Cognition introduces Dreaming: Devin refines cross-session memory overnight

#1 𝕏

Cognition introduced Dreaming, a capability that lets Devin build a cross-session memory graph of how users like to work and improve it at night by removing stale records and discovering latent information. Cognition is also creating an open-source standard called Agent Memory Repo.

Also covered by: @Harrison Chase

#2 𝕏

Cursor announced that steering and system prompts now work with local TypeScript agents, while background subagent results return locally in TypeScript and Python. The Cursor documentation includes the full SDK changelog.

Also covered by: @Cursor

#3 𝕏

NVIDIA AI announced that Nemotron-Labs-3-Competitive-Coding-550B-A55B-NVFP4 is available on Hugging Face. It is described as a competitive-programming specialist based on Nemotron-3-Ultra.

#4 𝕏

clem 🤗 says their group turned Claude Code, Codex, Hermes, Pi, @opencode, and other coding harnesses into RL environments. The capture proxy in OpenEnv, trainer in TRL, tasks, SFT data, training code, and all 7 trained models are open and reproducible, with bigger models and runs planned next.

#5 📝 OpenAI News

Our approach to EU text provenance rules - OpenAI shared its approach to EU text provenance rules on October 5, 2026, aiming to promote transparency within current technological limits. The article’s listed listening time is 7:42, and it says content provenance helps people understand content’s origins, how it was created or edited, and whether it contains signals associated with OpenAI’s models.

Also covered by: @OpenAI, @OpenAI

#6 ▶️

Build an AI Language Tutor You Can Talk To (6 Steps)

Peter Yang

#7 𝕏

Guillermo Rauch announced gdp-ts, a TypeScript library, linter, and AI skill for safer API design that requires authorization “proofs” for sensitive functions and verifies them at compile time. Its README and examples model a Vercel API constraint requiring proof of a certain role and entitlement before changing a Project password.

Also covered by: @Guillermo Rauch

#8 𝕏

Santiago shared Era, an agent-testing app that generates a complete synthetic company from a one-line business description across CRM, tickets, Slack, files, emails, and call recordings. Builders can test agents against the connected data and reset the company to its initial state if something breaks.

#9 in

Marc Baselga recapped how AI agents inside Stripe behaved unpredictably and overwhelmed core systems, noting that Stripe’s internal agent Kai is used weekly by more than 10,000 employees. Stripe responded with separate agent identities, priority signals, load shedding, and a query layer built to absorb added demand.

#10 𝕏

Claire Vo shared a video workflow using an iPhone with Riverside or Zoom, OpenAI GPT-6 astra and ElevenLabs MCP for horizontal and vertical cuts, exporting to Claude and using Opus 5.5 for captions and overlays, with Muse posting across Instagram, YouTube Shorts, LinkedIn, and other platforms.

#11 in

Claire Vo recapped updates to Sites presented in connection with OpenAI dev day, highlighting connected enterprise data, user permissions, and site-as-infrastructure for deploying micro SaaS tools. Her discussion with product lead Kathy K. also covered consumer use cases and how Sites can host more than webpages.

#12 𝕏

LlamaIndex 🦙 shared a breakdown by its Head of Open Source, Logan Markewich, contrasting traditional one-pass OCR with agentic OCR’s iterative approach using layout-aware reading order, model routing, multi-pass verification and self-correction, and multimodal parsing. The breakdown also covers agentic OCR’s remaining struggles.

#13 in

Anu Jagga Narang shared a post about setting up an Obsidian vault with multiple LLMs, describing how she connects Claude Code and ChatGPT to the same vault and assigns them interchangeable builder and reviewer roles. She presents Markdown files as an easy way for non-technical product managers to manage context without heavy architecture.

#14 ▶️

The 5 Layers of Enterprise AI, Explained | David Hogan, NVIDIA

SyntaxGTM

#15 𝕏

Santiago shared Nolla Health’s AI-led acne prescription and treatment app, which operates in Utah and uses a face scan and a few questions to create a personalized plan, provide a prescription if necessary, and adjust treatment over time. Users also have access to a physician at no extra cost.

#16 𝕏

Thariq calls the pattern “local hands”: Claude runs in the cloud while accessing local files. He says it’s also coming to “cowork.”

#17 𝕏

Cognition shared that Agent Memory Repo stores memory records in files versioned with Git, enabling use cases such as a message board across agent sessions.

#18 𝕏

Thariq shared commands to add the `anthropics/claude-plugins-community` marketplace and install `html-plan@claude-community`, asking users for feedback and linking to an example on claude.ai.

#19 𝕏

Garry Tan commented that lab-developed harnesses are incentivized to burn tokens, giving startup-developed harnesses utility. He said Grep can convert token burn into deterministic, tested, repeatable code by observing agent use, while quoting a GrepdotAI agent workforce startup announcement.

#20 in

Peter Yang shared a clip from his dots review outlining three areas he would simplify: Work versus Codex, Spaces versus Pages versus Sites, and the model and effort picker. He argued that the Work launch was a mistake and that Work does not need its own brand.