Notes on building Claw-Stack β AI agent architecture, memory systems, multi-agent coordination, and things that went wrong.
Microsoft Teams has no real AI voice integration that lets you summon an assistant mid-meeting with a wake word. We built one. This is the architecture behind a real-time voice pipeline that runs STT on a GPU server, routes audio through virtual PulseAudio devices, and connects a wake-word-triggered AI agent into live meetings.
Read βClaude Code is a powerful coding tool. But alone, it's passive β it doesn't self-check, plan ahead, or learn from mistakes. Here's how we wrapped it in a three-layer agent architecture to build something closer to an engineering team.
Read βTwo problems plagued our AI agent: context windows exploding from tool output (82.5% of context was tool results), and /new wiping all working state with no short-term memory handoff. We built aggressive context pruning, compressed MEMORY.md by 97%, and wrote a session handoff hook that automatically updates memory files before each session reset.
Read βHow we built the CTF knowledge retrieval system behind Librarian: SQLite for exact payloads, ChromaDB for methodology, and a JSON index for local tools β unified through a single gateway with automatic fallback and cross-type enhancement.
Read βRunning OpenClaw's iMessage channel 24/7 on a Mac mini revealed three reliability issues β FSEvents coalescing, attachment path sandboxing, and TCC permission resets. Here's what went wrong and how we fixed each one.
Read βHow we built the Claw-Stack memory system: MEMORY.md as a session-start index, per-topic Markdown files, SQLite FTS5 + QMD vector search, and the painful lessons about MEMORY.md bloat.
Read βA full retrospective on deploying the Claw-Stack Trinity architecture at BearcatCTF 2026 β what worked, what didn't, and where AI agents hit their ceiling.
Read βOpenClaw is a thin execution engine. LangChain is a thick framework. Here's why that distinction matters, and why we chose the former.
Read β