Weekly Bookmarks

These are some things I’ve wandered across on the web this week.

🔖 sliver

An ‘archival sliver’ of the web. A bit like a ‘data lifeboat’ for making or replicating web archives of small sets of pages. Uses shot-scraper to drive a web browser that generates screenshots of your URLs, but runs it through a pywb web proxy so it can produce a high quality archival version of what you download.

As well as archiving live web pages, this tools can leverage pywb’s support for neatly extracting URLs from other web archives and recording items with all the appropriate provenance information (see below for an example). This means it can work like hartator/wayback-machine-downloader but retain the additional information that the WARC and WACZ web archiving format suppor

🔖

Justice Department deletes press releases on charges against Jan. 6 rioters

The Justice Department has removed press releases detailing the charges against hundreds of individuals who participated in the Jan. 6, 2021 Capitol riot from its website, the department confirmed Friday.

🔖

Meryl Kornfield on deletion of justice.gov web content

The Trump admin is quietly deleting info about the Capitol attack from the DOJ website as it prepares to give funds to J6ers. This week, DOJ deleted a press release about one man with an ongoing child solicitation case who came to the Capitol with bear spray.

🔖

On Tools and the Normalization of Evil

The scale of theft is unreal, if one person or company plagiarizes something, lawsuits and court filings often ensue, or at the very least some reputational damage to the perpetrators, but not for Anthropic or OpenAI, not for Google or Microsoft, they steal from all of us and then they sell our work back to us. They want to keep us dumb and uneducated, they want us to rely on them. Learning is power, learning is resistance, knowledge provides independence.

🔖

LLMs and Buttondown

Our month-over-month growth rate in Q1 2026 was double our growth rate in Q4 2025. Buttondown has, roughly, grown a little less than 2x every year of its existence; this — its eighth year — is poised to shatter that, if trends hold.

Almost all of that incremental growth, meaning the growth in addition to our historical trend, I attribute to LLMs. We ask people when they sign up what brought them here, and an answer that went from surprising to banal to overwhelming over the course of Q1 was: an LLM. Users of all stripes cite an LLM as the reason that they ended up at Buttondown’s front door.

🔖

GitHub Breach Traced to Malicious ‘Nx Console’ VS Code Extension

GitHub has confirmed that a recent breach into its internal repositories was caused by a vulnerability in a Microsoft Visual Studio Code (VS Code) extension called ‘Nx Console.’

The security team at the Microsoft-owed software developer platform warned on May 19 that an attacker gained unauthorized access to 3800 internal repositories via a “poisoned” VS Code extension found on an employee device.

It was later confirmed by Jeff Cross, CEO of Nx that Nx Console, a popular VS Code extension, was the extensions that was poisoned extension and resulted in the GitHub breach.

🔖 ReS

Futurae

ReS Futurae est une revue francophone internationale dédiée à l’étude de la science-fiction sous toutes ses formes : littérature, cinéma, arts graphiques, jeux vidéo, musique, design et phénomènes culturels divers. C’est une revue académique, à comité de lecture et arbitrage par les pairs, fondée sur un partenariat avec la revue Science Fiction Studies : des traductions croisées d’articles acceptés dans l’une et l’autre revue seront publiées régulièrement. Dans le paysage académique francophone, ce sera la première revue de cette nature.

🔖

Starlight

Starlight is a documentation website framework for Astro.

🔖

Tell New York Times, The Atlantic, and USA Today to keep the crucial work of journalists in the Wayback Machine!

The freedom of journalists isn’t only the freedom to write, it’s also the freedom to have your work read and remembered for generations to come. 2026 is the first World Press Freedom Day in 30 years that journalists’ work at major media outlets including New York Times, The Atlantic, and USA Today is not being preserved by the independent, nonprofit Internet Archive. We are calling on you and on all news outlets to publicly commit to working with the Internet Archive to keep the news in the Wayback Machine.

🔖 Hands-On Large

Language Models

Through the visually educational nature of this book and with over 250 custom made figures, Python developers will learn the practical tools and concepts they need to use Large Language Models today.

🔖

Hiroshi Yoshimura

Hiroshi Yoshimura (吉村弘, Yoshimura Hiroshi; 22 October 1940 – 23 October 2003) was a Japanese musician and composer. He is considered a pioneer of ambient music in Japan.[2][3] His music lies mostly in the minimalist genre of kankyō ongaku, or environment music—soft electronic melodies infused with the sounds of nature: babbling brooks, steady rain, and morning birds.[4] However, not all Yoshimura’s work included nature sounds. His album Green (1986) only contained them in the United States release, as they were excluded in the Japanese version.

🔖

Yves Tanguy

Tanguy’s paintings have a recognizable style of nonrepresentational Surrealism. They show vast, abstract landscapes, mostly in a tightly limited palette of colors, occasionally showing flashes of contrasting color accents. Typically, these landscapes are populated with various abstract shapes, sometimes angular and sharp, sometimes with an organic look to them.

🔖

Announcing Web Serial Support in Firefox

Web Serial is a web API that allows a website to read and write to serial devices using JavaScript. See the MDN documentation for the details. While modern computers don’t typically include serial ports, serial devices connected to a USB port or paired via Bluetooth can advertise themselves as serial-capable devices so they appear as serial ports in the operating system.

The Web Serial API lets developers use the web platform to communicate with these devices. For example, websites can control devices or deliver firmware without requiring native applications or installers.

🔖

Weeds tend not to grow where they can’t take root

Destroying AI must include building counter-structures and nurturing a healthy, thriving social landscape that denies AI projects access to us in the first place. AI solutions like therapy & medical chatbots find space to thrive because of all the gaps in medical care we’ve normalized; we must make these interventions totally inscrutable in a future where care is always available, and people’s needs are not constantly being means-tested and scrutinized.

🔖 YesWeScan

Got an old USB scanner your computer can’t talk to? This web app is for you. Connect your scanner (see above) and get scanning.

🔖 Langfuse

Building AI applications and agents is very different from traditional software. Outputs are probabilistic, and teams need to reason about quality, cost, latency, and the tradeoffs between them. Langfuse Academy explains the AI engineering lifecycle to help you understand how the pieces fit together and what it takes to ship from prototype to production.

🔖

pocket_archive

Pocket Archive is a digital archival system and static site generator for small- to medium-(?) sized archives. It is designed to function in environments with unreliable connectivity and requires very low technical and human resources to set up, run, and use.

🔖 Memory in

the Age of AI Agents

Memory has emerged, and will continue to remain, a core capability of foundation model-based agents. As research on agent memory rapidly expands and attracts unprecedented attention, the field has also become increasingly fragmented. Existing works that fall under the umbrella of agent memory often differ substantially in their motivations, implementations, and evaluation protocols, while the proliferation of loosely defined memory terminologies has further obscured conceptual clarity. Traditional taxonomies such as long/short-term memory have proven insufficient to capture the diversity of contemporary agent memory systems. This work aims to provide an up-to-date landscape of current agent memory research. We begin by clearly delineating the scope of agent memory and distinguishing it from related concepts such as LLM memory, retrieval augmented generation (RAG), and context engineering. We then examine agent memory through the unified lenses of forms, functions, and dynamics. From the perspective of forms, we identify three dominant realizations of agent memory, namely token-level, parametric, and latent memory. From the perspective of functions, we propose a finer-grained taxonomy that distinguishes factual, experiential, and working memory. From the perspective of dynamics, we analyze how memory is formed, evolved, and retrieved over time. To support practical development, we compile a comprehensive summary of memory benchmarks and open-source frameworks. Beyond consolidation, we articulate a forward-looking perspective on emerging research frontiers, including memory automation, reinforcement learning integration, multimodal memory, multi-agent memory, and trustworthiness issues. We hope this survey serves not only as a reference for existing work, but also as a conceptual foundation for rethinking memory as a first-class primitive in the design of future agentic intelligence.

🔖 Is Grep

All You Need? How Agent Harnesses Reshape Agentic Search

This paper reports an empirical study organized into two experiments. Experiment 1 compares grep and vector retrieval on a 116-question sample from LongMemEval, using a custom agent harness (Chronos) and provider-native CLI harnesses (Claude Code, Codex, and Gemini CLI), for both inline tool results and file-based tool results that the model reads separately. Experiment 2 compares grep-only and vector-only retrieval while progressively mixing in additional unrelated conversation history, so that each query is embedded in more distracting material alongside the passages that matter. Across Chronos and the provider CLIs, grep generally yields higher accuracy than vector retrieval in our comparisons in experiment 1; at the same time, overall scores still depend strongly on which harness and tool-calling style is used, even when the underlying conversation data are the same.

🔖

Trevor Paglen and Holly Herndon on Making Art with AI and What the Discourse Is Missing

Neither Paglen nor Herndon are AI “skeptics”—they both use the various machine-learning technologies discursively bundled up as “AI” throughout their practices—but neither are they full-blown enthusiasts. So how is it changing their sense of what art is, and how we produce it? In the conversation that follows, I posed that question to them. “I think both Trevor’s practice and ours are looking at infrastructure in a really deep way,” Herndon said. “It was important in the early days, when we were beginning to experiment with this stuff, to see artists we had great respect for, like Trevor, working with it as well. It was like, OK, you’re not crazy—this is a really fruitful area to explore.”

🔖

Moving away from Tailwind, and learning to structure my CSS

I spent the last week or so migrating a couple of sites away from Tailwind and towards more semantic HTML + vanilla CSS, and it was SO fun and SO interesting, so here are some things I learned!

🔖 Wendt Center

Since 1975, the Wendt Center for Loss and Healing has helped people in the Washington metropolitan area rebuild a sense of safety and hope after experiencing the death of a loved one, life-threatening illness, violence, or other trauma. Nationally recognized for our expertise in grief, trauma, and mental health, we provide an array of holistic services for children, teens, adults, families, and our local communities.

🔖

OpenWebUI

Open WebUI is an extensible, feature-rich, and user-friendly self-hosted AI platform designed to operate entirely offline. It supports various LLM runners like Ollama and OpenAI-compatible APIs, with built-in inference engine for RAG, making it a powerful AI deployment solution.

🔖

FaultLine

Most AI memory systems trust the LLM to write whatever it extracts. FaultLine doesn’t — every fact passes a validation gate before it touches storage. It’s the only system in the field that treats the model as an untrusted writer by design.

🔖

fastino-ai / GLiNER2

GLiNER2 unifies Named Entity Recognition, Text Classification, Structured Data Extraction, and Relation Extraction into a single 205M parameter model. It provides efficient CPU-based inference without requiring complex pipelines or external API dependencies