Give Your Company a Memory: My Claude + Obsidian Vault

My company has one employee and a very good memory. Every decision, client note, legal document, invoice automation and half-finished idea lives in a single Obsidian vault that Claude reads and writes as naturally as I do. After a year of daily use it has become the most valuable system I run — more than any single product. Here's how it works and what I'd copy first if I were starting again.

Credit where due: the skeleton is Claudesidian, Noah Brier's open-source Claude Code + Obsidian starter, itself inspired by Dan Shipper's “Claude Code as a second brain” interviews — plus the broader idea, which I first absorbed from Andrej Karpathy's writing on LLMs as a new kind of computer, that files are the natural memory for a model. The starter gives you folders; everything below is what turned it into an actual company memory.

The architecture: PARA plus an immutable inbox for reality

The vault uses Tiago Forte's PARA method — numbered folders for Projects (time-bound, with a completion criterion), Areas (ongoing responsibilities), Resources, and Archive, plus an Inbox that must stay under 20 items. The one structural addition I'd defend anywhere: a raw/ folder for immutable source data — meeting recordings, calendar exports, emails, scraped articles. Never edited, only referenced. The vault stores full source text rather than summaries, because the model can search saved text at will; a summary is a decision about what mattered, made too early.

Range matters more than tidiness. My Projects folder holds tax planning next to product R&D next to client prospection next to furniture shipping. One memory, whole company — that's what makes the AI useful in the gaps between domains, which is where a founder actually lives.

The three ideas that make it AI-legible

1. Context files are an in-vault system prompt

Three living documents — Me.md (background, working style, preferences), Spectralgo.md (legal facts, services, positioning, stack), and a finance README — carry frontmatter marking them type: context. The vault's CLAUDE.md tells the model to read them first. Instead of re-explaining my company in every conversation, I maintain three files that make any session instantly situated. If you copy one idea from this post, copy this one.

2. CLAUDE.md is a constitution, not a readme

The instruction file at the vault root encodes invariants, not vibes: never move the numbered core folders, use mv not cp so links survive, preserve bidirectional links, always add frontmatter, write in a specific voice (“direct and confident statements… no unnecessary lead-ins”). The vault teaches the AI how to behave inside it — which is exactly what you do when you onboard a human.

3. Git is the persistence layer

Every session starts with git pull and ends with a commit. Memory lives in version-controlled files, so it survives across sessions, across machines, and across the model forgetting everything. When people ask “how do you give an AI long-term memory?”, the boring, working answer is: a repo.

The memory compounds: session logs and a decision register

After meaningful work, Claude appends a structured session log — topic, actions, decisions, follow-ups — to the daily note. Explicit rules say what not to log (quick questions, failed debugging), because an audit trail nobody reads is noise. The daily note itself is templated: mood, one Most Important Thing, and pomodoro counts tracked inline as [🍅:: 3/4].

The clearest proof this is company memory rather than a notes app: the SASU's legal decision register — the sole-shareholder decisions French law requires — lives in the vault as dated markdown files, each paired with its signed PDF. When I ask “why did we structure compensation this way?”, the answer is a versioned file, not a memory of a conversation.

Two brains: the vault thinks, Microsoft 365 executes

The failure mode of every second-brain system is becoming a beautiful graveyard of tasks. The vault's rule: long-form thinking and decisions live in the vault; time-sensitive execution lives in Microsoft 365 (To Do, Calendar, Teams), bridged by a small custom CLI the AI can drive. The docs even name the anti-patterns: “creating dozens of tasks without a verb, context, or due date (becomes task debt)” and “automating messages to clients without review (reputational risk).” Thinking tool and execution tool, connected but not confused.

Small automations that punch above their weight

• Link-repair scripts fix wikilinks after files move between PARA folders — the graph stays intact, so reorganizing is never destructive.
• A skill-discovery hook injects the list of available vault skills into context only when a prompt mentions them — progressive disclosure instead of a bloated system prompt.
• Scrapers (web articles, YouTube transcripts) save full text into raw/, feeding the token-maximalist philosophy: more searchable context beats cleverer prompts.
• An invoice generator lives in the same repo — because the memory that knows your clients and your legal details is the right place to automate the paperwork they generate.

Start here

1. Make a repo. Put PARA folders in it. Open it in Obsidian (or don't — files are enough).
2. Write your two context files: who you are, what your company is. Mark them type: context.
3. Write a CLAUDE.md with five invariants you actually care about.
4. End every session with a commit. Let the AI write the log entry.

The starter kit's README says it best: the bicycle feels wobbly at first, then you forget it was ever hard.