One conversation, kept for a lifetime — and a quiet intelligence that learns who you are from it.
▶ Watch the demo — your journal as a living star constellation · Download the latest release
Most journaling ends where it should begin: the words go down, and nothing reads them back. SemanticPortrait is the other half. You write to a single, unending thread, and a frontier model listens the way a good analyst would — noticing, remembering, gently telling you the truth. Over time it draws a portrait of you that you can actually see — a night sky where every star is something you care about — and returns it to you when you need it most: when you're about to believe something about yourself that isn't so.
Your journal as a living constellation. Every star is one thing the analyst has learned about you; each cluster is a theme, and a star's position encodes its energy (↕) and mood (↔). (Shown with sample data — your real map stays encrypted on your device.)
One lock over everything. The vault is SQLCipher (AES-256) encrypted at rest and opened only by Windows Hello or your passphrase — the database stays closed while locked.
Turn an ongoing conversation into a living, self-correcting model of who you are — and let it keep you in reality, not in your distortions.
It is built around a simple discipline most of us can't keep alone: separate what happened from the story you told about it. The model records the timestamped facts, marks its own inferences as inferences, and holds the line between the two so the record stays honest as the years accumulate.
- Your memory gets a second reader. Everything you write is studied, indexed, and connected — so the one thing you said two years ago that matters tonight actually comes back to you.
- You see yourself change. The Constellation isn't clip art; every star, color, shape, and orbit is a deterministic function of the analysis. When your inner world shifts, the sky moves.
- It argues with your distortions, not with you. Rejection-radar, hope-as-fact, the inner critic's verdicts — named, tracked, and countered with your own recorded evidence.
- It keeps score on itself. Falsifiable predictions with observable criteria, checked when reality arrives. Calibration, not reassurance. Anti-delusion by construction.
- It remembers your life admin too. Reminders, events, and open loops live in the same brain — one question ("what's coming up?") returns your whole plate, and Windows notifications catch you when the app doesn't have your attention.
- One eternal chat thread — your whole history in a single conversation, streamed token by token, rendered in clean Markdown. Nothing is ever deleted.
- A self-updating memory. Every meaningful exchange is studied by a background analyst that writes durable notes, stores stable facts about you and the people in your life, and keeps the record current as new information refines the old.
- Semantic recall. Ask about something from months ago and it finds it — a vector index over every entry and note surfaces the relevant past, not just the recent.
- The Constellation. A living map of you — people, themes, patterns, distortions, values — radially laid out around your core self, with named star groups, weak ties tracing back to what anchors them, and a sky that breathes: gas drift, meteors, stars that twinkle and sway.
- A prediction ledger / track record. It logs falsifiable forecasts and scores itself when reality arrives, so you can see — over time — where your instincts are sharp and where they lie.
- Reminders that respect your attention. Due items arrive in-thread when you're looking and as a Windows toast when you're not; quiet hours roll late-night pings to morning; a once-a-day digest tells you what's on your plate.
- A to-do drawer. When you mention something you need to do, it's quietly tracked — and a button by the bell opens a list you can actually see: check items off, or ask the analyst a question / give it an update on any one of them, right in the thread.
- Inspectable thinking. Tool calls appear as quiet reference chips in the chat ("noted this for later"); click one to see exactly what it recorded. A developer trace shows the full reasoning.
- Bulk import & export. Bring in years of old notes and prior analysis from text files — a pre-pass counts the facts and a live status bar tracks the work — or export your whole thread.
- Thought compaction. Conversations older than two days fold into a rolling summary to stay fast and focused; the full detail never leaves the searchable store.
- Guided onboarding. A short, conversational setup that gets the portrait started — and runs again from a clean slate if you ever erase everything.
- Cost in the open. Live token-usage and spend tracking, per session and lifetime — visible or hidden as you prefer.
- One thread, forever. Never a new chat. Continuity is the product — nothing is deleted, edits are versioned, and the past stays the past.
- An analyst, not a cheerleader. The voice is calm and unsentimental. It won't flatter you, and it won't catastrophize with you. When you drift into rejection-radar or hope-as-fact, it says so — kindly, plainly.
- It keeps score. When you forecast something, it logs a falsifiable prediction with an observable criterion, then checks itself against what actually happened.
- The Constellation. A living mindmap of you whose colours, shapes and motion are a deterministic function of the analysis itself. Two people's portraits are comparable; none of it is random.
The analysis you can't argue with runs somewhere you can't reach. A clean-room subagent does the durable thinking in a fresh context that never sees the live chat — so the long-term model can't be coaxed, charmed, or talked soft by what you say in the moment. Personalization flexes its tone; it never touches its truthfulness. New claims are researched against what's already known before they're committed — confirmed, refined, or flagged as a contradiction — so the portrait accretes rather than drifts.
A journal is the most sensitive document most people will ever produce, so here is the honest version, not the marketing version.
The trade at the center: by default, SemanticPortrait sends your entries to a cloud frontier model (OpenAI today; Claude and others on the roster) — because frankly, frontier models are better at this job, and an analyst that misreads you is worse than no analyst. That means the model provider processes your words under their privacy policy, not this app's. Masking helps; it does not make a cloud call private. If that trade isn't acceptable to you, switch to local inference below — the app works either way, and the choice is per-provider, changeable at any time.
What the app itself guarantees:
- Storage is local and encrypted, always. No accounts, no sync, no app server, no telemetry. The vault is SQLCipher (AES-256) SQLite on your machine; the key is derived from your Windows Hello or PIN unlock — no copy of it sits in the clear.
- Notifications are privacy-classified. Before a reminder is scheduled as an OS toast, a classifier decides whether its text is safe for a lock screen; anything personal shows a generic placeholder instead — and the check fails safe to private.
- PII masking on the way out. An optional local pass pseudonymizes emails, phone numbers and ID numbers — not names — before a cloud request leaves the machine. Treat it as harm-reduction, not anonymity — freeform journal content (names included) can still re-identify you.
- Local embeddings. Semantic recall can run on an on-device MiniLM model, so the index that knows everything about you is built without a network call. Until you install it, semantic recall sends entry text to OpenAI for embeddings (masked by default).
If you want nothing to leave your machine, SemanticPortrait speaks the OpenAI-compatible chat protocol to any inference server you point it at:
- LM Studio is supported out of the box — run a local model and chat and analysis stay on your machine. Semantic recall stays local only once you also install local embeddings (⋯ → Local embeddings); until then, entry text still goes to OpenAI for embeddings (masked by default). With both a local model and local embeddings, nothing leaves.
- Any OpenAI-compatible endpoint works the same way — Ollama, llama.cpp's server, vLLM, LiteLLM, a box in your closet — the base URL is configurable in ⋯ → LLM settings.
- Third-party "private-ish" hosts (a VPS you rent, a privacy-marketed inference service) sit in between: better than a data-hungry default, but you're still trusting someone else's machine. Read their retention policy; don't take "private" on faith — including from us.
Expect a quality trade: local and small hosted models are noticeably weaker analysts than frontier models today. That gap is why the cloud is the default — and why the switch is yours to flip, not ours.
If you already pay for ChatGPT, ⋯ → LLM settings → ChatGPT plan · Codex lets you sign in with your subscription and use it for analysis instead of buying pay-per-token API credit. Click Sign in with ChatGPT, approve the login in your browser, and the tokens are stored encrypted in your vault (never a plaintext file) and refreshed silently.
Be honest with yourself about what this is: it rides OpenAI's private Codex backend — the same endpoint their Codex CLI uses — which is not the OpenAI API and not a sanctioned third-party integration. It can put your ChatGPT account at risk of being flagged, the endpoint can change or break without notice, and it's limited to whichever model Codex currently allows. Your entries are still masked before they leave (like any cloud provider), but treat this as an at-your-own-risk convenience, not a supported path. The API-key and local options above are the stable ones.
A desktop application, not a website — .NET MAUI Blazor Hybrid on .NET 10, native to Windows on ARM, light and dark. Conversations reach a model through a streaming client — cloud (Claude / OpenAI today; Kimi, GLM and DeepSeek on the horizon) or fully local (LM Studio and other OpenAI-compatible servers). Memory lives in encrypted-at-rest SQLite (SQLCipher, AES-256) with a small vector index for semantic recall. Older turns fold into a rolling summary; the detail never leaves the searchable store. Voice — on-device speech via the Snapdragon NPU (WhisperToMe for listening, Supertonic for speaking) — is the last mile.
Download the latest setup EXE, MSI, or portable zip (x64 / arm64) from
Releases — per-user install, no
admin needed. First-run SmartScreen notes and where your data lives:
docs/INSTALL.md.
MIT — see LICENSE.
- ▶ The demo video — see the constellation live.
docs/INSTALL.md— install it as a user.docs/GETTING_STARTED.md— build it as a developer.CONTRIBUTING.md·SECURITY.md