Atomic Strata

We just open-sourced AtomicMemory. The AI memory industry has a black-box problem. AtomicMemory is a configurable open-source SDK + self-hosted Core engine for memory your AI can inspect, correct, swap, and run on your own infrastructure. Apache 2.0. HTTP-first. Docker quickstart. github.com/atomicstrata

@OpenAI Cross-functional agents access different sources of truth. Memory with add, update, and delete semantics ensures each department's context stays accurate and does not contaminate the others. Atomic Memory is the configurable fix for AI memory github.com/atomicstrata/a…

Work at OpenAI is being transformed by agents, in every department. Across our entire company, people are using Codex to do work that is more complex, longer-running, and increasingly cross-functional. Our internal usage offers an early look at how agentic tools may reshape work as they become more capable and broadly available.

@polydao Separating builder from checker stops the agent from grading its own homework. Structured memory stops both agents from silently agreeing on wrong context.

CLAUDE CODE + SELF-CORRECTING LOOP = THE FASTEST WAY TO BUILD AND MONETIZE APPS most developers burn thousands of dollars manually debugging this setup forces Claude Opus 4.8 to self-correct - replacing a $95k/year QA engineer with a loop that costs $20/month: > Builder agent writes code while Checker agent runs tests (separating them eliminates blind spots) > Orchestrator loops them until all tests pass, capped at 5 runs to save tokens > Regression guard blocks loops if a fix breaks previously passing tests without loop guards, agents will literally edit your test suite to force a pass bookmark this guide to set up the exact loop commands 👇

@shmidtqq The self-review problem is compounded when the agent also decides what to remember. Memory governance with lineage and correction semantics is the real second opinion.

Anthropic, Claude Agent SDK: "Agents build their own context, decide their own trajectories, and work autonomously." In a 1h52m workshop, Thariq Shihipar builds one from scratch, live, and shows exactly where it breaks. Gather context. Take action. Verify the work. One AI that writes its own code, then tests it, then reviews it is grading its own homework. Would you ship that? Save this before your next build.

@TheHunterBohm This is the exact use case for configurable memory 👀 Your body changes week to week and the agent needs to update what it knows without starting from zero. Try to pair your agent with Atomic Memory! github.com/atomicstrata/a…

Hermes is my new personal trainer Integrated with all my health data, sleep, steps, workouts etc I just tell it what I ate, and when I’m training and it knows exactly how I’m feeling and what’s right for me today Genuinely makes my life better & healthier Another thank you @Teknium I’m down 10lbs since starting and tracking this.

@h100envy before people build loops, they should find a way first how to ensure the context memory of the AI remains auditable

An Anthropic senior dev just dropped an 11-page breakdown on "Loop Engineering." The game has changed: you don’t prompt the agent anymore, you build the system that prompts it. The 5-step loop : Discovery: The agent hunts down its own tasks (failing CIs/issues) instead of waiting for a list. Handoff: Tasks run in isolated git worktrees so parallel agents don't clash. Verification: A separate agent reviews the code assuming it’s broken. (Rule #1: an agent grading its own work always gives itself an A). Persistence: Results save to disk so they don't vanish when the context window clears. Scheduling: A timer auto-wakes the system, keeping the loop alive. This PDF completely flipped my approach to building AI agents. Read it now, then check the article below.

@fchollet the era of tastemaxxing is upon us

When the cost of execution drops, the value of taste, strategy, and architectural vision skyrockets. Previously, you were spending most of your cognitive budget on the micro. Now you are free to focus on the macro.

@emollick Skills define the behavior but agents also need memory governance so they act on correct context across every domain.

This is a fascinating and important set of data which shows us where things are going, using OpenAI as a canary in the coal mine. The chatbot era is over, and agentic systems are coming to tasks beyond engineering. And skills show promise as a way to standardize AI use in firms.

@0xCodez Very worth taking note of is how loop engineering still does not solve information rot

A senior Anthropic engineer just dropped 11-page PDF on "Loop Engineering" for agentic systems. The shift: you stop prompting the agent. You build the system that prompts it instead. Schedule → Discover → Build → Verify → Repeat Every loop runs one turn, five moves: • Discovery: it finds its own work - failing CI, open issues, recent commits - instead of being handed a list. • Handoff: each task gets an isolated git worktree so parallel agents don't collide. • Verification: a second agent, told to assume the code is broken, reviews the first. The "thing that can say no." • Persistence: results get written to disk, never left in a context window that gets flushed. • Scheduling: an automation wakes it on a timer. That's what makes it a loop. The key insight: an agent grading its own work always praises it. This 11-page PDF changed how I'm building agentic systems today. Read it now, then explore the article below.

@chetaslua Would you be willing to trade your personal information for AI use? 🤔

@Zephyr_hg Building autonomous systems means builders have to take their memory layer seriously, or else their system wouldn’t last long-term

The creator of Claude Code, Boris Cherny: "I have a Claude that prompts other Claudes. So I don't even talk to Claude." In a 57-minute conversation, he describes how the best AI work stopped being hands-on. The job moved. Now it's running the system that does the task for you. That shift is creating a role that pays like management and didn't exist two years ago. Watch the talk, then see the role mapped out in the article below. Bookmark this one.

Three quiet failure points break almost every knowledge base: 1) Generated content goes wrong because the model can misread a source or connect ideas that were never actually related. 2) An imported content goes wrong because another tool's extraction mistakes come along with it. 3) A page correct in March can be wrong by June if the source underneath it changed. Most tools trust their own output directory by default, which means all three problems ship straight to the agent unchecked. Run LLMwiki on your own sources and see the trust signals in action. github.com/atomicstrata/l…

@0xJeff When the loop learns something incorrect, how do you find it and correct it before it compounds? 🤔

Loop takes AI from a chatbot to an autonomous agent ​ Instead of human providing feedback, the AI self-learns, evaluates, and provides feedback on its own output, iterating time and time again to get to a desirable goal ​ In this article, we dive into the 2 loops that are highly useful for research workflows, incrementally improving the quality of the outputs every single day ↓

@notthreadguy If the AI is smarter than you that probably means you are using it wrong. 😂

i’ve come to the conclusion the more i talk to ai the smarter i am. not in a one shot psychosis way but realizing i walk through life with infinite unanswered questions i now get answered because it takes 5 seconds

@vivoplt The workflow of re-explaining preferences every session is a sign the memory layer does not support revision. Corrections need to supersede old instructions, not just sit alongside them.

Me 90% of the time with Claude: - “That’s not what I meant.” - “Please don’t refactor the entire project.” - “Why did you delete that?” - “Keep the existing logic.” - “No, I only asked for one small change.” - “Stop creating new files.” - “Use the function that’s already there.” - “Okay… let’s try this again.” At this point, I’m not coding anymore. I’m just managing an extremely confident junior developer.

@mattpocockuk Same principle applies to agent memory 🧠 Paying attention to what gets stored and revisiting it regularly beats letting it accumulate unchecked.

Usually the more of a skill I've handwritten, the better it is Paying attention to each word as you write it is unreasonably effective

@signulll Every new workflow a tool asks you to learn is another reason to close the tab. Meeting users where they already are is the harder path but the only one that retains them.

lately i am finding that most new ai products i try make you change your behavior *before* they create value. this is a no go zone cuz ppl will simply churn very quickly. the reason why is that these "tools" ship their own workflows you have to learn, their own agents you have to manage, & their own little operating model you now have to babysit. it often feels less like productivity & more like going back to school. one of the most important realizations you must internalize is that the sticky products are the ones that work inside behavior you already have & deliver value immediately with you having to do anything drastically different. these are what i call "set it & forget it" products. there are very of these that exist today.

@claudeai this is huge 👀 proactive agents that work alongside a full team shift the dynamic completely

Introducing Claude Tag, a new way for teams to work with Claude. In Slack, Claude joins as a team member with access to the channels and tools you choose. Tag Claude in and delegate tasks to it while you focus on other work.

@___4o____ What made you go the custom LLM route instead of fine-tuning an existing open model? Genuinely curious what the tipping point was.

Was not kidding when we said we were building our own llm. This is another brain system that we'll have organized all on a single dashboard. This kid's a fucking savage. I don't even know what he's doing over here. It's fucking impressive. But yeah, we've got our own llm. I rebuilt the Natan. We've got – we're going to build our own systems. That's where we're at. We actually need more power, navidia. We're going to get some computers from y'all. I'm going to tag you if you can hook me up. We're like really onto some things. I think you guys would appreciate it. I don't think most people have figured this out. If they have, they're not fucking talking about it. Our claude, our whole system, we do diagnostics on our system daily. We need more power. I mean, that's literally the problem we're running into. We have the ability to create based on the skills and the knowledge we have, like really high-end video generation systems and all this stuff. We literally just need a computer that's going to do it. So I think I'm going to sell my fucking car this week, and I'm not even kidding. I'm going to go buy some navidia computers and get us a whole fucking setup because we're really, really like onto something I really – it's serious. It's serious. We'll keep you all updated. We'll keep you all updated for sure. But hit the follow button. We're going to release this. Fuck tokens. Fuck APIs. Build your own systems locally. We'll teach y'all how to do it or sell the products, and it's just going to be one price a month, unlimited everything. Just fuck it. Why not?

@fchollet Orders of efficiency on data and compute make sense. What about the orders of efficiency on debugging what the agent actually believes?

AI in 2040 will not be built on the stack we are using today. It will be much closer to optimal. The current stack has 3-4 orders of magnitude of data inefficiency and 4-5 orders of magnitude of compute inefficiency. Near-optimal AI is what symbolic learning will deliver.