A discreet, client-controlled assistant appliance that reads approved sources, prepares concise briefs, keeps a reviewable memory, and asks before anything meaningful is done.
The client keeps using email, calendar, files, notes, staff messages, travel confirmations, and household routines. The system connects to approved sources, builds a private memory from them, and presents briefs, search, corrections, and approvals in one private console. Messages provide context; the console is where decisions happen.
The design separates context, memory, and action. This makes each source useful without giving it more power than it should have.
iMessage, WhatsApp, and notifications feed context or alerts. Approval, durable memory, and system state stay inside the private console.
The private PWA, a secure web app, handles approvals, memory corrections, source health, audit review, and account-level controls.
Each connector has an allowlist, credential scope, read/write rule, freshness signal, audit trail, and off switch.
Raw sources remain immutable. The assistant compiles candidate facts into a personal wiki, then sensitive or durable changes go through review.
The system turns daily context into a maintained memory layer and a short decision surface.
The private Mac Mini hosts the appliance. Sources feed a local archive. A Karpathy-style compiled wiki turns repeated context into durable memory. The console governs approval, correction, and review.
The memory wiki is maintained by the assistant, reviewed by the client, and rebuildable from raw sources. Raw messages, emails, and model outputs stay as evidence; reviewed wiki entries become durable memory.
Behind the private console, small scheduled programs connect to approved sources, keep local records, and report when something needs attention. The AI works from prepared context; it does not hold credentials or freely operate external apps.
The private PWA presents the daily brief, approval queue, search, memory corrections, source health, and audit trail. The technical machinery stays behind the interface.
Today has 5 commitments. One conflict needs a decision. Three emails need attention. Paris has two unresolved items.
Draft reply to attorney. Proposed calendar move. Passport appointment options. Nothing has been sent.
"What am I forgetting before the trip?" The answer cites calendar, email, and saved preferences.
"I learned these four facts this week. Keep, edit, or discard."
Clear decisions, short summaries, source links, and a record of what the assistant did.
human surfaceConnectors, model routing, file search, scheduled jobs, source health, logs, and usage tracking.
system layerPermissions, external actions, memory changes, support access, and data leaving the device.
trust layerThe memory layer follows the Karpathy LLM Wiki pattern: raw sources remain preserved, the assistant compiles them into linked pages, and periodic reviews find drift, contradictions, duplicates, and stale facts.
Email, calendar items, files, transcripts, imports, and optional message context are stored or referenced with provenance. This layer is the evidence base.
The assistant maintains pages for people, places, routines, preferences, trips, decisions, and open loops. Each page links back to the sources that justify it.
Sensitive or durable facts wait for review. Weekly lint checks flag stale claims, contradictions, duplicate entities, orphan notes, and memories with weak evidence.
For a personal assistant, WhatsApp history may contain the richest everyday context: family logistics, staff coordination, travel details, informal decisions. But because there is no official personal WhatsApp history API, it should be offered as an optional local context source, clearly separated from the reliable core.
Where the client explicitly allows it, selected WhatsApp exports or the local WhatsApp Desktop store can seed a private message index on the Mac Mini. This creates a starting corpus while keeping the unofficial nature of the source clear.
An in-house connector can use a client-approved linked-device session to mirror selected chats locally as new messages arrive. It is built for context only: no send, no reply, no reactions, no group changes.
Messages enter as raw context requiring review. The assistant extracts candidate facts, decisions, dates, people, and open loops; the client reviews what becomes durable memory inside the private console.
For this client, privacy is part of the product. The system should make clear where data lives, which services are connected, who can access support, and what is logged.
The console is available through the client's private network and only to devices explicitly allowed into that network.
The console requires its own login, ideally passkey-based, with short-lived sessions and approval-specific confirmation.
The first release should observe and summarize. Writing email, changing calendar events, or contacting people comes later, one permission at a time.
Each connector defines what it may read, what it may prepare, and which actions are unavailable by design.
Support should follow a break-glass procedure: requested, approved, time-limited, recorded, and easy to revoke.
The first release proves privacy, source reliability, daily briefing quality, and memory review before any external action capability is added.
Client-owned hardware, encrypted disk, private network, app authentication, credential storage, zero-retention provider terms, backup policy, and break-glass support procedure. No accounts connected before this is done.
Calendar, email, contacts, selected files, and tasks feed one concise morning brief inside the private console. Notifications can say only that a brief is ready.
Meeting summaries, preferences, people, places, routines, open loops, and optionally selected WhatsApp context become reviewable memory candidates. The client corrects what the assistant thinks it learned.
Email drafts, calendar proposals, scheduling options, and travel prep appear as approval tickets. The console records the source, the proposed action, the approval, and the result.
Only after the core loop works for several weeks do we expand into the more personal operational layer: school logistics, household maintenance, travel orchestration, and staff coordination.
It runs in the client's environment, connects to approved sources, builds a Karpathy-style personal memory wiki, summarizes what matters, keeps decisions inside a private PWA, and records its work. The value is simple: reliable context, reviewable memory, and explicit approval before consequences.