Operator — Data & Storage

Where a customer's life lives, and what Claude reads when
Generated 2026-07-09 · lives at AI dashboard/public/operator-data-architecture.html · grounded in the running silo implementation

The private space — one folder per customer

🗂️ Your space = a folder

Every customer gets one isolated directory on the operator host. Claude runs with its working directory locked to that folder (claude -p, cwd-pinned) — it physically starts inside your space and reads nothing outside it. No shared database, no commingled tables.

isolation by filesystem

📝 Memory = plain markdown

Everything the operator remembers is written as human-readable markdown files — todos, notes, decisions, people. No vector store, no opaque blob. The customer can open, edit, or delete any memory with a text editor.

no black box

📖 The employee handbook

A per-customer CLAUDE.md holds the operator's persona, standing instructions, language, boundaries, and shortcuts ("add todo" → edit todos.md). It is loaded automatically at the start of every single conversation — this is why the operator never forgets who it works for.

always in context

👤 Profile

context/about-me.md — filled by the guided onboarding interview (/onboard): who you are, family, goals, preferences. Read on demand when a task needs personal context.

from onboarding
customers/ └─ anna/ ← one isolated space per customer ├─ CLAUDE.md ← employee handbook: persona, rules, shortcuts (auto-loaded) ├─ context/ │ └─ about-me.md ← profile from onboarding ├─ inbox/ │ ├─ todos.md ← the todo list (checkbox lines) │ └─ notes, drafts, plans ← everything captured by chat ├─ decisions/log.md ← "log decision: …" appends here, dated ├─ files/ ← photos + documents sent in chat, files it makes for you ├─ logs/ │ ├─ audit-2026-07-09.jsonl ← append-only audit: one line per turn │ └─ telegram-mirror.md ← readable chat mirror (rotated) ├─ .env.local.enc ← secrets, encrypted at rest (SOPS + age) ├─ .config/gws/ ← this customer's OWN Google credentials (isolated) └─ .session-state.json ← 60-min conversation window + rate limits

What each scoop stores

AbilityData it createsWhere it lives
💬 Chat & memory (base)Conversation audit, chat mirror, remembered factslogs/*.jsonl, telegram-mirror.md, markdown memory files
✅ Todos & notes (base)Todo checkboxes, freeform notesinbox/todos.md, inbox/*.md
🤝 Contact memoryPeople, context, follow-up threadscontacts.md in the space (+ customer's own Google Contacts if connected)
🌅 Morning brief / 📬 Mail digestNothing new — reads calendar/mail, writes only the message to youtransient; source data stays in the customer's Google account
📅 Calendar concierge / ✉️ EmailEvents and mail live in the customer's own Google accountcustomer's Google — reached via their own isolated credentials, never copied wholesale
📡 Market scannerWatchlist + seen-items list (dedupe)inbox/watchlists.md + seen-log in the space
🌙 Content factoryDrafts produced overnightdrafts/ in the space
🎨 Visuals / 📎 files in chatGenerated images, received photos and documentsfiles/ in the space
🏡 Family & householdShared lists, home baseline valuesmarkdown in the space (family members' locations are read live, never logged)
Design rule: the operator's copy of your life is markdown in your space; systems of record stay where they already live (your Google Calendar stays your calendar, your inbox stays your inbox). The operator holds credentials to act on them — not a mirror of their contents.

How context reaches Claude — per message

1 · Message arrivesYour chat message hits the listener bound to your bot. Unknown senders are dropped before Claude ever runs.
→
2 · Session checkWithin 60 min of the last message? The running conversation continues. Longer? A fresh session cold-starts — memory files, not chat scrollback, carry the context.
→
3 · Claude starts in your spaceclaude -p launches with cwd locked to your folder. Handbook (CLAUDE.md) loads automatically. Tools are gated to your tier.
→
4 · Reads on demandIt opens only the files the task needs — todos for "what's on my list", contacts for "who is Tom". Nothing is bulk-embedded or pre-indexed.
→
5 · Acts & recordsFiles edited in place, outbound actions previewed, one audit line appended, reply sent. Process exits — nothing lingers in RAM.
Model boundary: the message plus whatever files Claude opened travel to the model API for that turn, then fall out of context when the session ends. The model provider processes them transiently under its commercial no-training terms — the durable copy of your data only ever lives in your space. Tier gates decide reach: base tier can only touch files in the space · read tier adds read-only calendar/mail via your own credentials · full tier adds send/write, with previews on by default.

Secrets & credentials

🔑 Encrypted at rest

Bot tokens and API keys live in .env.local.enc — encrypted with SOPS + age. Decrypted into process memory only for the seconds a task runs. Never plaintext on disk, never in git, never in logs (provisioning scrubs tokens from output by design).

sops + age

🧍 Your Google, your credentials

Calendar/mail abilities use a per-customer credential directory (.config/gws/) — the customer signs into their own Google account. No shared service account, no operator-wide Google access. Revoke in your Google settings and the ability dies instantly.

isolated oauth

🔄 Rotation & kill switch

Any token can be rotated at runtime without touching history. An emergency stop disables the listener, logs the bot out, and renames the secrets file so nothing can auto-restart until a human turns it back on.

rotate / kill

The customer guarantees (and what backs them)

📤 Export = zip of your folder

Because everything is files, "export my data" is literally an archive of your space — readable the moment it lands. No proprietary format to escape from.

portability

🗑️ Delete = destroy the folder

Cancellation keeps your export link alive 30 days, then the space is destroyed. One folder to delete means deletion is verifiable, not a promise scattered across databases.

right to be forgotten

🚫 Never trained on

Data is processed under commercial API terms — not used to train models. There is also no cross-customer anything: no shared index, no "insights from users like you".

no training

🧾 Auditable by the owner

The append-only JSONL audit plus the readable chat mirror mean a customer (or their admin) can always answer "what did my operator do, and when" — line by line.

accountability

Where the space physically lives

🏠 Today — managed private host

Spaces live on the operator-managed host in Sweden (EU rules), one directory per customer, chat/creds/logs excluded from any code repository by ignore rules. Tailscale-style private networking for any admin surfaces — nothing customer-facing is publicly exposed.

current implementation

☁️ Roadmap — encrypted volumes

Per-customer encrypted volumes on EU cloud, so a host failure or move never touches plaintext. Same folder model, same export/delete story — just replicated and rebootable.

planned

🔌 Roadmap — bring your own storage

For customers who want physical custody: the space syncs to their Drive/NAS as the canonical copy. Pairs with "Bring your own AI" (sign in with your own model provider) for a fully customer-owned stack.

planned · pairs with BYO-AI
One-line answer to "where is my stuff?": in one folder that is yours — handbook, memory, files, logs — plus encrypted keys that let your operator act on accounts you already own. Claude visits the folder to work; it never takes the folder with it.