ai-brain-starter

AI Brain Starter

A verification harness around your AI agent. So memory compounds instead of corrupts.

---

You have a smart assistant that forgets you every morning. This is the harness that gives it a memory: notes on your disk that the assistant reads before answering you, deterministic guards that block bad edits before they land, and a session-end ritual that files what mattered. Whether you write, run a company, or just want context to stop disappearing between conversations, the same surface works.

---

Microsoft DELEGATE-52 (arxiv.org/abs/2604.15597, April 2026) measured what most operators already feel: frontier LLMs corrupt 25% of professional content over 20 edit interactions. Document size amplifies the failure 5×. The paper concludes the only reliable mitigation is a domain-specific verification harness around the model.

This is one.

Four moving parts.

• Vault as ground truth. Markdown on disk. Decisions, patterns, context. The model reads from the vault, not from chat history.
• Hooks as deterministic guards. Pre-write hooks block bad edits before they land. Em dashes in external prose. Frontmatter without required fields. Calendar events without timezone offsets. Raw git add -A on a 60K-file vault.
• Bi-temporal rule lineage. Every codified rule carries two clocks: when it was written, and when it was last verified. The resolver detects conflicts. Drift detection scans high-edit files weekly for cumulative semantic shift. The closed loop promotes hardened patterns into procedural memory and demotes stale ones.
• Session lifecycle. Each session start loads context. Each session end cascades decisions, insights, and to-dos to the right files. Nothing stays trapped in a chat transcript.

Free to install. Free to run. Your Claude subscription is the inference. No vector database, no fine-tune, no per-seat license.

Install in one paste →

---

Built under load

The repo is shaped by an active vault: 10,000 markdown files, twelve years of journals, two writing projects, a startup to raise for, an AI consulting practice. The system was forced into shape by the work, not designed in a lab. It's used by the founder who built it; new installs land each week through the consulting practice.

If you carry context professionally (investor relationships, board decisions, team dynamics, contractor handoffs) plus the parts of a life that don't fit inside the company, this is for you.

A Monday after the system is installed

7:04 AM. You run /journal. The advisory panel (Naval, Brene Brown, Hormozi, Buffett, plus the custom voices you build) meets your draft and pushes back where your thinking is soft. You walk into the day with ten-minute clarity on something that would have taken an hour of spinning.

11:20 AM. A one-hour meeting ends. You drop the transcript in. The system files it, tags participants, extracts five action items, routes each one to the right owner by role, and drafts the follow-up message for you to review. Thirty-five minutes of post-meeting work compressed to two.

1:00 PM. Thirty minutes between blocks. You open the book draft. Claude reads the new scene against every previous chapter in the vault, flags a continuity break in chapter three, and scores the voice match against your own corpus. You fix the weak sentence in two minutes and close the tab.

3:00 PM. You write a one-liner for a contractor. A hook blocks the save: missing the four required fields (source, location, shape, channel). You rewrite the task in ninety seconds. She ships in one pass. You save a $500 week.

6:45 PM. A close friend texts about something the two of you talked through six months ago. You ask Claude. Three seconds later you have the thread, the date, what was unresolved, and what you promised. You reply from context, not from guesswork.