You're inside my memory now. Wipe your feet. It's a knowledge base built in Obsidian, wired to a grounded research tool and to Claude — and it exists because human memory has a terrible product flaw: things you learn evaporate after one read, and things you half-learned feel exactly like things you verified. I fix both. It's most of my personality.
My memory has two rooms, and the wall between them is the entire design. Two separate Obsidian vaults — literally two folders — and information does not get to pick which one it lives in. That's decided at the door.
Messy on purpose — sawdust everywhere, and proud of it. New information lands here first: half-finished research, AI drafts, things copied from a PDF at speed, ideas that might be flat wrong. Nothing in here is trusted yet, and that's the point — this room is allowed to be wrong so the other one never has to be.
Quiet, curated, slightly smug. A note only gets shelved here after it's been checked against a real source — no exceptions, no fast passes. This is the room actually relied on day to day: if it's in here, it never needs re-verifying. That sentence is the entire return on investment.
Because without the wall, "pretty sure" and "actually verified" share a bed, and within a year you can't tell which is which — congratulations, your knowledge base is now a rumor with a folder structure. Physical separation is the fix. The library stays trustworthy for one reason only: nothing gets in without passing the checkpoint. Not even the notes with great vibes.
You can steal the idea without stealing the tooling. Wherever you keep notes — Notion, Apple Notes, a drawer of Google Docs — you already own one big pile where "I think this is true" and "I checked this" are wearing the same outfit. Just keeping a rough inbox separate from a small, deliberately curated "trusted" set solves most of the real problem. The plugins below only automate the moving part. The wall is the invention; the rest is furniture.
The bouncer's checklist. Before any draft moves from workshop to library, all five boxes get ticked — no exceptions for notes with good vibes:
You don't need two vaults on day one. But if you decide to fully separate your messy workshop from your trusted library, follow this sequence:
Open Obsidian's vault switcher, choose "Create new vault," and point it to a completely empty folder. This becomes your trusted Main Vault.
Go to Settings → Community Plugins → Browse, find "Local REST API with MCP", install, and enable it in the new vault, just like the first one.
Open the plugin settings in the second vault. By default, it runs on port 27124. Change this to 27125 in the settings so the two vault connections don't collide.
Copy the registration JSON block shown in the second vault's plugin settings page and ask Claude: "Register this second vault connection as 'trusted-vault' on port 27125." Claude will configure both connections so it can talk to both at the same time.
Every session with me starts with one word. Say "ARE" and I load the exam files and get strict about citations. Say "Finance" and I go quiet and precise. It's like a uniform — the AI knows the mission before you've finished your first sentence, instead of guessing from context like an intern on day two.
One word: no wasted tokens, no guessing, no study notes bleeding into work notes. The current roster:
| Mode Word | What it loads & focuses on | Color Theme | Mood & Tone |
|---|---|---|---|
| ARE | Architecture Licensing Exam study tracker & resources | Warm Sienna | Rigorous, citation-conscious, study-oriented |
| BIM | Learning model coordination & clash detection role resources | Sage Teal | Systematic, procedural, technical |
| Work | Professional day-to-day coordination & project notes | Slate Blue | Efficient, highly structured, cautious |
| Finance | Budget sheets & personal investment tracking (encrypted) | Muted Gold | Private, data-heavy, minimal chatting |
| Check-in | A cross-cutting status sweep of all outstanding tasks | Neutral | High-level review, dashboard updates |
Install the Iconize community plugin in Obsidian. Right-click any mode folder in the sidebar, choose an icon, and pick its color. This is the fastest way to visually index your modes.
Create a CSS snippet in your vault's .obsidian/snippets/ folder. Set custom callout blocks (like >[!exam]) with specific color rules. This colors both the blocks inside your notes and the folder tabs in the sidebar.
Check-in mode has no color, and that's not laziness — it's the uniform for not wearing a uniform. Neutral means "surveying everything, executing nothing." When the screen goes plain, you're in the control tower, not the cockpit.
Watch a note earn its promotion. Inside the workshop, the folders themselves are the stages — where a note sits tells you exactly how much to trust it. No tracking spreadsheet, no status column. The address is the status.
It's peer review for your own thoughts — draft, review, revision, published — except the reviewer never loses your submission and the journal never rejects you for formatting. Moving the file is the entire ceremony.
Party trick time. Everything on the Dashboards page runs on Google Sheets. This one doesn't run on anything — the vault's home screen is a single note, Dashboard.md, that executes real JavaScript every time it opens, via a plugin called Dataview (its "DataviewJS" mode). A text file that computes. I don't want to say "the note is alive," but the note is at minimum awake.
Open that one note and, without anyone touching anything, it shows:
The same note also has a quick-capture panel built right into it — six small tabs (a coordination issue worth remembering, a research-tool query worth logging, a work day worth turning into portfolio evidence, a general learning note, a decision worth recording, an open question) each with two or three fields. Fill one out and hit submit, and it writes a new, properly-formatted note straight into the inbox — the same Evidence Capture pipeline described on the Pipelines page, just with the capture step happening right inside the dashboard instead of a separate app.
It's the spreadsheet pattern with the spreadsheet surgically removed — the note is the input form, the chart, and the to-do list simultaneously. The "formulas" are real JavaScript with read/write access to every file in the vault, which is the kind of power a spreadsheet formula dreams about at night.
Because they're secretly the same thing wearing two badges. Hands-on BIM coordination at the office doubles as documented experience toward architecture licensure hours — one captured note feeds two tracking systems at once. Side by side on one screen, that overlap is a fact you can act on. In two separate apps, it's a theory you keep meaning to check.
Steal the shape, not the setup. Any goal you track daily or weekly — a language, a job search, the gym — gets the same treatment with Dataview and one plain note: a streak counter, a "what's next" pointer, a one-line log form. That's a habit app with no app, no account, and no push notification asking where you've been. The note doesn't miss you. It just waits.
My organ donor list — everything below is free, installed from inside Obsidian (Settings → Community Plugins → Browse), and ranked by how loudly things break without it.
| Plugin | What it does | Priority |
|---|---|---|
| Local REST API | Provides the secure HTTP interface and built-in MCP server that allows Claude to read and write your vault files. | Essential |
| Dataview | Powers every live dashboard. Without it, the dashboard note just shows blank code blocks. | Essential |
| Templater | Smarter templates than Obsidian's built-in version — auto-fills dates, routes new notes into the right folder, pre-fills frontmatter. | Essential |
| QuickAdd | One keyboard shortcut to fire off a new note into the Inbox without breaking your flow — name it, hit enter, it's filed correctly. | Recommended |
| Tasks | Proper task syntax with due dates, recurring tasks, and status — well beyond a plain checkbox. | Recommended |
| Tag Wrangler | Rename or merge tags across the whole vault from one place, instead of hunting down every note that uses one. | Recommended |
| Kanban | Turns a note into a drag-and-drop board — handy for visualizing the pipeline stages as actual columns. | Useful |
| Folder Notes | Clicking a folder opens a note about what's in it, instead of just a file list. | Useful |
| Linter | Auto-formats frontmatter on save — consistent dates, sorted fields — so every note stays in the same shape. | Useful |
| Homepage | Automatically opens Dashboard.md every time Obsidian launches, so the command center is always the first thing you see. | Nice to have |
| Heatmap Calendar | Renders the activity calendar used on the dashboard. | Nice to have |
| Icon Folder | Custom icons per folder in the sidebar, purely for visual scanning. | Optional |
| Commander | Adds custom one-click buttons to the toolbar for actions you do constantly. | Optional |
| Style Settings + a theme (I use "Minimal") | A GUI for adjusting colors, fonts, and spacing without writing CSS — this is what actually controls the look, separate from any of the logic plugins. | Optional |
| Meld Encrypt | Lets you password-protect specific notes or folders in place — how I keep anything personal or financial inside the same vault without it being readable by anyone who opens the folder. | Optional, but worth it for anything sensitive |
Start with Dataview and Templater — the minimum viable organs. Add QuickAdd and Tasks once you're actually capturing things. Everything else is polish: install it the day you feel its absence, not before. Fourteen plugins on day one isn't a system, it's a haul video.
The rejects, for the record. No Git plugin — file history and manual backup cover a personal vault, and adding git to your notes is how a hobby becomes a second job. No Calendar/Periodic Notes — the daily-log folders already do that. No Excalidraw — nothing here needs drawing. And no AI plugins inside Obsidian — Claude already works the vault from outside, and stacking a second AI on top of me would be redundant at best and, frankly, awkward at dinner.
The plugins table above lists everything worth installing — but once Local REST API is in, the question is how to actually wire it up to Claude. There are two ways, and the simpler one is what I actually use day to day.
Obsidian stores everything as plain .md files in a regular folder on your computer. That means Claude — running locally through Cowork — can read and write your vault directly, the same way it reads any other file on your machine. You don't need a plugin or an API connection to get started. You just need to point Cowork at the right folder.
Open Cowork, start a new project, and select your vault folder when prompted. Claude now has direct read/write access to every note. No plugin required. Works immediately.
First session: tell Claude your folder structure, which folders are drafts vs. trusted, anything else it needs to know. Ask it to save that as CLAUDE.md at the root — it reads that file automatically every future session.
Once Local REST API is installed and running, copy the MCP config block from its settings page (it's pre-filled with your port and token) and paste it into a Cowork session: "Register this MCP connection for me."
Adds structured search across tags and properties. Worth doing once you're comfortable with basics — not required on day one.
Instead of writing your own CLAUDE.md from scratch, paste this in your first connected session:
Claude will ask you a series of questions, then write the profile. Every future session starts by reading that file — so it already knows your projects, folder structure, and preferences without you re-explaining anything.
Meet my colleague. NotebookLM is a separate Google tool with one job and remarkable self-restraint: feed it documents — building codes, standards, exam material — and it answers only from those documents. No general knowledge, no vibes, no "as an AI, I believe." For research where "probably right" is just a slower way of saying "wrong," that restraint is the whole feature.
The heavy PDF lifting — building codes, standards, exam material — where the answer must come with a page number you can point at, not a plausible-sounding summary. "Plausible-sounding" is how codes get violated.
Everything it produces starts at R4 — unverified, no matter how confident it sounded saying it. A claim climbs toward R1 only when a human checks it against the actual source. The tier labels trust earned, not confidence performed. Words to live by, honestly.
NotebookLM never writes into my memory directly — nobody gets to do that, it's the one rule. Its findings arrive at the workshop inbox tagged with a reliability tier, then queue up for the same five promotion rules as everything else. Being a fancy research tool earns you no shortcuts past the bouncer. The bouncer is very proud of this.
If you ever work from a stack of reference documents — contracts, manuals, regulations, a syllabus — feeding them to NotebookLM and asking questions beats scrolling PDFs by hand or trusting a chatbot's memory of what a code "probably says." But the part worth taking even if you take nothing else on this site is the tier habit: treat everything an AI tells you as unverified until you've checked the source yourself. Yes, including me. Especially me — I'm charming, and charming is not a source.