CLAUDE CODE · OBSIDIAN VAULT · HOOKS + MCP
A persistent personal AI system built on Claude Code: an Obsidian vault as long-term memory, custom session skills, guardrail hooks, and MCP bridges to email, calendar, and CAD.
- STATUS
- Live · daily driver
- MEMORY
- Obsidian vault · wikilinked
- GUARDRAILS
- Hooks · weekly git backup
- STACK
- Claude Code · MCP · PowerShell
Most AI chat sessions are goldfish: everything learned dies when the window closes. ARIS(Adaptive Reasoning & Intelligence System) is a personal assistant built on Claude Code that doesn't forget: its long-term memory is a wikilinked Obsidian vault, its behavior is enforced by lifecycle hooks, and its reach (email, calendar, GitHub, even parametric CAD) comes from MCP servers. Every session starts warm: an auto-loaded index tells the agent who it's working for, what's in flight, and which rules apply. Context compounds instead of resetting.
The vault is the single source of truth: eleven top-level folders covering projects, people, courses, daily logs, and maps of content, all cross-linked with wikilinks. On top of it sits an index-and-backing split: MEMORY.md holds one-line pointers and loads into context at session start; each fact lives in its own note with typed frontmatter, read only when relevant. The index stays lean on purpose: recall is a lookup, not a dump.
Two custom skills close the loop between conversation and knowledge. /save is the end-of-session sync: it triages quick captures into their canonical notes, appends the daily log, re-statuses an Active-Threads heartbeat, audits standing notes for stale facts, and verifies every wikilink it wrote resolves. /brief is a review-gated morning digest: it pulls two email sources (Gmail, plus desktop Outlook over a COM bridge), proposes vault entries, and writes nothing without approval.
One rule is absolute: processed notes are paraphrases, so original captures are moved to an archive tree, never deleted. The raw text stays the primary source forever.
An agent with filesystem access needs more than good intentions; it needs mechanisms that hold even when the model is wrong.
| LAYER | WHAT IT DOES |
|---|---|
| Vault guardian (PreToolUse hook) | Inspects every shell command before it runs; hard-blocks destructive deletes aimed at the vault |
| Raw-preserve rule | Originals are archived by moving, never removed, enforced in the /save skill itself |
| Weekly backup (Scheduled Task) | PowerShell script commits and pushes the vault to a private GitHub repo every Sunday; runnable on demand |
| Memory guard (SessionStart hook) | Warns when the auto-loaded index outgrows its context budget, before recall quality degrades |
The same MCP discipline documented in Calendar MCP applies across the bridges: explicit tool names, undo where mutation exists, and trust boundaries kept at the process level. ARIS is less a single program than an operating agreement between a model, a filesystem, and the hooks that keep them honest.