The Architecture Behind Atlas

Atlas is a personal AI assistant you talk to like a person. Five layers make it work: you (on any device) chat with the Atlas App, which hands your message to the AI Engine. The engine picks a brain (LLM Provider), reads and writes to your private memory & vault, and reaches into your life through integrations and automations.

When you send a message, Atlas doesn't fire it raw at an AI. A tier selector first decides how much brainpower your message deserves — a quick "thanks" needs almost nothing; "plan my week from my calendar and inbox" needs your full memory and vault. Atlas then assembles a package: its personality, your relevant memories, and any vault documents that might matter, and hands it to wingmanChat, which calls the chosen provider and returns the reply.

Not every message deserves the same brain. A two-word reply shouldn't cost the same as a 2,000-word life plan. The tier selector routes each message to one of three tiers — each with a different model, a different slice of your history, and a different price. This is what keeps Atlas affordable to run at scale.

In a long conversation, your personality prompt and early messages are identical every turn. Without caching, you'd pay full price for them every time. With prompt caching, the provider stores that stable prefix once and reuses it — charging about 10% of normal for the cached part. For a personal assistant that you chat with all day, this is the single biggest reason it's economically viable.

Atlas talks to three AI providers — OpenAI, Anthropic, and Google — each on its own account and billing. You might start on Google's free tier, add OpenAI's free credits, and upgrade to Anthropic later. The system handles all of that gracefully, which matters for a consumer product where reliability is everything.

Toggle: an admin sets WINGMAN_PROVIDER_CONFIG (a JSON secret) to enable/disable any provider. A provider is "available" only if enabled and its key is set. Failover: if the chosen provider hits a quota or overload error, Atlas silently retries the next available one — so a free tier running dry mid-conversation self-heals instead of breaking.

This is what separates Atlas from a generic chatbot. Atlas doesn't just remember your last message — it holds a structured, long-term model of you, and that model is what makes every reply feel like it's coming from someone who actually knows you.

The five memory types

Atlas stores five kinds of memory, each with a purpose. Facts are stable truths about you — your name, your role, your family. Preferences capture how you like things done — short emails, bullet points over prose, mornings for deep work. Goals are what you're working toward — the promotion, the side project, the trip. Summaries are condensed versions of past conversations so old chats aren't lost. Journal entries are things that happened to you — the meeting that went sideways, the win you landed. Every memory carries an importance score (1–5), searchable tags, and a source date so Atlas can weigh what matters most and what's still current.

How memories are born

You never have to explicitly "save" a memory. As you talk, an auto-summarizer runs in the background every ~10 user messages: it reads the recent thread, extracts anything worth remembering long-term, classifies it into one of the five types, scores its importance, tags it, and writes it to your private memory store. Trivial chatter is skipped; meaningful context is kept. The same pipeline powers briefing absorption — paste in notes or an exported conversation from another tool and Atlas extracts every useful memory from it in one pass.

How memories are retrieved

When you send a message, Atlas doesn't dump your entire memory into the prompt — that would be expensive and noisy. A relevance scorer ranks your memories against your message: title keywords, tag overlap, and content similarity. Only the top few (typically 3 for a standard turn, more for a full-tier turn) make it into the context window. This is why Atlas can hold thousands of memories and still respond in seconds — it loads the needle, not the haystack.

The three companion stores

Memory is only the beginning. Atlas also maintains three companion stores that turn memory into action. Open items are the things you've committed to — extracted from chat, tagged by life bucket, tracked to completion. Decisions log what you chose, the context, the alternatives, and (later) the outcome — so you can revisit why you picked a path. The vault holds your documents — drafts, plans, articles, spreadsheets — with full version history and sharing controls. Together, memory + open items + decisions + vault give Atlas a complete, living picture of your life.

Cross-device continuity

Every memory, vault doc, open item, decision, and conversation syncs across your devices in real time. Atlas subscribes to live changes on the backend, backfills on window focus, and runs a 60-second safety interval — so the thread you started on your laptop this morning is waiting on your phone this afternoon, complete with its full history. A daily briefing is generated each morning from your open items and overdue tasks, so Atlas greets you with what matters instead of a blank screen.

A personal assistant that remembers everything about you is also a serious privacy responsibility. Atlas is built privacy-first — not as an afterthought, but as a core design principle. Your data is yours: you control it, you can see it, you can delete it, and no one else can touch it.

Layer 1: Access control — your data is locked to you

Every record Atlas holds about you — every memory, every vault document, every conversation, every open item, every decision — is row-level secured. The technical rule is simple: read: { created_by_id: {{user.id}} } — only the user who created a record can read it. Even though every Atlas user's data lives in the same database tables, the database enforces that you only ever see your own. There is no "admin can see all user data" path for personal records — your memories are yours alone.

Passcode-locked conversations

Some conversations are more sensitive than others — a personal health matter, a financial decision, a difficult relationship. Atlas lets you passcode-lock any conversation so it's hidden behind a gate screen until you enter the code. The passcode is SHA-256 hashed before it's stored — the plaintext is never saved, not in the database, not in local storage. Even someone holding your unlocked phone can't read a locked conversation without the code. And the hash is verified client-side, so the server never sees the plaintext passcode either.

The private vault

Your vault documents — drafts, plans, articles, spreadsheets — are also RLS-locked to you. Each document has full version history (every save creates a snapshot you can roll back to) and granular sharing controls: you can grant another Atlas user view, comment, or edit access to a specific document, and you can revoke it at any time. Shared access is tracked in a WingmanDocumentAccess entity — you always know who can see what.

Layer 2: Your rights — see, edit, delete

Privacy isn't just about keeping others out — it's about giving you control. You can see every memory Atlas holds about you (they're listed in the sidebar and the memory panel). You can edit any memory to correct it. You can delete any memory, conversation, or vault document at any time. And if you delete your Atlas account, all your data is removed — the right to be forgotten is built in, not a support ticket.

What Atlas does NOT do

Atlas does not train AI models on your data. Your conversations and memories are not fed into a training pipeline. Atlas does not sell your data to third parties. Atlas does not show advertising — the subscription model means the user is the customer, not the product. Atlas does not share your data with other users — ever. Your memories, vault, and conversations are invisible to everyone but you.

Data retention

Conversations are capped at 40 messages and 40,000 characters per record to keep them manageable — older messages are summarized into long-term memories rather than kept verbatim forever. Memories are kept indefinitely but are always editable and deletable. Vault documents and their version history persist until you delete them. Archived conversations (cleared but restorable) are kept in local storage for up to 100 entries.

Encryption

All data in transit is protected by TLS. All data at rest is encrypted by the platform's managed storage layer. Passcodes are SHA-256 hashed. OAuth tokens (for Gmail, Calendar, Drive, Sheets) are stored by the platform's connector system and never exposed to other users or to the AI models.

Third-party AI providers

When Atlas thinks, it sends your message (plus the assembled context — personality, relevant memories, vault snippets) to an AI provider: OpenAI, Anthropic, or Google. Each provider has its own privacy policy governing how they handle API inputs. On the Free and Pro tiers, Atlas uses its own provider keys, so your prompts are sent under Atlas's API account. On the Enterprise tier, BYOK (Bring Your Own Key) lets your team supply its own provider keys — your prompts go directly to the provider under your own account, never touching Atlas's billing or infrastructure. This is the strongest privacy posture for organizations that need it.

Compliance readiness

Atlas is designed with GDPR and CCPA principles in mind: data minimization (we only store what's needed to be your assistant), transparency (you can see everything), right to access, right to rectification, right to erasure, and data portability (you can export your data in multiple formats). Formal compliance certifications will follow as the product scales.

Atlas connects to your Google workspace through OAuth — Gmail, Calendar, Drive, and Sheets. Once authorized, it can read your inbox for commitments, look at your calendar to plan your day, save the documents it drafts to your Drive, and sync your tasks to Sheets. This is what turns Atlas from a chat window into an assistant that's actually in your life.

Atlas is commercialized on three tiers. Free runs on Google's free-tier AI to let people try it with zero friction. Pro unlocks the mid and premium brains, unlimited memory, all integrations, and cross-device sync — priced to cover the direct-API cost plus margin. Enterprise adds BYOK so teams bring their own AI keys, turning the subscription into pure software revenue insulated from high-usage customers.

The same engine handles a one-line reminder and a full week-planning pipeline. The difference is which tier fires, how much of your memory is loaded, and how many calls chain together.