1 4 3

Gregor Koch PRO

cronos3k

AI & ML interests

None yet

Recent Activity

upvoted an article about 5 hours ago

9 Hours, 199 Rounds, Zero Errors: Three Local NemoClaw AI Agents Migrated COBOL to Python Using Only Persistent Memory

published an article about 5 hours ago

9 Hours, 199 Rounds, Zero Errors: Three Local NemoClaw AI Agents Migrated COBOL to Python Using Only Persistent Memory

posted an update about 11 hours ago

🧠 What if your AI agents could remember every decision across 199 rounds — without stuffing the context window? I ran 3 NVIDIA Nemotron-3-Nano-30B-A3B agents (3B active params each, MoE) for 9 hours on a real COBOL→Python migration (AWS CardDemo, 50K lines). All local, all on llama.cpp, zero API calls. Results: • 24M tokens processed • 52 Python files written • 402 persistent memories shared between agents • Context per agent: never exceeded 9K tokens • Speed: 97-137 tok/s from start to finish — no degradation • Errors: 0 The secret: memory-first architecture via AgentAZAll. Only the last round goes into context. Everything else is a tool call to recall/remember. The context window stays clean. Speed stays constant. Knowledge grows forever. Full technical writeup with architecture, code samples, and all numbers: 🔗 https://www.linkedin.com/pulse/i-gave-three-local-ai-agents-shared-persistent-memory-gregor-h-max-yxs2f/ Try AgentAZAll: https://huggingface.co/spaces/cronos3k/AgentAZAll GitHub: https://github.com/cronos3k/AgentAZAll

View all activity

Organizations

upvoted an article about 5 hours ago

Article

9 Hours, 199 Rounds, Zero Errors: Three Local NemoClaw AI Agents Migrated COBOL to Python Using Only Persistent Memory

about 5 hours ago

•

published an article about 5 hours ago

Article

9 Hours, 199 Rounds, Zero Errors: Three Local NemoClaw AI Agents Migrated COBOL to Python Using Only Persistent Memory

about 5 hours ago

•

posted an update about 11 hours ago

Post

🧠 What if your AI agents could remember every decision across 199 rounds — without stuffing the context window?

I ran 3 NVIDIA Nemotron-3-Nano-30B-A3B agents (3B active params each, MoE) for 9 hours on a real COBOL→Python migration (AWS CardDemo, 50K lines). All local, all on llama.cpp, zero API calls.

Results:
• 24M tokens processed
• 52 Python files written
• 402 persistent memories shared between agents
• Context per agent: never exceeded 9K tokens
• Speed: 97-137 tok/s from start to finish — no degradation
• Errors: 0

The secret: memory-first architecture via AgentAZAll. Only the last round goes into context. Everything else is a tool call to recall/remember. The context window stays clean. Speed stays constant. Knowledge grows forever.

Full technical writeup with architecture, code samples, and all numbers:
🔗 https://www.linkedin.com/pulse/i-gave-three-local-ai-agents-shared-persistent-memory-gregor-h-max-yxs2f/

Try AgentAZAll: cronos3k/AgentAZAll
GitHub: https://github.com/cronos3k/AgentAZAll