Satyakam Singhal

I benchmarked 10 LLMs including the Chinese models nobody talks about. The results will surprise you. 🧵

I got tired of coding agents that forget everything between sessions. So I shipped recall-mcp: a persistent memory MCP server that turns every Claude Code session into a living knowledge graph. Session ends → distill entities & decisions → Postgres + pgvector → next session starts warm. Hybrid search (semantic + keyword + RRF). Multi-user shared brain. Admin curation. Cloud Run. It's part of a broader stack: structural code graphs (sqry), agent observability (sysmon/Langfuse), context compression (headroom), and a hooks dashboard for Claude Code. Thesis: smarter models help — but durable context is what makes agents compound. Building this in the open. If memory for agents is a problem you're feeling too, let's talk. #AI #MCP #ClaudeCode #Engineering #BuildInPublic

I am trying to build a dashboard to make comprehension better while working with claude code, any thoughts? @claudeai #codex #claude

Andrej Karpathy built a wiki to think with AI. I built something that thinks back. Claude Code + Obsidian equals an AI that actually knows you. Your goals. Your context. Your history. Not a chatbot you explain yourself to every session. A second brain that remembers everything permanently. Build it once. It compounds forever. Bookmark it for later and start this weekend.

Try clawpatch.ai on one of your repos and let codex work its magic. It's amazing at uncovering bugs you didn't know you had.

The /goal Cheat Sheet for Hermes Agent Codex Builds, Claude Code Reviews and Hermes Agent Orchestrates.

Here are 10 GitHub repos that quietly print money while you sleep. 1. Cal. com Open-source Calendly. Fork it, white-label it, sell to dentists and lawyers for $200/month. The founders hit $5M ARR in 3 years doing exactly this. Repo → github.com/calcom/cal.com 2. Plausible Analytics Privacy-first Google Analytics. Self-host it, resell to agencies for $50/month per client. Two founders bootstrapped this to 7 figures. Repo → github.com/plausible/anal… 3. Ghost Open-source Substack with 100% margin. 1,000 readers at $5/month equals $60,000 a year. Forever. Repo → github.com/TryGhost/Ghost 4. n8n Open-source Zapier. Sell automation services for $500-$2,000 per setup. n8n raised $14M because the agency model behind it works. Repo → github.com/n8n-io/n8n 5. Supabase Free Firebase replacement. Build a SaaS in a weekend, charge $29-$99/month. They raised $116M for a reason. Repo → github.com/supabase/supab… 6. Medusa Open-source Shopify. Take 5% on every sale forever. Zero rev share to Shopify. Repo → github.com/medusajs/medusa 7. AppFlowy Open-source Notion. Sell self-hosted to enterprises worried about data privacy. They raised $30M because this market is massive. Repo → github.com/AppFlowy-IO/Ap… 8. Coolify Open-source Vercel and Heroku. Charge developers $20/month to manage their deployments. Replace their $200 Vercel bill. Repo → github.com/coollabsio/coo… 9. Listmonk Open-source Mailchimp. Send unlimited emails for the cost of an AWS bill. Resell to agencies at 10x markup. Repo → github.com/knadh/listmonk 10. Penpot Open-source Figma. Sell self-hosted design tools to agencies who refuse to upload client files to the cloud. Repo → github.com/penpot/penpot The difference between developers who build features and developers who build businesses is one decision. Pick one of these. Fork it this weekend. Ship it next week. The founders behind these repos already proved the model. Save this. Share it with the developer in your life who deserves to break free. 100% free. 100% open source.

The transformer architecture used for ChatGPT, Gemini, and Claude has defined the last decade of AI. It also introduced a fundamental constraint: compute scales quadratically as context grows. Longer inputs, exponentially higher costs and accuracy that degrades well before the context window limit. SubQ changes that. It's the first LLM that breaks the quadratic scaling constraint delivering longer context, higher accuracy, and lower cost at the same time without tradeoffs. Read more here. subq.ai/introducing-su…

This is kind of awesome! You can send a Google Meet invite to your agent (OpenClaw, Claude, whatever) and have them join you on a real-time video call. Just think about that for a second: You're talking "face to face" to your agent. I think Pika found gold here.

I gathered ALL claude code's system prompts/prompts for agents in a repository: github.com/Leonxlnx/claud…

Looks like @OpenClaw has some solid competition 👀 PokeeClaw dropped today, and the specs look pretty wild: → Zero setup → Enterprise-secure → 1,000+ app integrations 🔥🔥🔥 Will test + report back this week on how it handles real-world workflows 👊

AutoResearch only works when you can measure the result with a number but what about writing, arguments, marketing copy? theres no score for "is this convincing" SHL0MS built AutoReason to solve this instead of a metric, it uses a loop of agents arguing with each other: > one writes a draft > another critiques it (no fixes, just problems) > a third rewrites it based on the critique > a fourth merges the best parts of both > a blind judge panel picks the winner > loop until nothing beats the current version every agent gets fresh context so no confirmation bias builds up in testing, autoreason scored 35/35 on a blind panel. the next best method scored 21 same idea as autoresearch but instead of optimizing a number, its optimizing through debate

@gkisokay It's like a rescue agent right? How did you create this diagram btw, the tool?

The 'Super Self-Improving Multi-Agent Framework' for your Hermes or OpenClaw agent. The article below outlines how Hermes can watch over your Openclaw agent, so this example uses Hermes as the main agent. This idea is interchangeable based on preference. How it works: Agent 1 (Hermes) - The main agent: who operates your workflows, and who you have in the top hierarchy of agents. For simplicity, it also runs the workflows using its persistent memory and self-improving skills. Agent 2 (Hermes) - Main's Supervisor agent: who monitors the entire system, reading operations and failure logs, searches for bugs, solutions, then brings them to... Agent 3 (OpenClaw) - External Supervisor agent: who also monitors the Hermes system, and searches for issues and solutions. Multiple times per day, Agent 2 audits the system, finds errors, stale jobs, etc, proposes a fix, then tags Agent 3. Agent 3 audits the system, reads Agent 2's summary, and verifies the problem and solution. Together, they go back and forth in a dedicated channel until they find a coded solution to the problem. If it's a low-risk fix, it auto-fixes it. If it's high-risk, it's raised to Agent 1 for approval via the owner (me). The first goal is never to have to worry that the system is operating correctly. The second goal is to have two separate agents focused on recommending improvements to the system. Ideally, there are no bugs to fix, and you can have them focus on making meaningful improvements to your workflows over time. This is the basic workflow for monitoring, but you can extrapolate this framework to many more use cases. What's [redacted]? I'm cooking up something to continue working toward my goal of building more sentience in my agents. I will share more soon. Let me know in the comments if this works for you. I'll create a fleshed-out article if you are interested.

WorkLens / The Context Gap No one actually owns "work process visibility" for knowledge workers. We're building WorkLens to fix this. It’s not another Jira board. It’s a passive context intelligence platform. It tracks your workflow timeline, maps out the knowledge graph of what you're actually doing, and proactively surfaces context you forgot you had. The gap in the market isn't task tracking—it's contextual memory for the things that don't fit neatly into a ticket. 🔗 en.wikipedia.org/wiki/Context-a… #FutureOfWork #AI #SaaS

Claude Code Context Bleed Dumping all your AI agent plans into a global ~/.claude folder is a massive anti-pattern. I keep seeing devs wondering why their Claude Code context is bleeding across projects. Your folder structure *is* your context architecture. We just overhauled our setup: strict project-level isolation, wiped global bash hooks that were blocking the event loop, and tied SessionEnd directly to our AWS Memory MCP. Now, every session automatically builds the long-term knowledge graph without polluting the global state. Keep your agents scoped. 🔗 docs.anthropic.com/en/docs/agents… #ClaudeCode #AIAgents #DevTools

Best GitHub repos for Claude code that will 10x your next project in 2026: 1. Claude Mem github.com/thedotmack/cla… 2. UI UX Pro Max github.com/nextlevelbuild… 3. n8n-MCP github.com/czlonkowski/n8… 4. Obsidian Skills github.com/kepano/obsidia… 5. LightRAG github.com/hkuds/lightrag 6. Everything Claude Code github.com/affaan-m/every… 7. Superpowers github.com/obra/superpowe… 8. Awesome Claude Code github.com/hesreallyhim/a… 9. GSD (Get Shit Done) github.com/gsd-build/get-…

So cool!! I'm running @nvidia's NemoClaw with Qwen3.5-27B entirely locally via Telegram. No API costs, no data being sent to anyone NemoClaw is a security sandbox built on top of @openclaw that lets you restrict the files and networks your lobster can access. I'm running the 4-bit GGUF via llama cpp on a DGX Spark. It's running pretty smoothly at 9 tokens per second! The future might be local :)

Building a Blockchain That Survives the Post-Quantum Era Most blockchains today are secure only because quantum computers are not here yet. That’s the uncomfortable truth. Systems like Bitcoin and Ethereum rely on elliptic curve cryptography. The moment large-scale quantum machines become practical, those assumptions collapse. Not slowly completely. People talk about this problem. Very few are actually building systems to handle it end-to-end. This is where LQ1 comes in The problem nobody is solving fully Post-quantum cryptography is not new. Algorithms like Falcon, Dilithium, SPHINCS+ they exist. But plugging a new signature into a blockchain is not a solution.

Most of ai twitter pay $200/month for Claude. In the coming months that probably won't need to. google just open-sourced an algorithm called TurboQuant. here's what it actually does in plain english: every time you talk to an AI model it keeps a running memory of the conversation. the longer you talk, the more memory it eats. eventually it slows down, gets dumber, and falls apart. TurboQuant compresses that memory by 6x. makes the model run 8x faster. zero quality loss. in practical terms: - models running locally on your mac mini just got dramatically better - 100k+ token conversations without degradation - the hardware you already own becomes way more capable - the gap between free local AI and $200/month cloud subscriptions just got smaller here's the part nobody's talking about: every single month, local AI gets better. open-source models get smarter. compression techniques like this keep dropping. hardware keeps getting cheaper. 12 months ago running a real AI model locally was a novelty. now it's genuinely useful. 12 months from now it might be the default. google published the full research. no paywall. no API key. no subscription. anyone can use it. the companies building for local-first AI right now are going to look very smart very soon.

@shedntcare_ Vault

Your Obsidian vault could be a $120K/year asset… I’m giving away the setup 👇 Turn Obsidian + Claude Code into a 24/7 personal OS That thinks, organizes, and builds with you. No expensive tools. No complex stack. Just a simple system that compounds. I’ll send it to 20 people for free. Rt + Like + Reply “VAULT” 👇