AI News Archive: June 17, 2026 — Part 5
Sourced from 500+ daily AI sources, scored by relevance.
- China’s Coding Model, Kimi K2.7 Code, is 6x Cheaper Than Claude. It also Grades Its Own Homework
China’s best open-weight coder is here; it’s a fraction of the cost, and you cannot independently verify a single number it ships with… Continue reading on Towards AI »
- iPhone 18 to Launch With More RAM to Enable Support for More Advanced Siri AI Features: Report
Apple is expected to launch all the iPhone 18 series models with 12GB of RAM, including the standard iPhone 18, according to a report. This is said to allow Apple’s most advanced AI foundational models to run on the rumoured handset, consequently widening the Siri AI adoption and increasing the demand for the upcoming iPhone 18 series handsets. While most AI feature...
- Transforming global water cycle observations via synergistic AI and remote sensing
Science Advances, Volume 12, Issue 25, June 2026.
- AI nation: Is it boom, bubble or both for South Korea?
AI nation: Is it boom, bubble or both for South Korea? Nikkei Asia
Score: 65🌐 MovesJun 17, 2026https://asia.nikkei.com/opinion/ai-nation-is-it-boom-bubble-or-both-for-south-korea - Anthropic becomes first AI startup to join the Frontier carbon removal coalition
Anthropic has joined the Frontier coalition, which received another $915M in pledges to fund carbon removal projects.
- GLM-5.2: Built for Long-Horizon Tasks
GLM-5.2: Built for Long-Horizon Tasks
- Nvidia's Huang pledges AI will boost manufacturing jobs. A test will come in Texas
Nvidia's Huang pledges AI will boost manufacturing jobs. A test will come in Texas Houston Chronicle
Score: 64🌐 MovesJun 17, 2026https://www.houstonchronicle.com/business/article/nvidia-s-huang-pledges-ai-will-boost-22307340.php - Flagright secures $12.5M Series A to scale AI compliance platform
Flagright,the AI operating system for financial crime compliance, has raised a $12.5million Series A funding round led by Infinity Ventures, with participationfrom Sella Direct Ventures and continued ...
Score: 64💰 MoneyJun 17, 2026https://tech.eu/2026/06/17/flagright-secures-125m-series-a-to-scale-ai-compliance-platform/ - Databricks launches AI co-worker, Genie One
Databricks launches AI co-worker, Genie One IT Pro
Score: 64🌐 MovesJun 17, 2026https://www.itpro.com/technology/artificial-intelligence/databricks-launches-ai-co-worker-genie-one - Allbirds Completes Its Moonshot Pivot to AI. The Stock Surges 47%.
Allbirds Completes Its Moonshot Pivot to AI. The Stock Surges 47%. Barron's
Score: 63🌐 MovesJun 17, 2026https://www.barrons.com/articles/allbirds-smartbird-stock-surges-ai-870fc7d6?mod - Apple investors are running out of patience with its AI promises
Apple investors are losing patience with the company’s AI strategy. The stock is coming off its worst week since February after the annual Worldwide Developers Conference failed to convince Wall Street that a long-promised upgrade cycle is any closer to arriving. “There’s a bit of fatigue with Apple and AI,” Tim Chubb, chief investment officer at Girard, […] This story continues at The Next Web
Score: 63🌐 MovesJun 17, 2026https://thenextweb.com/news/apple-investors-are-running-out-of-patience-with-its-ai-promises - From RAG to ontology: Databricks bets on context as the key to trusted AI agents
From RAG to ontology: Databricks bets on context as the key to trusted AI agents InfoWorld
- By the numbers: the AI inferencing market and the infrastructure decisions that define it
As AI moves from development into production, the infrastructure decisions organizations make around inferencing are emerging as a primary driver of competitive differentiation. A new infographic from Futurum Research, sponsored by Lenovo, distills the most critical market data and technical considerations shaping this shift. The global AI inference market is projected to reach $48.8 billion by 2030 at a 46.3% CAGR, with hybrid and edge deployments growing at 65% , significantly outpacing public cloud. Unlike training workloads, inference is continuous, real-time, and latency-sensitive, making the infrastructure mismatch costly: organizations that rely on general-purpose architectures can face 2x higher costs per million tokens compared to those running inference-optimized environments. The infographic highlights five primary constraints memory bandwidth, latency sensitivity, power density, accelerator utilization, and operational tuning and underscores that right-sized infrastructure decisions, supported by specialized AI services, directly shape business outcomes. See below to find out more. Lenovo AI Inferencing Infographic finalpdf Download
- New GSMA report shows how Africa’s top mobile operators are building AI language models in African languages
Six of Africa's biggest mobile operators, including Airtel, MTN, and Orange, are working with GSMA to build AI language models for the continent's 2,000-plus languages.
- Training humanoid robots for high-risk applications using large and small AI models
Training humanoid robots for high-risk applications using large and small AI models research.csiro.au
- How DeepSeek Handles 1 Million Tokens With a Fraction of the Memory
A simple explanation of FlashMemory-DeepSeek-V4 and Lookahead Sparse Attention. Source The race toward million-token context windows has created a new problem for AI systems: memory . Modern large language models can theoretically read entire books, lengthy research reports, massive codebases, and months of conversation history. But as context windows grow, storing and managing all that information becomes increasingly expensive. In many cases, memory consumption becomes a bigger bottleneck than the computation itself. This is the problem a team of researchers from Tencent, Tsinghua University, and HKUST set out to solve in their paper, FlashMemory-DeepSeek-V4: Lightning Index Ultra-Long Context via Lookahead Sparse Attention . Rather than storing every piece of information a model encounters, they propose a smarter approach: predict what will matter in the future and keep only that . The result is a significant reduction in memory usage while maintaining, and in some cases actually improving, performance. In this article, we’ll explore the long-context memory problem, unpack the intuition behind Lookahead Sparse Attention (LSA), and examine why FlashMemory-DeepSeek-V4 could be an important step toward truly scalable AI. Why Ultra-Long Context Is So Expensive To understand why this paper matters, you need to understand one key bottleneck: the KV cache . When a language model processes text, it doesn’t simply read and forget. For every token it has seen, it stores two vectors — a Key and a Value — so it can reference that information later without recomputing it. This is the KV cache, and it’s what gives the model its “memory” during a single session. The problem: the KV cache grows linearly with context length. The longer the document, the more Key-Value pairs pile up in GPU memory. At 128K tokens, it starts to strain the system. At 500K or 1M tokens, it becomes the single biggest memory cost in the entire inference pipeline. A useful analogy: Imagine reading a 500-page book and photocopying every page as you go, so you can refer back to anything later. At chapter five, your stack of copies is manageable. By chapter forty, it weighs more than the book itself. That’s essentially what happens inside a large language model at scale. What makes this worse is that most of what gets stored is never actually used. The researchers found something striking in real-world inference logs: over 90% of requests with contexts longer than 64K tokens could be resolved using only the last 8K tokens. The model was carrying enormous amounts of history it seldom touched. Yet the fix isn’t simply discarding the rest. A simple sliding window approach of keeping only recent tokens fails on tasks that genuinely require reasoning over the full document. You need both : the ability to maintain global context and the efficiency to not store everything. That’s the hard contradiction FlashMemory sets out to resolve. The Core Insight: Not Everything Is Worth Remembering Most long-context research focuses on the same question: how do we process all this more efficiently? FlashMemory asks a different question: does the model need to remember all of it in the first place? When you answer a question about a long document, you rarely need every sentence in equal measure. Some passages are critical. Many are somewhat relevant. Most contribute nothing to the current moment. Traditional attention mechanisms treat them all the same, store everything, attend to everything, and incur the memory cost for everything. Lookahead Sparse Attention directly challenges that assumption. Understanding Lookahead Sparse Attention (LSA) Think about how students prepare for an exam. A strong student doesn’t memorise every sentence in the textbook. They identify key concepts, highlight important sections, and make strategic notes about what’s most likely to come up. LSA brings this logic into memory management for language models. Before the model processes each new step, a lightweight component called the Neural Memory Indexer proactively predicts which portions of the stored context are likely to be important in the next several decoding steps. Only those predicted high-value chunks get loaded into active GPU memory. Everything else is offloaded to CPU memory, still accessible if needed, but not taking up precious GPU space. Here’s the contrast in plain terms: Standard attention: Store everything in GPU memory At each step, attend to all of it Pay the full memory cost no matter what Lookahead Sparse Attention: Predict which chunks will matter next Load only those into GPU memory Pay only for what actually gets used The keyword is lookahead . The system isn’t waiting to see what gets attended to and then deciding to prune. It makes the prediction ahead of time , every 64 decoding steps, so the right memory is already resident when computation begins. This turns memory management from a passive cost into an active, intelligent decision. The Lightning Index: A Search Engine for Memory Pair LSA with another concept introduced in the paper: the Lightning Index . Without an index, finding a specific piece of information in a massive context means scanning through everything. With an index, you can jump directly to what’s relevant. The Lightning Index acts like a search catalogue for the model’s own memory. Instead of exhaustively attending to all stored context, the model uses the index to identify and retrieve the most query-relevant chunks efficiently. It dramatically reduces what needs to remain active while ensuring nothing critical is unreachable. Together, LSA predicts what to keep, the Lightning Index organise how to retrieve it, these form a tiered memory architecture: Heavily Compressed Attention (HCA) layers maintain a 128:1 compressed global view of all context, always present, lightweight LSA-filtered CSA chunks provide fine-grained recall, loaded on-demand based on the indexer’s predictions A local sliding window keeps the most recent tokens always available The model never loses awareness of the full document. It just stops hauling all of it into GPU memory at once. Backbone-Free Training Here’s something I found genuinely elegant about this paper. Training the Neural Memory Indexer doesn’t require loading the massive DeepSeek-V4-Flash backbone, a 285-billion-parameter model with 13B active per token, at all. The indexer is structured as a dual-encoder retrieval model, the same architecture used in dense passage retrieval systems. The backbone model’s compressed key vectors are pre-extracted and frozen offline. The indexer only needs to learn how to map the current hidden state to those frozen targets. The trainable parameters account for less than 0.1% of the full model. The entire Memory Indexer converges within a single H20 GPU hour. For comparison, the team ran approximately 500 training experiments in a single week to find the optimal configuration, something that would have been computationally impossible under traditional joint fine-tuning on the full backbone. That’s the kind of decoupled design that makes research cycles fast and deployment practical. The Results: Less Really Is More Efficiency techniques usually come with trade-offs. Reduce memory aggressively, and performance suffers. Reduce computation and quality drops. Across three long-context benchmarks — LongBench-v2 , LongMemEval , and RULER — the results are: source: https://arxiv.org/abs/2606.09079 The model slightly outperforms the full-memory baseline. On the hardest subset (LongBench-v2-L, 493K tokens), FM-DS-V4 beats the baseline by +1.9% while running on just 10% of the memory. The researchers argue this happens because LSA acts as an attention denoiser . By filtering out thousands of low-relevance historical chunks, the model’s attention is no longer diluted by noise. It focuses more sharply on what matters, which reduces factual hallucinations in long-context tasks. For contrast, the naive alternatives completely collapse: Recency Only (last 8K tokens): average accuracy drops to 33.3% Random 10% (random sparse selection): average accuracy of 38.7% Neither can match LSA’s 77.5%. The indexer isn’t guessing — it’s learned to route intelligently. What This Means for Real AI Applications This matters beyond academic benchmarks. Ultra-long context has always felt like a capability that exists in theory but costs too much to use in practice. The KV cache overhead at 500K+ tokens makes it prohibitively expensive at scale. That affects anyone building: Document QA systems over large corpora Legal or financial contract analysis tools Coding agents that reason over full repositories Long-horizon conversational agents Research tools that synthesise across many papers If you can reduce KV cache memory by 90% without sacrificing accuracy, then longer contexts become economically viable for production. More requests can run concurrently on the same hardware. Latency improves. The infrastructure cost of deploying these systems comes down significantly. A Note on Limitations This paper is honest about what it doesn’t know yet. The project was suspended due to organisational changes at Tencent before some planned ablations could be completed. A few hyperparameters, including the 64-step trigger interval and 0.5 classification threshold, were selected from early exploratory runs rather than systematic sweeps. The optimal 3-layer configuration (layers 10, 12, 20) came from a 500-run Pareto search, but finer-grained analysis remains as future work. Open questions worth following: How does LSA hold up at 2M+ tokens? Does the backbone-free training strategy transfer to other model architectures? What’s the optimal trigger interval for different task types? The project lead has explicitly invited collaboration for the next phase, compute sponsorship, scaling tests, and research integration. This work feels like the beginning of something, not the end. Final Thoughts FlashMemory-DeepSeek-V4 makes a deceptively simple argument: you don’t need to remember everything to understand everything. By training a lightweight Neural Memory Indexer to predict which historical chunks are query-critical, LSA reduces KV cache memory to just 13.5% of the full baseline and to over 90% at 500K tokens, while slightly improving accuracy. The backbone-free decoupled training makes it practical to build and iterate on without the overhead of loading massive models during indexer optimisation. The future of long-context AI may not belong to models that remember everything. It may belong to models that know what’s worth remembering. What’s your take, do you think intelligent memory selection is more promising than simply scaling context windows larger? Drop a comment below. Resources: Paper: FlashMemory-DeepSeek-V4 on arXiv HuggingFace: huggingface.co/papers/2606.09079 Video explainer: YouTube walkthrough How DeepSeek Handles 1 Million Tokens With a Fraction of the Memory was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
- Clearlake-backed Quest Software Acquires Anetac to Advance Identity Security for the Agentic AI Era
Clearlake-backed Quest Software Acquires Anetac to Advance Identity Security for the Agentic AI Era markets.businessinsider.com
- Elon Musk’s AI tool Grok was used in strikes against Iran: US govt
Elon Musk’s AI tool Grok was used in strikes against Iran: US govt The Straits Times
- Wrong-Way AI Trade Costs Florida Stock-Picker $50 Billion
It was 180 degrees off the hot trade on Wall Street. Snub Nvidia Corp., the soaring king of artificial intelligence chips. Embrace Adobe Inc., a sinking casualty of AI disruption. Polen Capital did just that, and more, until it was …
- Microsoft Mulls Using DeepSeek for Copilot Cowork
Microsoft Mulls Using DeepSeek for Copilot Cowork The Information
Score: 61🌐 MovesJun 17, 2026https://www.theinformation.com/briefings/microsoft-mulls-using-deepseek-copilot-cowork - Atlassian MCP + Claude Code: The Beginning of a New Workflow
Atlassian MCP + Claude Code: The Beginning of a New Workflow Atlassian Community
- AI decodes the language of genes: A new look inside the “control center” of plants
AI decodes the language of genes: A new look inside the “control center” of plants EurekAlert!
- Forget basic transcriptions: An upcoming wearable wants to power autonomous AI agents
Moving past note-taking, Plaud's new wearable device will act as an active pipeline for digital models.
- Why Companies Are Already Blowing Through Their 2026 AI Budgets in Just 2 Months
As OpenAI’s Sam Altman admits sudden AI cost spikes have caught businesses off guard, a quiet corporate crisis is taking shape.
- Who owns the control plane? Google Cloud Next 2026 and the real contest in agentic AI
I recently spent some time reflecting on the announcements from Google Cloud Next 2026, as well as a series of vendor briefings and a handful of enterprise architecture engagements, where the same question kept coming up across different venues: once an organization has agents, who governs them? For two years, the enterprise AI conversation has been a conversation about models — whose is largest, whose is cheapest, whose context window stretches furthest. Virtually no one was talking about data and semantic context. After getting some perspective, I was forced to consider that model obsession might have finally fizzled out under the grim reality of non-existent ontologies and limited to no semantic context for enterprise data. The interesting question is no longer which model an enterprise runs. It is who controls the connective context layer — the agentic control plane — that decides what those agents know, what they are allowed to do and who is accountable when a thousand of them are running at once. Whoever owns that layer owns the next decade of enterprise AI and judging by the “marketecture” of every major vendor at Next 2026, the industry has reached the same conclusion. The urgency here is clearly not a slide-ware exercise. Gartner has reported an exponential surge in enterprise inquiries about multi-agent systems and predicts that 40% of enterprise applications will embed task-specific agents by the end of 2026, up from less than 5% a year earlier. Yet the same analysts deliver an equally important counterweight: Gartner also expects more than 40% of agentic AI projects to be canceled by the end of 2027, citing escalating cost, an expanded risk surface and governance that no one built in advance. The 2026 Gartner Hype Cycle for Agentic AI makes the diagnosis plain — governance, security and FinOps capabilities are proliferating precisely because enterprises are alarmed about accountability and control as agents grow more autonomous and interconnected. Exponential demand colliding with non-existent guardrails is the environment Google walked into. So, what is the path forward to a control plane an enterprise can actually trust? What Google actually brought to Next 2026 I’m not ardent supporter of single-ecosystem architectures. That’s not the world we live in and interoperability has always prevailed as the final arbiter of truth. Beneath all the agent drama, however, Google’s message was fundamentally architectural. The company repositioned Gemini less as a standalone model and more as the connective and contextual tissue binding data systems, applications and agent runtimes, and assembled Big Query, Alloy DB, Spanner and its managed Spark service into a new category it calls the Agentic Data Cloud. As Constellation Research observed, the standardization of data on Apache Iceberg has put the data layer itself in play, and Google responded by stacking its assets into a cross-cloud lakehouse and a knowledge catalog, complete with migration tooling pointed squarely at Snowflake and Databricks. Three pillars define the offering. The first is a federated data layer built on the principle of reach, not relocation. By integrating Cross-Cloud Interconnect directly into the data plane and pairing it with the Apache Iceberg REST Catalog, Google lets agents query data residing on AWS or Azure as though it were local, with no egress fees and extends bi-directional federation in preview to Databricks’ Unity Catalog, Snowflake’s Polaris and the AWS Glue Data Catalog, according to Google’s own technical briefings and independent analysis . Google data cloud managing director Yasmeen Ahmad summarized in Google’s Next ’26 announcement with characteristic economy: you don’t move the data, you connect it. The second pillar is a semantic layer — the Knowledge Catalog, an evolution of Dataplex — which uses Gemini to tag assets, infer relationships and map business meaning so that agents are grounded rather than, as one analysis put it, fast but blind. Critically, its retrieval is permission-aware, meaning agents can only retrieve and act on assets they are explicitly authorized to see — a design choice that fuses context delivery and access control into a single operation. The third pillar is a build layer, the Data Agent Kit, which ships as portable skills, MCP tools and IDE extensions that drop into VS Code, Claude Code, Gemini CLI and Codex, deliberately declining to impose a new proprietary interface. This is a credible and, to Google’s credit, a mostly real offering. A control plane, however, is a claim, not a feature, and the term deserves more focus and detail than vendors typically provide. An agentic control plane is not a product it is a semantically governed set of domain services and underlying structured and unstructured data. How we federate agentic access and data with intention and governance means everything. What an interoperable control plane requires A control plane governs how a system behaves rather than performing the work itself. For agents, a genuine control plane must deliver at least five functions, and an interoperable one must deliver them across vendor, model and cloud boundaries rather than only within a single domain or scope. The first is identity. Agents are a new class of non-human actors, and an enterprise must be able to authenticate them and manage their actions. Microsoft’s competing Agent 365, unveiled at Ignite 2025, is built explicitly around a registry of which agents exist, plus access control and security — as an Ignite 2025 industry analysis noted, that identity is foundational. The second is context and semantics, the half of the problem the data clouds have collectively rushed toward. The third, and the most consistently underplayed, is action governance — control not merely over what an agent can read, but over what it can do: the writes, the state changes, the transactional operations. The fourth is observability and lifecycle management, the simulate-evaluate-monitor-optimize loop across an agent fleet, where Google’s integrated offering is, by most accounts, the most complete a hyperscaler has yet shipped. The fifth is economics; the reason so many projects are forecast to fail is partly cost, and FinOps for agentic AI is now an expressly named discipline on Gartner’s Hype Cycle. Interoperability cuts across all five, and here the industry has done something genuinely impactful and useful: it has agreed on protocols. The Model Context Protocol, originated by Anthropic and since donated to the Linux Foundation under multi-vendor governance, standardizes how an agent connects to tools and data. The Agent2Agent protocol, originated by Google and likewise moved to the Linux Foundation, governs how agents discover and delegate to one another across organizational boundaries. Forrester predicts that 30% of enterprise app vendors will launch their own MCP servers in 2026, and Gartner’s Anushree Verma positions standardized protocols as the enabler of the seamless interoperability that, by 2028, will let networks of specialized agents collaborate dynamically across applications. What is key here — and what enterprise leaders miss — is that open protocols deliver portable messages, not a portable control plane. Two agents can exchange tasks across clouds in A2A all day long, but identity, semantics, action governance, observability and cost remain platform functions. A2A and MCP have ensured that the communication protocol has been commoditized. The final frontier and the competitive moat is not the communication and access protocol, it is the semantic context and the business ontology Where Google is strong, and where leaders should look twice Google deserves real credit for embracing open standards where it counts. It adopted MCP across its own services, repositioned Apigee as an MCP bridge that turns any standard API into a governed agent tool and built its federation story on the open Iceberg REST Catalog rather than a proprietary format. Technology Business Research (TBR) characterized this as a meaningful strategic shift: a company historically defensive about keeping data inside BigQuery now signals that it cares less about where data physically resides than about ensuring Gemini is the semantic context layer generating value on top of it. That repositioning is exactly the lock-in risk an enterprise must carefully consider, and two limitations matter significantly and deserve an architect’s attention. The first is that federation is not the same as unified control. In my view, TBR’s analysis is totally on point: The Knowledge Catalog addresses upper-stack governance but is not an operational catalog in the way that Databricks’ Unity Catalog, Snowflake’s Polaris and AWS Glue are — those systems govern the underlying Iceberg tables. Google reads into them; it does not replace them. The second is that the focus of lock-in has simply moved up the stack to the semantic context and ontology layers. Moor Insights & Strategy and others all have cautionary tales that exiting Google-managed semantics, Gemini agents or BigQuery abstractions may prove harder than migrating the data itself. The semantics and the orchestration are now the sticky layer. I think this is a logically coherent and impressive strategy, but for every gain, something is lost. That loss is exactly the moment where an enterprise either preserves its independence or succumbs to lock-in for convenience and expedience. There is a maturity gap also worth mentioning here as well. One widely-circulated analysis of Next 2026 carried its verdict in the title — the agent stack is ready, the semantic engine isn’t — arguing that the Knowledge Catalog, however promising, is not yet the governed business-context layer a true enterprise operating system demands and remains more aspirational than operational. With much of the federation and catalog functionality still in preview, optimism is the right approach from my perspective, not “all-in” commitment. Meanwhile, the competition is playing a different game The competitors are not building the same artifact, and the differences are instructive. The data-cloud catalogs — Databricks Unity Catalog, Snowflake Polaris and Cortex, AWS Glue, Microsoft Fabric — govern data and, increasingly, semantics; the entire field now accepts that agents need context, not merely access, as industry analysis of the field documents. Their structural limit is that catalog constraints are frequently informational rather than strictly enforced, and metric-oriented semantic layers model measures rather than actions or state changes. They excel at conversing with data and remain weaker at agents that act. The agent-management planes, exemplified by Microsoft’s Agent 365, approach the problem from fleet control — registry, identity, observability — and are excellent for organizations living inside Microsoft 365, bounded by that same dependence. Palantir Foundry represents a genuinely different category. Where catalogs register tables and semantic layers define metrics, Foundry is built around an ontology that models entities, their typed relationships and the actions that can be taken against them — semantics in service of operational execution, not merely analytics. That distinction is the single most important idea for anyone designing an agentic control plane today. As Atlan frames it, a semantic layer hands agents governed metrics, which solves half the problem; agents that reason across domains and act need a knowledge-representation layer underneath — what things are, how they relate and what operations are possible. A control plane that governs reads but not writes, metrics but not actions, is fundamentally limiting for agents, and semi-autonomous action is the entire point. It is also worth noting, the semantic layer itself is now standardizing: the Open Semantic Interchange initiative, launched in late 2025 by Snowflake, dbt Labs, Salesforce and a coalition of partners under an Apache 2.0 license, finalized its v1.0 specification in early 2026. Just as MCP and A2A commoditized the agent communication protocols, OSI aims to commoditize semantic portability. A blueprint for enterprise leaders As always, the path forward is clear enough to state but extremely demanding to execute. An interoperable agentic control plane is not a product an enterprise purchases from a single vendor; it is an architecture it composes — open standards at the commoditized layers, owned assets at the differentiating one. Drawing on both the Next 2026 announcements and recent architectural engagements, I would urge leaders to prioritize four design commitments. First, standardize on open formats at the storage layer. Apache Iceberg and its REST Catalog deliver genuine data portability, and this is the one element of Google’s model worth adopting wholesale, precisely because the broader industry already has. Second, standardize on open protocols at the agent layer— A2A between agents and MCP to tools and systems — so that a Gemini agent, a Claude agent and a partner’s agent can interoperate without any one of them owning the others. Third, own the semantic and ontology layer in the middle. Don’t just model metrics but entities, relationships and the typed actions agents may perform; this is what delivers semantic portability and keeps the vendor lock-in at bay and in the enterprise’s own hands rather than a vendor. Fourth, own the control-plane core components — identity, registry, observability and policy — so that governance remains independent of any single platform. Any vendor relationship that requires managed semantics will make it intentionally harder to migrate data. The architectural response should never be to place those semantics in any single vendor’s control in the first place. Federation buys data portability; an owned ontology buys semantic portability; open protocols buy agent portability. Composed together, they close the gap that the analysts identified. The vendors will continue to make the case that the control plane is a product. The analysts — Gartner on governance and failure rates, Forrester on protocol proliferation, TBR and Moor on the limits of federation — are collectively telling enterprise leaders something more useful: it is an architectural decision, and the organizations that treat it as one, that build adaptive governance before their agentic minions outpace them and that preserve the option to change their minds, will be the ones still in command of their AI a decade from now. Google Cloud Next 2026 is a genuinely strong architecture, and there is no doubt that it is the most complete agentic control-plane offering any hyperscaler has yet shipped. It is also the clearest illustration to date of why no enterprise should outsource its control plane to anyone. The shift from owning models to owning the control plane is not just underway; for organizations serious about operating at the speed of an agent-driven business, it is inevitable. The winning move at this point is to show up with an architecture, not a purchase order. This article was made possible by our partnership with the IASA Chief Architect Forum . The CAF’s purpose is to test, challenge and support the art and science of Business Technology Architecture and its evolution over time, as well as grow the influence and leadership of chief architects both inside and outside the profession. The CAF is a leadership community of the IASA , the leading non-profit professional association for business technology architects. This article is published as part of the Foundry Expert Contributor Network. Want to join?
- 'Yesterday, a user was the weakest link. Today these agents are becoming the weakest link': Zscaler CEO Jay Chaudhry on why he believes zero trust can secure the AI agents of the present, and the future
Zscaler CEO Jay Chaudhry explains why zero trust architecture is the best way to secure AI agents
- Grounding Dify Agents in Real Data: MongoDB Atlas and Voyage AI Are Now Native to Dify RAG Workflows
Dify integrates MongoDB Atlas and Voyage AI for grounded AI agents with real business data
- What it will take to make data centres more sustainable and fit for an AI future?
The data centre energy demand needed to support AI growth is creating opportunities across many industries, but it's also putting a strain on resources.
Score: 60🌐 MovesJun 17, 2026https://www.weforum.org/stories/2026/06/build-data-centres-future-sustainable/ - June Pixel Drop brings AI video tools, smarter multitasking, and Android 17
June Pixel Drop brings AI video tools, smarter multitasking, and Android 17
- When Americans choose Chinese AI
Developers say DeepSeek is good enough for a fraction of the cost. “You don’t need God to write your email.”
Score: 60🌐 MovesJun 17, 2026https://restofworld.org/2026/when-americans-choose-chinese-ai/?utm_source=rss&utm_medium=rss&utm_campaign=feeds - "Dangerous" AI models are coming no matter what
AI models with advanced hacking capabilities will soon be the norm.
Score: 60🌐 MovesJun 17, 2026https://arstechnica.com/ai/2026/06/dangerous-ai-models-are-coming-no-matter-what/ - AI will create more jobs for humans, not replace them, Amazon founder Bezos says
The Amazon founder, who now has robotics and space travel companies, thinks AI will create a labour shortage.
Score: 60🌐 MovesJun 17, 2026https://www.bbc.com/news/articles/ceqdrw2yy3vo?at_medium=RSS&at_campaign=rss - Bringing more agent harnesses and frameworks to Cloudflare, starting with Flue
The Agents SDK is now a runtime any agent framework can build on. Today we're opening up the Agents SDK primitives, with Flue as a first framework targeting Agents SDK, and rolling out agents in the dashboard.
- Plaud reaches USD 100 million ARR in two years as AI hardware gains traction
The company expects conversations to become the starting point for AI workflows.
Score: 59🌐 MovesJun 17, 2026https://kr-asia.com/plaud-reaches-usd-100-million-arr-in-two-years-as-ai-hardware-gains-traction - CYGNVS launches command center for crises caused by a company’s own AI
Cyber resilience company CYGNVS Inc. today launched its AI Incident Command Center, a platform built to help organizations manage operational crises caused by their own artificial intelligence deployments. The product extends CYGNVS’ existing out-of-band incident platform to a new category of risk: failures in the models and agents companies are putting into production. Those failures […] The post CYGNVS launches command center for crises caused by a company’s own AI appeared first on SiliconANGLE .
Score: 59🌐 MovesJun 17, 2026https://siliconangle.com/2026/06/17/cygnvs-launches-command-center-crises-caused-companys-ai/ - India has wide AI talent gaps in deployment, governance, security: Quess report
India faces significant AI talent gaps in deployment, governance, and security roles, with GenAI deployment showing the widest deficit. While the country boasts the second-largest AI talent pool globally, the demand is shifting towards production-ready skills, particularly in the three-to-five-year experience band, as AI becomes a horizontal capability across industries.
- AI avatars in Google Vids are now available for free users
Google Vids announced today that it’s making AI avatars available to free accounts, with the Workspace video creation tool now used by 7 million users every month.
- CMU Researchers Train Robots With Internet Videos
CMU Researchers Train Robots With Internet Videos Robotics Institute Carnegie Mellon University
Score: 59🌐 MovesJun 17, 2026https://www.ri.cmu.edu/cmu-researchers-train-robots-with-internet-videos/ - Claude’s biggest Voice Mode upgrade yet is finally rolling out
Anthropic is making Claude easier to talk with.
Score: 58🌐 MovesJun 17, 2026https://www.androidauthority.com/claude-voice-mode-upgrade-rolling-out-3678314/ - MiniMax Sparse Attention (MSA): a Two-Branch Block-Sparse Attention Trained on a 109B-Parameter MoE With a 3T-Token Budget
MiniMax Sparse Attention (MSA): a Two-Branch Block-Sparse Attention Trained on a 109B-Parameter MoE With a 3T-Token Budget MarkTechPost
- Alibaba Cloud launches data centres in France amid Europe’s data sovereignty push
Alibaba Cloud announced on Wednesday that it has launched its first data centres in France, as it continues to expand its footprint in Europe amid growing calls on the continent for data sovereignty. The move came as a response to increasing demand from customers in the region, the cloud computing services unit of Alibaba Group Holding said in a statement. “The expansion of our cloud infrastructure into France reinforces our ongoing commitment to empowering European businesses with sovereign,...
- Meta pursues muddled AI strategy
With massive borrowing to build data centers and without and enterprise business like Anthropic’s or OpenAI’s, the company’s AI strategy still doesn’t seem to make sense.
Score: 58🌐 MovesJun 17, 2026https://www.semafor.com/article/06/17/2026/meta-pursues-muddled-ai-strategy - Amazon AI exec predicts first 'commercially useful' quantum computers in 5-7 years
Quantum computing is becoming an increasingly competitive field, with tech giants including Microsoft, Google and IBM developing the technology.
- AI coding agents taught robots how to install GPUs and cut zip ties
Nvidia's self-improvement program for robots enlists teams of AI coding agents.
Score: 57🌐 MovesJun 17, 2026https://arstechnica.com/ai/2026/06/ai-coding-agents-can-autonomously-direct-robot-training/ - Ex-Cisco researchers launch Tenet Security to lock down rogue AI agents
Former Cisco artificial intelligence security researchers have launched a new company to tackle a problem that barely existed a year ago: securing the autonomous AI agents enterprises are handing the keys to their most critical systems. Tenet Security Inc. today formally launched a platform designed to stop malicious AI agent behavior before it reaches production […] The post Ex-Cisco researchers launch Tenet Security to lock down rogue AI agents appeared first on SiliconANGLE .
Score: 57🌐 MovesJun 17, 2026https://siliconangle.com/2026/06/17/ex-cisco-researchers-launch-tenet-security-lock-rogue-ai-agents/ - Exclusive: Meta head of product for 'AI for work' transformation is leaving company
Exclusive: Meta head of product for 'AI for work' transformation is leaving company Reuters
Score: 57🌐 MovesJun 17, 2026https://www.reuters.com/world/meta-head-product-ai-work-transformation-is-leaving-company-2026-06-17/ - Jabil raises annual profit forecast amid strong data center demand
Jabil has raised its profit forecast for 2026, fueled by a surge in demand from AI-powered data centers. This boom in infrastructure investment is proving advantageous for Jabil and similar firms. Additionally, the company reports unexpectedly robust results in its automotive and connected living divisions. As a result, Jabil's stock experienced a modest uptick in premarket trading.
- Autonomous AWS Agent Automates Modernization of Codebases
Autonomous AWS Agent Automates Modernization of Codebases DevOps.com
Score: 56🌐 MovesJun 17, 2026https://devops.com/autonomous-aws-agent-automates-modernization-of-codebases/ - Gnani.ai Doubles Down On Sovereign Voice AI Models With Prisma v2.5 Launch
Enterprise voice AI startup Gnani.ai has launched its latest model called Prisma v2.5. A speech-to-text (STT) model, Prisma spans across…
Score: 56🤖 ModelsJun 17, 2026https://inc42.com/buzz/gnani-ai-doubles-down-on-sovereign-voice-ai-models-with-prisma-v2-5-launch/ - Context Windows Are the New RAM: Memory Architecture for Agentic Systems
Your agent isn’t dumb. It’s just forgetful. Here’s how to fix that. There’s a quiet crisis happening inside every production AI agent right now. It isn’t hallucinations. It isn’t latency. It isn’t even cost — though that one stings. It’s memory. More precisely, it’s the complete lack of a coherent memory model. Most agentic systems today are architected the same way early computers were before virtual memory existed: one flat address space, no hierarchy, no eviction strategy, no concept of what’s hot versus cold. Shove everything into the context window and pray it fits. That approach worked when agents were glorified chatbots answering single-turn questions. It doesn’t work when you’re running a multi-step research pipeline, a code review agent that spans dozens of files, or an autonomous workflow that must remember what it decided three hours ago. The engineers who figure out memory architecture — real memory architecture, borrowed from decades of systems thinking — are going to build agents that actually work in production. Everyone else will keep hitting the 128K token wall and calling it a “limitation of the technology.” It isn’t. It’s a design failure. The RAM Analogy Is More Literal Than You Think When computer architects in the 1960s and 70s were designing memory hierarchies, they faced a surprisingly similar problem to what AI engineers face today. Compute was expensive. Fast memory was extremely expensive. Slow memory was cheap but too slow to be useful on its own. And programs needed to operate on more data than could fit in the fast tier. Their solution was the memory hierarchy: registers → L1 cache → L2 cache → RAM → disk. Each tier trades speed for capacity. The operating system’s job is to keep the most relevant data in the fastest tier at any given moment — a problem called cache management . Now look at what an LLM agent actually operates on: The active context window — fast, immediately accessible, brutally limited in size, expensive per token Retrieved chunks — slower to access, require a retrieval step, but can cover far more ground Persistent storage — slowest, but unlimited, and survives across sessions This is a memory hierarchy. We just haven’t been treating it like one. The context window is not a document. It’s a cache. And like any cache, it has a hit rate, an eviction policy, and a capacity budget. The moment you start thinking about it that way, a completely different set of engineering tools becomes available to you. What Actually Lives Inside a Context Window Before you can architect memory well, you need to understand what’s consuming your context budget right now. In a typical agentic system, context tokens are spent roughly as follows: System prompt and persona — 500 to 3,000 tokens. Often bloated with redundant instructions, lengthy examples, and rules that could be expressed in a third of the space. Tool definitions — 200 to 2,000 tokens per tool. Ten tools with verbose schemas can silently consume 15,000 tokens before a single user message lands. Conversation history — the big one. In multi-turn agents, the raw transcript of every prior turn gets appended verbatim. By turn 20, you’re carrying 40,000 tokens of history that’s 80% irrelevant to what the agent needs to do right now. Retrieved context — chunks pulled from RAG, document stores, or memory databases. Often retrieved with poor precision, so you pull 8 chunks when 2 would have sufficed. Working scratchpad — the agent’s chain-of-thought, tool call outputs, intermediate reasoning. Useful in the moment, wasteful to carry forward. The current task — what the user actually asked. Often less than 200 tokens, buried under everything above. Now ask yourself: if this were an operating system’s memory map, would a kernel engineer approve it? Almost certainly not. You’re running with no cache eviction, no priority scheme, and no separation between hot and cold data. The Four-Tier Memory Architecture Here’s the architecture pattern that serious agentic systems are converging on. Think of it as the memory hierarchy for AI — each tier has a distinct role, access cost, and management strategy. Tier 1: In-Context Working Memory This is your L1 cache. It’s what the model can “see” right now — fast, zero-latency, directly attended over. Treat it like a precious, scarce resource. What belongs here: the current task, the most recent tool outputs, active intermediate reasoning, and the minimal prior context needed to maintain coherence. What does not belong here: full conversation history, raw document dumps, every tool schema whether or not it’s relevant to this step. The discipline required here is aggressive pruning. Summaries instead of full transcripts. Filtered tool schemas instead of the entire definition list. Compressed scratchpad outputs rather than verbose chains. A useful mental model: the in-context window is your agent’s short-term working memory. Cognitive science tells us humans can hold roughly 7 ± 2 items in working memory. Your agent’s context has a higher ceiling, but the principle holds — the more you cram in, the worse the reasoning quality becomes, independent of token limits. Attention dilutes. Tier 2: Episodic Memory (Session Store) This is RAM. It persists across turns within a session, but not necessarily across sessions. It holds the running narrative of what has happened: decisions made, steps completed, information discovered. The key insight is that episodic memory should store summaries and conclusions , not raw transcripts. At each major decision point or completed subtask, the agent writes a compressed episode record: Episode 7 — Research phase complete - Queried 3 sources on topic X - Key finding: Y is the dominant approach as of 2025 - Contradicting claim found in source Z — flagged for verification - Next planned action: synthesize findings into outline When the agent needs historical context, it retrieves episode summaries, not raw history. This keeps Tier 1 clean while preserving continuity. Tier 3: Semantic Memory (Vector Store) This is your disk — slow relative to context, but vast. Semantic memory holds facts, knowledge, and learned associations retrieved via embedding similarity. Most engineers know this as RAG. But RAG as typically implemented is like having a disk but no filesystem — you can store things, but retrieval is a shotgun rather than a scalpel. Better semantic memory systems include: Typed namespaces — separate indexes for different knowledge types (domain knowledge, user preferences, tool documentation, past task outcomes). Don’t mix everything into one flat vector space. Confidence-weighted storage — not all retrieved facts are equally reliable. Tag memory entries with a source, a timestamp, and a confidence score. Stale or low-confidence entries should be surfaced with appropriate hedging. Write-back from episodes — when an agent discovers something durable (a user’s preferred output format, a domain fact that took three steps to uncover), it should write that back to semantic memory explicitly, not just leave it buried in a session log. Tier 4: Persistent Procedural Memory This is the BIOS — rarely accessed, but foundational. Procedural memory holds how to do things: task templates, successful past approaches, learned heuristics. If your agent successfully completed a complex multi-step data pipeline last week, that workflow pattern is worth storing. Not the specific data — the approach. Next time a similar task arrives, the agent can retrieve the procedural template and adapt it rather than reasoning from scratch. This is where agentic systems start to compound. An agent that learns its own successful patterns becomes meaningfully better over time, not just through model updates, but through accumulated operational knowledge. The Cache Management Problem: What to Keep, What to Evict Having four tiers is necessary but not sufficient. You need eviction policies — rules for deciding what gets promoted to the expensive fast tier and what gets demoted or discarded. Several strategies from classical systems apply directly: Recency (LRU — Least Recently Used) — the oldest unused context gets evicted first. Easy to implement, works reasonably well as a default, but blind to importance. Relevance scoring — compute a similarity score between each context chunk and the current task. Low-relevance chunks get evicted even if they’re recent. This is the AI-native equivalent of priority-based scheduling. Importance tagging — at write time, tag certain context items as non-evictable: explicit user constraints, hard requirements, safety guardrails. These survive regardless of relevance score. Summary compression — before evicting a chunk, summarize it into a compact form and store the summary instead. You lose detail but preserve the gist. This is analogous to a compressed cache tier. Decay with anchoring — implement time-decay on memory relevance scores, but anchor certain facts that the user has explicitly confirmed. User-confirmed information should have near-zero decay. A practical implementation cycles through these at each major agent step: score everything in the episodic store, evict or compress below-threshold items, retrieve high-relevance items into Tier 1 for the current action. The Write Problem: Agents That Learn Most agentic memory architectures focus heavily on the read side — how do we retrieve the right context? — and almost entirely ignore the write side: how does the agent know what’s worth remembering? This is a mistake. An agent that reads memory well but writes nothing useful is just a sophisticated cache that never gets warmer. Principled write strategies: Task completion writes — at the end of any completed subtask, always write an episode record. This is your commit log. Contradiction detection writes — when the agent encounters information that conflicts with existing memory, write a conflict record. Don’t silently overwrite; surface the inconsistency. User correction writes — when a user corrects the agent, that correction should propagate to memory, not just affect the current turn. This is where most systems fail — they treat corrections as one-off context rather than durable knowledge updates. Uncertainty writes — when the agent makes a decision under uncertainty, log the uncertainty alongside the decision. “Chose approach A because B and C were unavailable, but B would be preferred if accessible” is worth storing. It changes how you interpret the outcome later. Practical Implementation: A Minimal Architecture Here’s what a working four-tier memory system looks like in code terms, without the academic abstraction: python class AgentMemory: def __init__(self): self.working_memory = [] # Tier 1: assembled per-action self.episode_store = EpisodeDB() # Tier 2: session-scoped SQLite or Redis self.semantic_store = VectorDB() # Tier 3: Pinecone, Weaviate, pgvector self.procedural_store = TemplateDB() # Tier 4: structured task templates def assemble_context(self, current_task, token_budget=4096): context = [] remaining = token_budget # Always include: current task + hard constraints context.append(current_task) remaining -= count_tokens(current_task) # Retrieve: relevant episodes (scored by similarity) episodes = self.episode_store.retrieve( query=current_task, limit=5, score_threshold=0.7 ) for ep in episodes: if remaining > count_tokens(ep.summary): context.append(ep.summary) remaining -= count_tokens(ep.summary) # Retrieve: semantic knowledge facts = self.semantic_store.query(current_task, top_k=3) for fact in facts: if remaining > count_tokens(fact): context.append(fact) remaining -= count_tokens(fact) return context def commit_episode(self, action, outcome, metadata): summary = self.summarize(action, outcome) self.episode_store.write(Episode( summary=summary, timestamp=now(), importance=metadata.get("importance", 0.5), task_type=metadata.get("task_type") )) # Write durable facts to semantic store if outcome.contains_durable_knowledge: self.semantic_store.upsert(outcome.extracted_facts) This isn’t production code — it’s a structural sketch. But it illustrates the key point: context assembly is an active process with a budget constraint, not a passive accumulation. What You’re Actually Paying For Let’s be direct about the economic dimension here, because it changes the calculus on how much engineering effort to invest. At current pricing for frontier models, a 128K context window request costs roughly 20 to 50x more than a 4K request, depending on the model. For an agent running 100 steps per task with naive context management, that’s a cost multiplier of 20–50x compared to an agent with well-managed working memory. That’s not a minor optimization. For any agent running at scale, memory architecture is a direct cost driver — as significant as model selection. Beyond cost: attention quality degrades as context grows. There’s substantial evidence, both empirical and from architecture fundamentals, that models perform worse on tasks buried deep in large contexts compared to tasks presented with minimal surrounding noise. The “lost in the middle” phenomenon is well-documented. Bloated context isn’t just expensive; it’s actively harmful to reasoning quality. Good memory architecture is therefore not a nice-to-have. It simultaneously reduces cost and improves quality. That’s a rare double win in systems engineering. The Deeper Point: Agents Need a Memory OS We are, right now, at roughly the stage of computing history where memory management was manual. Early C programs called malloc and free by hand, and memory bugs were endemic — leaks, overwrites, use-after-free. The innovation of garbage collection and managed memory didn't make programs slower; it made them more reliable and let engineers focus on the actual problem. Agentic AI systems need the same transition. Right now, every team building a serious agent is reinventing their own ad hoc context management strategy. Some are using sliding windows. Some are manually summarizing. Some are just running into walls and calling it a model limitation. What the field is converging toward — and what will become standard infrastructure within the next two to three years — is a memory operating system layer: a runtime that sits between the application logic and the LLM API, managing tier promotion and demotion, handling eviction policies, providing atomic writes and reads across the hierarchy, and exposing a clean interface that lets agent logic focus on tasks rather than token budgets. The engineers building that layer today are working on what will become invisible infrastructure tomorrow — the way TCP/IP, virtual memory, and the filesystem are invisible today. You don’t think about them. You just trust they work. The agents that will define the next five years of AI products won’t be remembered for their model choices. They’ll be remembered for making memory invisible — and therefore making intelligence feel effortless. Key Takeaways The context window is a cache, not a document. Treat it like one. Every serious agentic system needs four memory tiers: working memory, episodic memory, semantic memory, and procedural memory. Context assembly should be active and budget-constrained, not passive accumulation. Eviction policies — recency, relevance scoring, importance tagging, summary compression — are not optional at scale. The write problem matters as much as the read problem. Agents that learn from their own operations compound in capability over time. Memory architecture is simultaneously a cost reduction and a quality improvement. No other optimization delivers both. Context Windows Are the New RAM: Memory Architecture for Agentic Systems was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.