How it works
Your exports go in. A private brain comes out — plain Markdown your AI can read, built on your machine with zero network calls. This page is the whole mechanism, shown on the two demo brains.
Export · Build · Ask
Every AI brain ships empty. This is the step that fills it — from data you already hold, on a machine you already control.
Every platform that matters must hand you a machine-readable copy, free — GDPR Article 20, the EU Digital Markets Act, CCPA. How, for each of the 27
Studio runs the brain builder on your machine: detect, match, build. Zero network calls; nothing uploaded. Or ask Claude Code or Codex through the Agent Skill.
Explore the graph, ask a plain question and get an answer with receipts, or hand an agent a job.
data/ ├── linkedin/ │ ├── Connections.csv │ ├── Positions.csv │ └── messages.csv # never parsed ├── facebook-export.zip ├── instagram-export.zip └── takeout.zip ✕ four formats, zero links, the same person three times
_memory/ _notes/ 00-me/ 10-people/ 15-organizations/ 20-reputation/ 30-voice/ 35-shopping/ 40-career/ 45-jobs/ 47-travel/ 50-mirror/ 60-learning/ 80-search/ 85-places/ 90-synthesis/ 95-goals/ 96-agents/ _HEALTH.md _SUMMARY.md CLAUDE.md Dashboard.md graph.json Home.md ✓ one graph — typed, linked, one note per person
The build
The transform is deterministic, not generative: the same archive builds the same brain every time, and you can read the code that decides. AI is used only on the residue — files no mapping recognises.
00-me or 00-org.# Build report Built: 2026-10-03 Subject: person · Emitters: obsidian Sources: amazon, facebook, github, google, instagram, linkedin, reddit, spotify, strava, tiktok, whatsapp, x ## Log - [amazon] purchase: 64 from Retail.OrderHistory.1.csv - [amazon] purchase: 18 from Digital Items.csv - [amazon] purchase: 2 from Subscriptions.csv
The ontology
An ontology is just the list of kinds of things in your brain and how they connect: people, companies, posts, places, purchases — each a type of note — and labelled links like works at. Studio derives it from your notes, deterministically; no AI decides a type.
Resolved means found in two or more exports and matched by name — so the friend who is also a connection and a follower is one note. From the John demo brain: 329 of 438 people.
Identity
LinkedIn: 437 people · 34 posts · GitHub: 142 people · 10 repos · Facebook: 138 friends · 40 check-ins · Instagram: 127 followers · 21 posts. Built into one brain they become one you: 438 people, 329 of them Resolved across two or more exports, matched by name.
In the demo, Hedy Toth appears in 4 exports — Facebook, Instagram, LinkedIn, Strava. Every other tool keeps those as 4 separate records. Here they are one note — built on your machine, without an email address or a phone number ever reaching disk.
name-onlyMatching is on the normalised name only, and precision-biased on purpose: variant spellings and initials stay separate rather than being guessed together, and no AI decides a merge. The flip side: two different people with one name share a note. Between two brains, conflicting profile links block a merge. Better matching is roadmap, not shipped.
--- type: person title: Hedy Toth sources: - facebook - instagram - linkedin - strava status: cold company: "[[Acme Robotics]]" role: facebook friend strength: 2 last_contact: 2025-09-19 connected_on: 2024-11-21 ---
Agent-ready
Typed frontmatter, resolved wikilinks, real Obsidian Properties. An agent doesn’t have to guess what a note means — and when it answers, it cites the notes it stood on.
Typed frontmatter on every note — person, organization, place, deal — with ISO dates and enums, not free text. A note merged from several exports lists them in sources.
Every [[link]] resolves to a real note, and graph.json carries the typed, weighted edges — so an agent walks the graph the way you click through it.
00-me → 10-people → 40-career → 85-places → 90-synthesis; a company brain gets company-named layers (40-pipeline, 65-documents). Stable paths a script can rely on.
Each brain ships with a guide for Claude Code (an AGENTS.md for Codex): the read order, the privacy rules, how to keep its memory.
Receipts, not vibes
Ask a plain question. A bare model has nothing to count. Point the same model at the brain and the answer comes with the chart, the folder it counted and the exports behind it.
A chart is drawn only from numbers counted in the brain. A chart that cannot say what it was counted from is not drawn.
The Receipts row names the folder it counted and the exports behind it; Show the count opens the numbers.
Open Studio in your browser and ask the demo brains, or ask Claude Code or Codex over your own.
Runsdeveloper preview
Every agent runs the same daily: a goal you set, a dated note opened first, discovery, screening, the shortlist, actions behind your gates, the log — and a weekly review. When it works, Studio already follows its session on your machine. Runtime draws what it sees — nothing is sent anywhere new.
reports/<Daily> <date>.md in the agent’s own layer before anything else, and keeps it current — that note is the run’s home, beside the chat and on its line.pipeline.jsonOne record per item, written by the agent’s own scripts; from it, one file in its folder of your brain — its funnels, their stages, the items in each — and the dashboard and KPI note. Counted, never typed.this computerThe past is what your chats kept, not a full audit log, and a screenshot lives only as long as the session. The dailies are a developer preview and run while the computer is awake.
Brain recall
Brain recall starts from your question and follows the links in your brain to the notes that matter — then shows you which ones. It runs on your machine, makes no network calls, and gives the same answer every time.
works_at, member_of, attended — so a note two hops away still surfaces. Computed as Personalized PageRank.$ retrieval.py john-brain --top 8 \
"who could introduce me to an investor?"
0.1895 fundraising (goal)
0.1822 Find warm paths to investors (routine)
0.1170 Priya Nair (person)
0.1170 Northwind Capital (company)
0.0528 Job pipeline (dashboard)
0.0406 Home (moc)
0.0149 Kiln Ceramics (company)
0.0145 Tidewater Marine (company)
When Studio lights the graph, every pulse comes from the real run: an edge lights only if it truly connects two notes.
No embeddings, no vector store, no model trained on your data. It is graph retrieval over links you can open and read.
Studio, the browser, the sbl command and retrieval.py rank a brain identically, so an answer is reproducible.
Next month
Download a newer one and press Update (or run the build with --refresh): the brain is updated, not rebuilt.
_GENERATED.json is the manifest: a person or a place keeps its file and its backlinks across updates.<name>.new.md._notes/ and the agents’ layers keep their own manifests and are never overwritten._UPDATE_REPORT.md lists added, updated, pruned and kept-because-edited. Running it twice changes nothing.roadmapHosted, scheduled sync — the always-fresh version — is on the roadmap. The local Update shipped.
Proof, not polish
Every build ends with a _HEALTH.md: orphan notes, unresolved links, snapshot age, open conflicts. This is the one that ships with the company demo — unflattering on purpose.
known_contacts: 0.# Brain health Nodes: 704 · edges: 1417 · cross-source edges: 74 ## Snapshot freshness (per source) Exports are snapshots — sharp on history, blind to this week. The latest record date each source actually carries: (no dated records found) ## Links
The supply
Every platform that matters must hand you a machine-readable copy of your own data, on request, for free. The hard part was never getting it out — it is that what comes back is unusable, so it gets downloaded once and never opened.
A structured, commonly used, machine-readable export. Not a favour or an API deal that can be revoked — an obligation, enforced with nine-figure fines.
The largest platforms must move to continuous, real-time portability. The pipe this depends on is being widened by regulators.
It is your export, requested by you, processed on your machine. No scraping, no terms to break, no access anyone can switch off.
Custody, not promises. Gathering it here doesn’t create the exposure — it ends it. The build makes no network calls, so the question stops being “where does the vendor say this goes?” and becomes “which machine is it on?” The privacy model
Integrations
A brain is plain Markdown on your disk, so every door is open: browse it in Obsidian, hand it to Claude Code or Codex, chat over it with your own model, write it out for GBrain. Nothing here locks you in — that is the point.
Typed frontmatter that Obsidian reads as native Properties, every wikilink resolved, a Home note as the map. Dataview and Bases read it as-is; no community plugin is needed. Plugins like Obsidian Copilot or Claudian work on it unchanged, because it is just Markdown.
One open-source engine, packaged for Claude Code and OpenAI Codex (the Agent Skills open standard). Install it once and your AI knows how to build, update and use a brain — locally, with zero network calls. It also runs as a plain command line.
Studio’s chat and the Avatar answer with the AI you choose. Keys stay on your machine; calls go straight to the provider; how a call is billed is up to how you signed in.
The agents run through Claude Code; on Codex they are a narrower, flattened skill with no browser tooling. The hosted browser Studio explores the demo brains only.
GBrain is a brain layer — search, synthesis, team serving. Second Brain Link can write a company brain in the open GBrain repository format (--emit gbrain), so a team starts it full. Not affiliated with the GBrain project; the emitter is built and not yet validated against a running GBrain.
A stdio MCP server ships with the engine — no open port, nothing listening on the network. Point Claude Desktop, an IDE or your own agent at it: sbl_ask, sbl_query, sbl_inbox, sbl_routines, sbl_sources, sbl_build — with the same ranking Studio uses.
Do I need Obsidian? No. Obsidian is the nicest way to browse the files, Studio is ours, and Claude Code or Codex read them directly. Any Markdown editor works.
Can I add my own integration? Yes — it is MIT. A new source is usually a JSON mapping, an output is a small Python module beside the Obsidian and GBrain ones, and an agent is a plugin built for Claude Code and Codex from one source. The repository
How it sits beside Glean, Microsoft 365 Copilot and GBrain — dated, cell by cell: the comparison
Start
Free, local and open source. 27 sources in; one brain on your machine out.
Second Brain Studio · Mac (Apple silicon and Intel), Windows and Linux · free · MIT
Just looking? Open Studio in your browser — two demo brains, nothing to install