raw chunks and events. Our method reconciles fine-grained retrieval with the contextual integrity of interactions. Experiments across three closed-source and open-source models on the LoCoMo benchmark demonstrate that AnchorMem significantly outperforms baselines. Code is available at https://github.com/RayNeo-AI-2025/AnchorMem.

Memory embeddings are crucial for memory-augmented systems, such as OpenClaw, but their evaluation is underexplored in current text embedding benchmarks, which narrowly focus on traditional passage retrieval and fail to assess models' ability to handle long-horizon memory retrieval tasks involving fragmented, context-dependent, and temporally distant information. To address this gap, we introduce the Long-horizon Memory Embedding Benchmark (LMEB), a comprehensive framework for evaluating embedding models on complex, long-horizon memory retrieval. LMEB comprises 22 datasets and 193 zero-shot retrieval tasks spanning four memory types: episodic, dialogue, semantic, and procedural. These memory types differ in terms of level of abstraction and temporal dependency, capturing distinct aspects of memory retrieval that reflect the diverse challenges of the real world. We evaluate 15 widely used embedding models, ranging from hundreds of millions to ten billion parameters. The results reveal that (1) LMEB provides a reasonable level of difficulty; (2) Larger models do not always perform better; (3) LMEB and MTEB measure orthogonal capabilities. This suggests that the field has yet to converge on a universal model capable of excelling across all memory retrieval tasks, and that strong performance on traditional passage retrieval does not necessarily transfer to long-horizon memory retrieval. LMEB provides a standardized and reproducible framework that fills a key gap in memory embedding evaluation and supports future advances in long-term, context-dependent retrieval.

As Large Language Models (LLMs) are increasingly used for long-duration tasks, maintaining effective long-term memory has become a critical challenge. Current methods often face a trade-off between cost and accuracy. Simple storage methods often fail to retrieve relevant information, while complex indexing methods (such as memory graphs) require heavy computation and can cause information loss. Furthermore, relying on the working LLM to process all memories is computationally expensive and slow. To address these limitations, we propose MemSifter, a novel framework that offloads the memory retrieval process to a small-scale proxy model. Instead of increasing the burden on the primary working LLM, MemSifter uses a smaller model to reason about the task before retrieving the necessary information. This approach requires no heavy computation during the indexing phase and adds minimal overhead during inference. To optimize the proxy model, we introduce a memory-specific Reinforcement Learning (RL) training paradigm. We design a task-outcome-oriented reward based on the working LLM's actual performance in completing the task. The reward measures the actual contribution of retrieved memories by mutiple interactions with the working LLM, and discriminates retrieved rankings by stepped decreasing contributions. Additionally, we employ training techniques such as Curriculum Learning and Model Merging to improve performance. We evaluated MemSifter on eight LLM memory benchmarks, including Deep Research tasks. The results demonstrate that our method meets or exceeds the performance of existing state-of-the-art approaches in both retrieval accuracy and final task completion. MemSifter offers an efficient and scalable solution for long-term LLM memory. We have open-sourced the model weights, code, and training data to support further research.

Continuous knowledge updating for pre-trained large language models (LLMs) is increasingly necessary yet remains challenging. Although inference-time methods like In-Context Learning (ICL) and Retrieval-Augmented Generation (RAG) are popular, they face constraints in context budgets, costs, and retrieval fragmentation. Departing from these context-dependent paradigms, this work investigates a parametric approach using Low-Rank Adaptation (LoRA) as a modular knowledge memory. Although few recent works examine this concept, the fundamental mechanics governing its capacity and composability remain largely unexplored. We bridge this gap through the first systematic empirical study mapping the design space of LoRA-based memory, ranging from characterizing storage capacity and optimizing internalization to scaling multi-module systems and evaluating long-context reasoning. Rather than proposing a single architecture, we provide practical guidance on the operational boundaries of LoRA memory. Overall, our findings position LoRA as the complementary axis of memory alongside RAG and ICL, offering distinct advantages.

Semantic caching has emerged as a pivotal technique for scaling LLM applications, widely adopted by major providers including AWS and Microsoft. By utilizing semantic embedding vectors as cache keys, this mechanism effectively minimizes latency and redundant computation for semantically similar queries. In this work, we conceptualize semantic cache keys as a form of fuzzy hashes. We demonstrate that the locality required to maximize cache hit rates fundamentally conflicts with the cryptographic avalanche effect necessary for collision resistance. Our conceptual analysis formalizes this inherent trade-off between performance (locality) and security (collision resilience), revealing that semantic caching is naturally vulnerable to key collision attacks. While prior research has focused on side-channel and privacy risks, we present the first systematic study of integrity risks arising from cache collisions. We introduce CacheAttack, an automated framework for launching black-box collision attacks. We evaluate CacheAttack in security-critical tasks and agentic workflows. It achieves a hit rate of 86% in LLM response hijacking and can induce malicious behaviors in LLM agent, while preserving strong transferability across different embedding models. A case study on a financial agent further illustrates the real-world impact of these vulnerabilities. Finally, we discuss mitigation strategies.

Large language models (LLMs) often produce incorrect or outdated content after being employed. Efficient and accurate knowledge updates without costly retraining are a major challenge. This problem is particularly challenging in lifelong settings, where complex, unstructured knowledge must coexist without interference. We introduce RILKE (Representation Intervention for Lifelong KnowledgE Control), a robust and scalable method that treats knowledge control as interventions within the model's representation space. Leveraging representation-space expressiveness, we identify two key properties enabling RILKE to achieve fine-grained control over complex, unstructured knowledge while maintaining general utility with frozen base weights. During training, RILKE learns paraphrase-robust and edit-localized modules that limit each update to a low-dimensional subspace to minimize cross-edit interference. At inference, a query-adaptive router selects the appropriate module to guide the model's generation. Across LLaMA and Qwen models, RILKE scales effectively to large-scale benchmarks, demonstrating high edit success and strong paraphrase generalization while preserving general utility with modest memory overhead. These results show RILKE is an effective and scalable solution for lifelong knowledge control in LLMs.

Multimodal large language models (MLLMs) can inadvertently memorize privacy-sensitive information during training. While existing unlearning methods can remove such content, they often severely degrade the model's foundational capabilities, such as general image understanding. This critical shortfall motivates our investigation into benign memory forgetting, the precise removal of targeted, privacy-sensitive knowledge while rigorously preserving unrelated capabilities. To pioneer and evaluate progress toward this objective, we introduce S-MLLMUn Bench, the first benchmark designed to jointly and quantitatively assess an unlearning method's efficacy in knowledge erasure and the preservation of image understanding. Furthermore, we propose the Sculpted Memory Forgetting Adapter (SMFA), a new framework that enables benign memory forgetting. SMFA confines forgetting to designated memory regions, maintaining overall model performance. By initially fine-tuning the model to replace sensitive outputs with refusals, SMFA generates a memory forgetting adapter, followed by a retaining anchor-guided masking mechanism that safeguards unrelated knowledge. Extensive experiments on S-MLLMUn Bench demonstrate that existing methods fail to achieve benign forgetting, whereas our proposed SMFA serves as an effective baseline, successfully achieving targeted knowledge erasure without compromising the model's foundational visual capabilities. Code and data are available at https://github.com/zeng-zhen/S-MLLMUn.

Processing long documents with large language models (LLMs) is expensive: a single query over a 100K-token document can cost from tens of cents to over a dollar in API fees, depending on the model, and memory grows linearly with context length. We introduce BudgetMem, a training-free architecture that keeps only high-salience content under an explicit memory budget. Unlike token-level neural compressors such as LLMLingua, BudgetMem makes chunk-level keep-or-discard decisions from interpretable features: entity density, TF-IDF importance, position, numerical density, discourse markers, and question presence. Across four benchmarks, BudgetMem matches the uncompressed baseline on template-generated structured documents (F1 = 0.859 vs. 0.855) while discarding 70% of chunks. As a pre-retrieval filter on the same documents it outperforms LLMLingua-2 by a wide margin (0.859 vs. 0.554), because token-level compression destroys the phrasal structure that retrieval depends on; the advantage holds across model scale and family (Llama-3.2-3B, Qwen2.5-7B), under dense retrieval (0.918 vs. 0.530), and in LLMLingua-2's intended direct-feed mode (0.797 vs. 0.720). On real NLP research papers (Qasper), BudgetMem is competitive but not free, trailing the full baseline by 7.4% and LLMLingua-2 by 5.5% relative F1 at a 72% storage reduction. Removing discourse markers or numerical density hurts most, although TF-IDF alone scores as well on template documents. The pipeline needs no trained models or GPU for compression and runs on a $10/month Colab instance.

Adapting language models (LMs) to new tasks via post-training carries the risk of degrading existing capabilities -- a phenomenon classically known as catastrophic forgetting. In this paper, toward identifying guidelines for mitigating this phenomenon, we systematically compare the forgetting patterns of two widely adopted post-training methods: supervised fine-tuning (SFT) and reinforcement learning (RL). Our experiments reveal a consistent trend across LM families (Llama, Qwen) and tasks (instruction following, general knowledge, and arithmetic reasoning): RL leads to less forgetting than SFT while achieving comparable or higher target task performance. To investigate the cause for this difference, we consider a simplified setting in which the LM is modeled as a mixture of two distributions, one corresponding to prior knowledge and the other to the target task. We identify that the mode-seeking nature of RL, which stems from its use of on-policy data, enables keeping prior knowledge intact when learning the target task. We then verify this insight by demonstrating that the use on-policy data underlies the robustness of RL to forgetting in practical settings, as opposed to other algorithmic choices such as the KL regularization or advantage estimation. Lastly, as a practical implication, our results highlight the potential of mitigating forgetting using approximately on-policy data, which can be substantially more efficient to obtain than fully on-policy data.

FastText remains a practical choice for industrial word representation because it can synthesize vectors for out-of-vocabulary words from character n-grams. Its original hash-bucket implementation, however, couples two engineering compromises that become painful at large scale: unrelated n-grams collide into the same row, while increasing the bucket count quickly turns the input matrix into the dominant memory cost. This paper presents a memory-efficient FastText variant based on an exact-then-compress principle: first give every observed word and n-gram an explicit identity, then compress only those rows whose learned vectors and lexical structure justify sharing. Concretely, we replace hash buckets with collision-free double-array trie indexes and compress the resulting n-gram matrix through structurally constrained prefix and suffix merging followed by mark-compact row reorganization. Unlike arbitrary hashing, the proposed method shares rows only after a high cosine-similarity test, preserving interpretable n-gram identities while reducing the number of live rows. We describe the full training and serving pipeline, including UTF-8 aware n-gram enumeration, double-array trie lookup, memory-mapped model loading, and vector reconstruction for words and sentences. On a large Chinese vocabulary benchmark with 30.1M words and 287.4M extracted n-grams, the compressed model reduces memory from 145.2GB to 28.9GB, improves load time from 12.3 minutes to 3.2 minutes, and preserves downstream quality within 0.3 points of a hash-free model. We position the result as a compact lexical memory layer for LLM-era retrieval systems and release the implementation as an extended FastText prototype.