Agent governance is turning into its own category. Runtime policy engines. OWASP risk frameworks. Enterprise-grade interceptors. Big companies are shipping serious infra for this. I govern my agents with pure markdown files. The problem is the same though. Your agent forgets what it decided two sessions ago. Silently drops context. Fixes one thing, breaks two others. Other approach: intercept every action at runtime. Mine: make the agent write down what it did before it leaves. Pure markdown. No runtime, no database, zero dependencies. Over 9 versions the framework found 13 bugs in its own operating procedures. Diagnosed and patched each one by itself. One version the SOP hit 565 lines from all the patches. My agent started failing basic tasks because its own rulebook was too long. Had to cut 78% of it. Got more reliable immediately. Another version the agent wrote memory to the wrong directory for 4 days. No errors. Just silently saving to a path nobody reads. The part I really like is the framework keeps getting smarter about me. It tracks how I make decisions, not just what I decided. Different context depending on whether I'm doing strategy work or debugging code. Feels less like a tool and more like a second brain that actually remembers. More versions in. Still just markdown files.

Someone opened an issue asking for OpenCode support. Looked at it. OpenCode uses .agents/ (plural) instead of .agent/ (singular). That's it. Half a day later: install.sh, convert-skills.sh, docs, QA all done. v3.0.3 tagged. This is the part people don't talk about with "pure markdown, zero dependencies" It's not just about being lightweight. It's that adding a new tool literally means changing a directory path. No API adapters. No SDK updates. No CI pipeline changes. 7 tools supported now. The 8th won't be any different.

I spent a day inside Claude Code's leaked 512K lines of source code. Not for the architecture. Not for the agents. For 7 plain-text rules buried inside the TypeScript. Rules that solve problems every AI power-user hits: 1. Your AI remembers stale info from 2 weeks ago 👉 Tag every memory with a date. Flag anything >7 days. 2. Your memory file bloats to 500 lines of garbage 👉 Never save what you can re-derive. If you can grep it, don't store it. 3. You delegate to sub-agents but they have zero context 👉 Iron law: understand it yourself first, then write a self-contained spec. Never forward raw chat history. 4. "Last Thursday", which one? 👉 Always write absolute dates. 2026-04-01, not "Thursday." 5. Your AI only remembers mistakes, never what worked 👉 Record both corrections AND confirmations. 6. All memories dumped in one giant file 👉 4 categories: user prefs / feedback / project / reference. 7. End-of-session summaries are useless walls of text → Structured format: Rule → Why → How to apply. 💫 These aren't code. They're just rules that you can just copy them into your CLAUDE.md/AGENT.md We did, it's called MUSE v3.0