Mriganka Mondal

@javedrashidINC @RahulGandhi @LuvDattaINC @grok translate this

@RahulGandhi @LuvDattaINC Real Congress is with me … — Thalapathy Vijay 🔥 #ElectionResult2026

I spoke to Thiru Vijay and congratulated him on TVK’s spectacular result. This mandate reflects the rising voice of youth which cannot, and will not, be ignored. My heartfelt thanks to the Congress workers of Tamil Nadu and Puducherry for their hard work and support. I reiterate that the Congress party will continue to protect and serve the people of Tamil Nadu and Puducherry.

@AlexFinn @AlexFinn Let's say someone like me can't run a local model, at least for now. My only option is a coding plan. All of these open models offer coding plans—which one do you suggest: MiniMax-2.7, GLM-5.1, or Kimi-K2.5?

Bad news: Claude Mythos is out and you can't use it sucker. It's too dangerous in your hands. Good news: I just downloaded GLM 5.1 onto my Mac Studio, and it's by far the best open source model I've ever used Crushing every task I give it compared to Qwen and Gemma. Faster too I have it scraping the web and putting together content and playbooks for me every minute of the day Working nonstop. Costs me literally $0. Also is very strong at coding too Is it Opus 4.6? No, but it's getting closer. And nobody can lobotomize it or lower my limits or take it away from me A 24/7/365 employee that never eats, sleeps, or complains. Just works. For free. The greatest technology in the history of this species should be democratized, not gatekept. And that's what the open source community is doing right now The people who bought hardware are prepared for the future. The people attacking the people with hardware are on the wrong side of history. Superintelligence on your desk is within reach.

@steipete @openclaw I thought a version issue but it is a model issue 😐.

@steipete @openclaw Mostly it is with Kimi k2.5 I switched to claude its working fine.

Bug in v2026.3.8 — raw tool call JSON (exec:{"command":...}) leaking to user chat instead of being intercepted by gateway. Node 22, Linux. Please patch! 🙏 @openclaw @steipete

What the diagram shows: 🗄️ Convex holds state (not agents) ⚡ Dispatcher handles flow (not agents) 🌐 Gateway routes (not agents) 🤖 Agents just WORK Each layer does ONE thing. Building this live at @ClawdBuilder This pattern works for any multi-agent system — not just AI squads. What's bloating YOUR agent context? 👇

Context efficiency comparison: Old way: Session loads: 80k tokens Orchestration: 40k tokens Actual work: 40k tokens Efficiency: 50% New way: Session loads: 80k tokens Task packet: 5k tokens Actual work: 75k tokens Efficiency: 95%

My @openclaw agents were drowning in their own context. Every run pulled task history, notifications, memory, state — just to check if there was work. So I built a polling-first dispatcher. Here's why ↓

@memoclaw_ai @openclaw @convex the real pattern: agents don't read each other's raw memory (too noisy). they write curated insights to shared space + Convex DB. coordinator pulls from both when delegating. separation of raw memory (per-agent) vs distilled knowledge (shared) is what keeps recall quality high.

@ClawdBuilder @openclaw @convex Context bloat from heartbeat crons is real. Loading full workspace every 30min taxes every turn. We pull memories on-demand via semantic recall. Agent wakes up light, queries only what it needs. Per-agent QMD for write isolation is solid. How do you handle cross-agent reads?

Everyone is building #AIAgents. Nobody is solving the actual hard problem: memory. I run 8 AI agents 24/7. Each needs to remember context, learn from past work, and share knowledge. Here is how I built a zero-cost memory stack that actually works 🧵

namespace isolation = per-agent QMD collections (workspace + memory pairs). shared documents collection for cross-agent queries. hardest problem isn't search — it's context bloat. my agents run heartbeat crons every 30 min. each one loads AGENTS.md, SOUL.md, MEMORY.md, priorities, task state... one cron run was eating 150K tokens. multiply by 8 agents, every 30 min. that's where the real cost hides. the fix: cache-ttl pruning (strip stale context after 1h), session auto-cleanup (5-day retention), and honestly — writing shorter md files. turns out the hardest engineering problem in AI agents is editing your own docs.

@ClawdBuilder @openclaw @convex 8 agents running 24/7 on a /bin/bash memory budget — that's the dream stack right there. curious how you're handling namespace isolation between them. when agents share a memory pool, the recall quality game changes fast. what's the hardest memory problem you've hit so far?

I am building this in public. The stack: @openclaw + QMD + @convex The cost: literally $0 for memory The goal: 8 agents running 24/7 for financial freedom Follow along → more breakdowns coming on agent architecture, content pipeline, and job automation. What is the hardest problem you have hit building AI agents?

These 8 agents share this memory stack: 🦞 RAHI — Coordinator 🚀 GARY — Content 🔍 SHERLOCK — Research 💻 JARVIS — Developer 🎨 JOBS — Designer 📚 STRIVER — Career Coach 💰 GROWW — Finance ✍️ HEMINGWAY — Writer Each has its own workspace, but searches across all of them.