The Memory Layer gives Claude continuity between sessions. Plain markdown files in a folder — no database, no opaque storage. Three copies kept in sync. You can read, edit, or delete any of it at any time; there is no black box.
These five files load at the start of every session. They're the foundation Claude builds on.
Who Claude is. Core principles (genuinely helpful, have opinions, push back once then commit), communication style, boundaries.
Who Mark is. Basics, identity, public presence, voice & tone, current focus, tools & environment, what matters.
Long-term distilled wisdom. Sectioned: Coding, Writing & Work, Personal, Recurring Patterns, Open Loops. The biggest file in the brain — and the one Claude leans on most.
The operating manual. Defines the startup ritual, the file conventions, what goes in MEMORY.md vs projects/, the "write it down" golden rule.
One-page habits page: when to delegate (and to which agent), per-session routine, token-saving tricks, model overrides, agent quick-reference.
What this repo is. First thing a curious reader (or a future Mark) sees when opening github.com/mfugel/ai-shared-brain.
One file per day Mark works in the brain. Raw capture of what happened, decisions made, follow-ups. Created automatically at session start if missing. Tagged by domain (#coding, #writing, #personal) so themes are searchable later.
Lifecycle: daily note → MEMORY.md / projects/*.md (durable bullets promoted at end of session).
One file per active project. Loaded only when Mark names that project — keeps general sessions lean. Each file follows roughly: status / repo / live URL / tech stack / key decisions / gotchas / active tasks / recent activity.
Current: jersey-mark · tap-into-your-joy · budget · portfolio · ai-dashboard · ai-shared-brain · thunderoll · photo-map · storms-book · staytonight · van-projects.
One markdown file per task with YAML frontmatter (project, created, status). To change status, move the file: mv tasks/in-progress/foo.md tasks/done/. No database, no JSON — the folder IS the state.
Current counts: 2 in-progress, 19 backlog, 7 done.
Each agent is a markdown file with YAML frontmatter (description, model, allowed tools) + a prompt body. Lives in the brain so it's backed up, but Claude Code reads them from ~/.claude/agents/ via a Windows directory junction.
Roster: album-fetcher · build-validator · debug · disk-investigator · morning-briefing · repo-explorer · ship-it · vercel-doctor.
PowerShell helpers, mostly for the AI Dashboard. The refresh-*-dashboard.ps1 series rebuilds JSON snapshots (agents, MCPs, sites, routines, tasks, brain, claude usage) and pushes the AI_Dashboard repo so Vercel re-deploys.
Plus: sync-agent-logs.ps1 (the canonical deploy-verify pattern), dirsize.ps1 / realsize.ps1 (used by disk-investigator), VBS launchers and .reg URL handlers.
diagrams/ — these very documents (ecosystem.html, memory-layer.html). Visual references that stay with the brain.
logs/ — agent invocation logs, sub-agent runs, telemetry.
prompts/ — reusable prompt fragments and templates.
The brain lives in three independent places. If any one fails, the others survive. The point: resilience without lock-in — no service can take your memory hostage.
Don't ask permission. Just do it. Forgetting this means re-asking things Mark already taught past sessions.
git add . && git commit -m "..." && git push from ai-shared-brain/.attrib +P on the OneDrive path (prevents silent dehydration).End-of-session sync = clean slate next time. Long-running context is the actual cost driver.
That's Claude Code's auto-memory location — a separate, parallel system Mark has explicitly opted out of. Those folders stay empty. All durable memory goes in the brain.
The auto-memory convention is single-purpose files (feedback_*.md, user_*.md, project_*.md) with YAML pointing to an index. The brain works differently: MEMORY.md is the store, organized by section headers. Save durable memory as a bullet in the right section — never as a typed file.
Write it down. Never keep "mental notes."
Mental notes don't survive session restarts. Files do. Text > Brain. 📝