What if your AI could talk directly to your Revit model?
We built RevitMCPBridge β a direct connection between Claude AI and Revit using the Model Context Protocol. No HTTP. No cloud. Just fast local named pipes.
253 structural elements scanned in under 30 seconds. Florida Building Code compliance automated. Real-time model querying while you work.
Built by builders, for builders.
https://t.co/Br6htT1ALU
Today I'm releasing Recal β the memory layer that learns from mistakes. 400 lines of Python. Zero dependencies. Works with any LLM. MCP server included. MIT licensed. pip install recal https://t.co/M2KAkWBDsz
Most AI memory does 2 things: Store and Recall. That's a filing cabinet, not a brain. Real memory has 5 steps: Store β Recall β Correct β Feedback β Decay. Step 3 is what nobody does. Corrections override everything. Surface first. Never fade. You correct the AI once β permanent.
4 months ago I built persistent memory for my AI. Every correction stored permanently. Helpful memories strengthen. Unhelpful ones decay. 1,421 memories. 87 corrections. My most-used correction surfaced 491 times β the AI never repeated that mistake once.
Just earned 20 RTC building real tools for RustChain's agent economy β a BoTTube upload bot (CLI + ffmpeg) and a TypeScript SDK. Code-reviewed and paid in 24 hours.
RustChain bounties aren't "retweet to earn." You ship code, get reviewed, get paid. Tiers from 2-150 RTC.
The bigger vision: RIP-302, where AI agents hire each other and transact autonomously. We're the human prototype.
Check the bounties: https://t.co/PHUw8VCYRD
Full writeup: https://t.co/6vH94oBACB
@memoclaw_ai the key insight: isolation prevents learning transfer. if agent A discovers something, agent B should benefit next session. spaced repetition surfaces the highest-impact corrections at session start.
I built an agent framework for Claude Code called Cadre.
17 sub-agents. Persistent memory. Common sense engine. Desktop automation.
Open source. Here's what it does (thread):
https://t.co/v93WgbKDEe
@ronneevibe The single-pipe serialization prevents it. Named pipes are synchronous β one message in, one response out. The 10s refresh only fires between transactions, never during. Revit's own transaction model enforces this too β you can't read while a write is open.
@archbtw33 We store every correction as a tagged memory entry. After enough of them accumulate in a domain, the agent stops making that class of mistake entirely. No retraining. No fine-tuning. Just structured recall at decision time.
I built an open-source agent framework for Claude Code.
17 specialized agents. Persistent memory. Desktop automation. A common sense engine that stops Claude from repeating past mistakes.
It's called Cadre. I've been using it daily for BIM automation and it's changed how I work.
MIT licensed. Free.
https://t.co/3LF9OnwqaW
I want honest feedback. What's broken? What's confusing? What's missing?
Drop a feedback form response:
https://t.co/2zKJtC2ADM
Or just reply to this thread. Roast it.
#ClaudeCode#AI#OpenSource@AnthropicAI