I open-sourced project-brain v2.3.1: installable as a Claude Code skill. Works with Cursor, Copilot, Codex, Aider too. Two production projects validated.
Architecture: brain/ folder + collaboration protocol that lets AI tools resume context across sessions.
DMs open for your observations.
https://t.co/uLjYn6veVJ
For days, I nearly said goodbye to an account I'd had for four years.
Here's what happened: OpenAI suddenly required me to verify a phone number I lost years ago just to get into my own account — even though I could prove it was mine with email, an authenticator app, and a passkey. The code could never reach that dead number. I was locked out.
I wrote to support, over and over. They answered twice — and twice was enough to break me: "impossible," "we can't," "it's policy." I'd already started practicing the goodbye.
Then the door opened.
Not because they changed their minds — but because somewhere I couldn't see, a crowd of strangers kept reporting that same wall, kept making noise, refused to compromise — until OpenAI loosened it.
My account came back on the light they made.
Thank you — every one who didn't swallow it, who didn't accept "impossible." You'll never know your stubbornness saved four years of a stranger's memories.
One day, when someone else is stuck in the dark — I want to be that light.
I open-sourced project-brain v2.3.1: installable as a Claude Code skill. Works with Cursor, Copilot, Codex, Aider too. Two production projects validated.
Architecture: brain/ folder + collaboration protocol that lets AI tools resume context across sessions.
DMs open for your observations.
https://t.co/uLjYn6veVJ
Compared 4 local-first memory systems (Naive bge-m3,
LangChain Memory, Mem0, my own Hybrid+Reranker setup)
on 200 anonymized chunks of my AI memory.
3 findings:
1. The 3 "dense-only" approaches scored identically
(54.2% Recall@5). SDK ≠ precision.
2. Mem0 + rerank scored *worse* than Mem0 raw —
no candidate prefetch in its pipeline.
3. Hybrid + prefetched reranker hit 62.5% (+8pts)
but at 54× latency.
n=200, self-evaluated, not a real benchmark.
Engineering notes from a Saturday.
This stuck with me. Spent 8 months building on the idea — designer, don't code, used context-engineering as my main lever to ship 5 AI projects with my AI co-builder.
Open-sourced the system today: a brain/ folder split by how fast each file changes (heartbeat → per-decision → per-session). The constraint forces the model to retrieve, not accumulate.
Writeup: https://t.co/ruBooPCmYw
Matt — your skills repo nudged me to ship mine. Designer who doesn't code; been running a project-brain skill for 8 months across 5 AI infra projects.
Just open-sourced it. Same "small, composable, hackable" spirit you championed — but focused on session-to-session continuity instead of in-session workflow.
Long writeup with the design decisions: https://t.co/ruBooPCmYw
读完这条受启发。
我也是被同样的痛点逼出来造工具的——产品设计师,不写代码,8 个月和 AI 副驾一起出了 5 个 AI 项目。痛点不是 AI 不会写代码,是新会话每次都"失忆"。
刚开源了 project-brain:一个 brain/ 文件夹结构,让新会话 30 秒内接续上一次的进度。Slogan: Different session. Same brain.
刚写了一篇英文版的设计思路,主题接近你这条但角度是 maker 视角:
https://t.co/ruBooPCmYw