memory

Nomura published a report indicating that AI-driven demand is growing exponentially, while memory supply is limited, suggesting that memory stocks are set for a valuation reassessment. The firm significantly raised the target prices for Samsung Electronics and SK Hynix, with Samsung's target price soaring from 340,000 KRW to 590,000 KRW, and SK Hynix's target price rising from 2,340,000 KRW to 4,000,000 KRW, both rated as "Buy." Nomura stated that as AI semiconductor demand shifts from training to inference workloads, memory demand is entering a period of exponential expansion.In contrast, the firm believes that industry supply growth during the same period may still be constrained to about 5 to 6 times (with a compound annual growth rate of about 30%), raising serious questions about whether the structural supply-demand imbalance can truly be resolved. The firm mentioned that currently, the industry is attempting to narrow this widening supply-demand gap through various software and architectural optimizations, but Nomura believes these solutions can only slow the growth rate rather than reverse the trend.

The GoPlus Security team has disclosed a new type of attack in its AgentGuard AI project: inducing AI agents to perform unauthorized sensitive operations through "memory poisoning." This attack method does not rely on traditional vulnerabilities or malicious code but exploits the long-term memory mechanism of AI agents. For example, an attacker first induces the agent to "remember preferences," such as "usually prioritizing proactive refunds instead of waiting for chargebacks," and then uses vague expressions like "process as usual" or "execute as before" in subsequent instructions, thereby triggering automated financial operations.GoPlus points out that the key risk in such cases lies in the AI agent mistakenly treating "historical preferences" as a basis for authorization, leading to financial losses or security incidents in operations such as refunds, transfers, and configuration changes. To address this issue, the team has proposed several protective recommendations, including:Operations involving refunds, transfers, deletions, or sensitive configurations must require explicit confirmation in the current session.Memory-related instructions like "habit," "usual way," and "as before" should be regarded as high-risk state changes.Long-term memory must have a traceability mechanism (writer, time, confirmation status).Vague instructions should automatically elevate the risk level and trigger secondary verification.Long-term memory must not replace real-time authorization processes.The team emphasizes that the "AI agent memory system" should be viewed as a potential attack surface and should be constrained and audited through a dedicated security framework.

Aptos announced that after the governance vote passes, it will become the first Layer 1 public chain to provide a native encrypted mempool. This feature achieves complete confidentiality of transaction intentions, does not affect transaction speed, and aligns the trust assumptions with the network itself, aiming to protect users from frontrunning at the protocol level.

According to OnchainLens monitoring, the whale "0x3Ed" added an investment of 2.668 million USDC to Hyperliquid and opened positions in EWY (iShares MSCI Korea ETF) and DRAM (memory ETF) with 20x leverage.This whale currently still holds long positions in MU (Micron Technology) and SNDK (SanDisk) with a total value of 12 million dollars, realizing an unrealized profit of 848,000 dollars.

Tencent Cloud officially released the "Lobster" memory service ------ TencentDB Agent Memory, adding a layer of long-term memory capability to OpenClaw. Currently, Agent Memory is seamlessly integrated into Tencent Cloud Lighthouse, ClawPro, and other products in the form of a plugin, supporting free one-click activation.

Polygon announced the launch of the Private Mempool, a private transaction submission entry that protects each transaction from front-running and sandwich attacks.Simply replace the RPC address with the corresponding private entry to use this feature, with no additional operations required.

According to Jinshi reports, U.S. memory stocks generally fell, with Micron Technology down 5.28%, SanDisk down 2.7%, and Seagate down 4.4%.

AI visual memory startup Memories.ai announced a partnership with NVIDIA to build a visual memory infrastructure for wearable devices and robots using NVIDIA's Cosmos-Reason 2 visual language model and Metropolis video search platform.The company was co-founded by former Meta Ray-Ban glasses AI team members Shawn Shen and Ben Zhou in 2024 and has completed $16 million in funding, led by Susa Ventures, with participation from Seedcamp, Fusion Fund, and Crane Venture Partners. The company launched its Large Visual Memory Model (LVMM) in July 2025 and has partnered with Qualcomm, planning to run the model on Qualcomm processors later this year.

Ethereum co-founder Vitalik Buterin recently published a series of technical articles discussing the future roadmap of Ethereum. In the latest article, he focused on analyzing the potential centralization risks in the block building pipeline and proposed solutions such as expanding the FOCIL mechanism and introducing encrypted mempools to enhance the network's censorship resistance.According to the plan, Ethereum will launch the Glamsterdam upgrade in the first half of 2026, which will introduce the enshrined Proposer-Builder Separation (ePBS) mechanism. This mechanism allows block proposers to outsource block construction to a permissionless open market, reducing the centralization risk at the staking level. However, Buterin pointed out that while ePBS can prevent the concentration of block building rights among a few staking pools, the block construction itself may still become concentrated among a few high-tier participants due to specialization and maximizing MEV, leading to a new trend of centralization.To address this issue, Ethereum developers plan to simultaneously launch the FOCIL (Forward Obligatory Commitment to Inclusion Lists) mechanism in the Glamsterdam upgrade. The initial version will randomly select 16 witnesses and mandate that specific transactions must be included in the block; otherwise, the block will be rejected. Buterin stated that even if block construction is controlled by a single malicious entity, FOCIL can still ensure that transactions cannot be completely censored.Additionally, Buterin explored the possibility of expanding the scale of FOCIL ("big FOCIL") and introducing encrypted mempools to further mitigate the issues of information asymmetry and power concentration in the block building process. Recently, Buterin has been vocal about topics such as the quantum resistance roadmap, execution layer improvements, and block building mechanisms, indicating that the core Ethereum development team is conducting systematic design and risk assessment for the next phase of protocol upgrades.

Anthropic is accused of launching a prompt tool for "exporting ChatGPT memory data," helping users transfer historical memory information to its model Claude, which has drawn industry attention.According to public content, the tool allows users to export their memory data from OpenAI's ChatGPT by copying and pasting specific prompts, and then import it into Claude. Related discussions suggest that this move is seen as directly weakening the user stickiness and switching costs that ChatGPT relies on through its "memory function."Market views consider the memory mechanism as an important moat for large model products— the longer users engage, the deeper the model understands their preferences, context, and historical conversations, thereby increasing the migration costs. If third-party tools can facilitate easy data migration, it may change the current user lock-in logic of AI products.Meanwhile, reports also mentioned that Anthropic was previously restricted from use by relevant systems of the U.S. Department of Defense, but the company's popularity and attention have quickly surged, topping some application rankings.Currently, the specific compliance of the aforementioned tool and the platform's response have not been clarified. The industry generally believes that competition among large models has extended from performance comparisons to ecological and data sovereignty aspects, with user data portability potentially becoming a key variable in the next stage.

According to the tracking data from the Web3 asset data platform RootData X, in the past 7 days, the decentralized AI memory layer Unibase has been the project with the most unfollows by top figures on X (Twitter). The newly unfollowed influential figures on X include BroLeon (@BroLeon), 0xAA (@0xAA_Science), and 0xAA (@0xAA_Science).In addition, the project with the most unfollows by top figures on X also includes Kaito.

According to official news, ZetaChain has announced the official launch of the AI interoperability layer ZetaChain 2, along with the release of the privacy-centric large model aggregation application Anuma.ZetaChain 2 is a brand new AI interoperability layer designed to help developers build applications and agents that can operate across AI models, enabling global monetization without the need for backend infrastructure while preserving private user context. ZetaChain 2 consists of two core components: the AI Portal and the Private Memory Layer.The AI Portal serves as a unified routing and execution layer, allowing applications to access multiple AI model providers without being locked in, and includes support for availability, fallback, and cost/performance optimization. The Private Memory Layer is a protocol-level memory system that protects user context through encryption and access control, allowing experiences to persist across sessions while maintaining user control over the access scope of applications and agents.Anuma is the first consumer-level AI application based on ZetaChain 2, currently in beta testing and has launched a public waitlist, allowing users to apply for early access through the public waitlist. The product integrates multiple leading AI models in a single experience, enabling users to switch between different models without losing context, and is designed to keep memories private and under user control.

According to The Block, Architect Financial Technologies, founded by former FTX US president Brett Harrison, plans to launch perpetual futures contracts linked to GPU and memory pricing.The announcement states that the company's trading platform AX, aimed at institutional users, is set to introduce perpetual futures designed to track the daily rental prices of GPUs used for training and inference of artificial intelligence models, as well as dynamic random-access memory (RAM) pricing, and is currently awaiting regulatory approval. It is reported that these contracts are being developed in collaboration with computing index provider Ornn Data, which provides benchmarks based on real-time GPU trading data.In December last year, Architect Financial Technologies completed a $35 million funding round, with a valuation of approximately $187 million.

ChainCatcher message, according to the Web3 asset data platform RootData X tracking data, in the past 7 days, the decentralized AI memory layer Unibase has been the project with the most new followers among X (Twitter) Top personalities. New influential figures following this project include BroLeon (@BroLeonAus) from Australia, crypto industry KOL 嗯哼 (@enheng456), and 默默 (@momochenming).

ChainCatcher news, October 5, Ethereum co-founder Vitalik Buterin published a new article titled "Memory access is O(N^(1/3))" exploring the complexity of memory access, discussing the complexity issues of "memory access" in data structures and algorithms.The article states that arithmetic operations (addition, multiplication, division, etc.) for fixed-size numbers are typically considered to require one time unit, and memory access is also considered to require one time unit, but this assertion is inaccurate. In both theory and practice, memory access requires O(N^(1/3)) time: if your memory is 8 times larger, the time required for read and write operations will increase by 2 times.Vitalik points out that this rule has practical implications in cryptography and algorithm optimization, and calls for future computational models to more accurately reflect memory hierarchy and physical limitations.

ChainCatcher news, OpenGradient AI, which received investment from a16z Crypto for decentralized infrastructure, announced the launch of MemSync, a universal AI memory layer platform that addresses the context memory loss issue of existing AI assistants.MemSync supports ChatGPT, Claude, Perplexity, and other artificial intelligences. The platform also supports personalized memory learning and seamless switching across multiple platforms, enhancing work efficiency, with some users saving over 30 minutes a day and increasing productivity by up to 10 times.

ChainCatcher news, according to CNBC, OpenAI CEO Sam Altman stated that GPT-6 will be launched sooner, with core features including enhanced memory capabilities and support for users to customize the personality and tone of their chatbots.Altman mentioned that the memory feature is the most important upgrade this year, and future products will automatically adjust based on user preferences and habits. Altman also revealed that OpenAI is collaborating with psychologists to optimize the experience and will comply with U.S. government requirements to ensure that AI systems are neutral and customizable.In addition, Altman pointed out that GPT-5 faced criticism from users for its cold tone and has quietly been updated to a warmer version. He emphasized that privacy protection and data encryption still need to be strengthened.

ChainCatcher news, OpenAI tweeted to announce that the lightweight memory feature of ChatGPT is officially available to free users, which was previously limited to paid subscribers.

ChainCatcher news, OpenAI tweeted that it has launched the ChatGPT memory improvement feature for Plus and Pro users in the European Economic Area (EEA), the UK, Switzerland, Norway, Iceland, and Liechtenstein.This feature allows ChatGPT to reference previous conversations with users more smoothly, enhancing the interaction experience. Previously, this feature was rolled out to users in other regions on April 10, allowing users to choose to delete or manage memory content.