THE LAW THAT BENDS EVERYTHING TOWARD THE CENTER
The Method
Fifteen bodies orbit a vault. This is the law that makes them fall: how the system turns raw signal into memory, and how the agent works inside it.
THE ENGINE · ONE CLOSED LOOP
The vault is not a place to store notes. It is an engine with five stations on one orbit: signal is captured, ingested into the graph, distilled into memory, recalled before the agent answers, and handed back as documents that fall back in as memory. Nothing leaves the loop.
01
Capture
A daily note, a clipped link, a line from a meeting. Nothing is formatted or filed yet; the only job is to get raw signal into the well. It lands as plain markdown you can read with any editor.
02
Ingest
The agent reads what fell in, tags it, and links it to what is already there: a person to their company, a decision to the project it moved. A loose note becomes a connected one, found later by its neighbors, not only by its name.
03
Memory
What matters is distilled into durable files: rules the agent should follow, decisions with their trade-offs, learnings from a single session. This is the core of the well: plain files under _system/, versioned in git, owned by you.
04
Recall
Before it answers anything, the agent reads memory: the vault context, your preferences, the active rules, the recent sessions. It walks the graph outward from what you asked. The answer starts from what you already know, not from a blank slate.
05
Documents
What comes back out is a document you can trust: a PRD, a decision record, a brief grounded in your own prior thinking. Every document you keep falls back in as new memory.
↺ WHAT’S LEARNED GOES BACK INTO MEMORY. THE LOOP CLOSES.
THE AGENT IN THE WELL · HOW EXO WORKS
Exo is an agent that lives in your Obsidian sidebar (a real chat surface, not a terminal), powered by the Claude or Codex CLI you already have installed. Your vault is its working directory. Any agent can write files. Exo’s difference iswhere it looks first.
1 · RECALL
A prompt lands on Exo. Before answering, it reads the memory core: vault context, your preferences, the rules in force, the recent sessions. It never starts from a blank slate.
2 · READ
It walks the vault graph outward from your question: backlinks, the neighborhood, the notes that link to the notes you named. A question about a company pulls in the person who works there and the decision made about them.
3 · WRITE-BACK
The answer comes back as something you can keep: inserted into the note, or captured as a decision or learning through gated tools. The trusted output becomes memory.
Because the agent runs on the CLI you own, Exo adds no network surface of its own: your prompt and its context go to Anthropic or OpenAI under your own login, and nowhere else. Sensitive actions (edit, write, unlisted shell) stop forAllow once / Always / Deny, and every turn shows exactly which notes it touched, with a diff and a two-step revert.
DETECTED CAPABILITIES
What the agent does inside the vault, grouped by what it lets you do. Everything here ships today.
Memory & recall
- Boots every conversation with vault context, your preferences, the active rules, and recent sessions
- Reads memory before it answers: the graph, the rules, the record
- Writes back through gated tools: capture a decision, log a session, capture a learning
- Warm sessions across turns, with a live context-window meter
Graph-aware tools
- An in-process toolset that knows the vault: search, read, backlinks, neighborhood, tags
- Search rides the Omnisearch index (BM25 + fuzzy) when present, a built-in scorer otherwise
- Creates tag- and frontmatter-aware notes, appends, updates frontmatter, adds links
- Every touched note is a link back into the graph
Control & trust
- Sensitive actions (edit, write, unlisted shell) prompt Allow once / Always / Deny
- Tool calls surface as cards: a diff preview for edits, command + output for shell
- A touched-notes footer after each turn, with per-note diff and two-step revert
- Exo adds no network surface of its own; it drives the CLI you already trust
Knowledge-work native
- Insert a reply straight into the active note, or copy it as markdown
- Replies are wikilink-ified: mentions of real notes become clickable links
- A / palette for your custom prompts, commands and skills; @ to attach a note or folder
- Context appears as document cards above the composer; chats name themselves
Connections & extend
- Runs your vault’s .claude/ commands and skills as first-class prompts
- Connections: one-tap import of MCP servers and skills already installed elsewhere
- A session cockpit, a task board, and open loops built on the same memory folder
- Two backends, switchable per conversation: Claude or Codex, your own login
Onboarding
- A fresh vault gets a one-time picker: full scaffold, only the essentials, or nothing
- The memory folder is configurable: a neutral _exo/, or an existing _system/ layer
- Re-runnable and idempotent: it fills in only what is missing, never overwrites
- No imposed folder scheme: how you organize notes stays entirely yours
THE MEMORY LAYER · _system/
Underneath everything sits one folder. It holds the vault context, your preferences, and three kinds of durable knowledge that compound at different rates: learnings captured from a single session,rules that survive because they held across several, anddecisions recorded with their trade-offs and a date to revisit.
None of it is a chatbot’s hidden scratch space. It is markdown, versioned in git, readable by any tool that can open a folder. The agent can only write to it through gated tools that stamp what it wrote. Your memory is a folder you own, not a feature you rent.
REACH FROM YOUR PHONE
The agent is desktop-only; it drives a local CLI. But you are not. A request typed on your phone is written to a queue that Obsidian Sync already carries; the desktop drains it, answers read-only, and writes the answer back through the same synced folder.
It needs no always-on server, no cloud relay, no second account. The two devices meet in a folder. When the desktop is asleep, the request waits; when it wakes, it answers.
ONE METHOD, TWO SCALES
A PERSONAL SECOND BRAIN
One person’s thinking, compounding. People and companies you deal with, concepts you’re working out, decisions you’ve made and why, all cross-linked, so the next question starts where the last one ended, not from scratch.
A COMPANY BRAIN
The same loop, wider. Decisions become shared memory with their trade-offs intact; the graph of people, accounts and projects becomes the team’s working model of what it knows. An agent keeps reading and writing it, so it stays current instead of rotting like a wiki.
The structure that holds either, four folders and the memory layer, has its own page: the brain itself, and a vault you can clone.
WHY THIS HOLDS
A chat window remembers one conversation; close the tab and you re-explain yourself. A database keeps memory too, but behind an API, with the assistant bolted on beside the writing surface. Other agents run over a markdown vault, the closest neighbors, but ask for an always-on server, or record your screen and microphone around the clock to feed themselves.
The method here asks for none of that. The agent lives where you write. The memory is plain files you own, local and versioned. Nothing is always on; nothing is watching. Files you own, not features you rent.
| AXIS | Exo + Obsidian | Notion AI | Hermes + OMI | Chat memory |
|---|---|---|---|---|
| Where the agent lives | Your editor sidebar | Bolted-on assistant | CLI + Discord | Its own app |
| Your memory is | Plain files you own | Behind an API | Markdown files | One conversation |
| Who holds the data | You · local + git | The vendor | You | The vendor |
| Always-on server | None | Cloud | Required | Cloud |
| Records you | No | No | Screen + mic, 24/7 | No |
| Graph-native | Yes | No | Partial | No |
| Runs on a local vault | Yes | No | On its server | No |
NEIGHBORS SURVEYED: NOTION AI · HERMES + OMI · GENERIC CHAT MEMORY. AXES ARE STRUCTURAL, NOT SCORED.
APPROACH VECTOR
Two moves get you the whole method at once: the sample vault to see the shape, and Exo to put the agent in the sidebar.