Drop files to attach
Images, PDFs, DOCX, TXT, CSV…
Third-person facts starting with "The user…" · Enter to save · Shift+Enter for newline
| # | Fact |
|---|---|
|
⚠
|
Uses LM Studio to distill facts
Drop a file or click to browse
PDF · DOCX · TXT · MD · PNG · JPG · WEBP · GIF · HTML · CSV
| # | Fact |
|---|---|
|
⚠
|
PDF, Word, text, or image → LM Studio extracts facts
Irreversible database reset
Keep
Remove
Archive
Remove (trivial)
Write the combined fact:
AI suggestedCount tokens to see how many facts fit in an LLM context window.
Filters (uses Memory tab filters)
Tag: Entity: No filters — counting all facts
Total Tokens
Facts
Avg Tokens/Fact
Download all facts (including archived) as a backup.
JSON = full backup with metadata · CSV = spreadsheet-friendly
Import a previously exported JSON backup. Semantic dedup runs automatically — no doubles.
These facts are pinned at the top of every chat prompt so the AI always knows who it's talking to — even if ingested conversations mention other people by name.
Use this to clarify names the AI might confuse you with — e.g. "Akash is my colleague, not me."
Search for facts that were incorrectly attributed to another person's name and archive them in bulk. This is useful if ChatGPT history or chat logs contain conversations about other people.
Archive hides from AI but keeps for audit · Delete removes permanently
Replace every occurrence of an entity name in fact text and metadata. Use this when the AI stored facts under the wrong name (e.g. "Lee" instead of "Vish").
🔍
No facts found
No questions yet.
Questions are extracted from your chats automatically, or you can generate them from your existing knowledge.
Facts extracted from your answer
No new facts extracted (answer may already be known)
mp3, wav, m4a, webm, mp4… · no size limit
⚠ Error
✓ Audio safely saved on server:
The file was fully received before the error. Retry freely — it won't re-upload from scratch.
Upload did not complete — audio was not saved. Check your connection and try again.
large-v3=best quality, large-v3-turbo=fast+accurate, tiny=fastest
✓ Done
🎧
Drop an audio file, then transcribe
Step 1 → review transcript → Step 2 → review facts → Step 3 save
You can navigate away — the server keeps running
Edit below to fix names or errors before extracting facts.
✓ Facts saved to memory
Identify each speaker. Click "That's me" for your voice.
Analyzing meeting transcript...
One place to configure the model and LM Studio URL. Applies to chat, fact extraction, brain dump, document import, audio ingestion, and bulk processing.
Must match the identifier LM Studio shows for the loaded model.
Higher = more memory context but slower. Default: 15.
Appended after the memory context. Leave blank to use the default behaviour.
large-v3-turbo = fast + accurate (recommended) · large-v3 = best quality · tiny = fastest
auto = first available GPU · pick a specific GPU or force CPU
TickTick Open API bearer token — enables task creation, search, and completion via chat.
Perplexity Sonar API key — enables AI-powered deep research with citations via chat.
Save settings first, then click Connect to authorise with Microsoft.
Save settings first, then click Connect to authorise with Google.
From Google Cloud Console → APIs & Services → Credentials → API Keys. Enable Places API (New) first.
Slack User Token (xoxp-…) — messages are sent as you, not a bot. Get it from api.slack.com/apps → OAuth & Permissions → User Token Scopes. Required scopes: channels:read, chat:write, channels:history, groups:read, groups:history, im:read, im:history, users:read.
Enables hands-free calling from voice chat. The provider calls your phone first, then bridges to the contact. Your verified number shows as caller ID. Use tailscale funnel 7860 for the webhook URL.
Say these phrases during a voice call to pause/resume AI processing. Must be at the start of your utterance.
How long to wait after you stop speaking before sending. Lets you pause and think without triggering a response. Set to 0 for instant (old behavior).
Required for Apple/Safari push. Use mailto:you@example.com — any real email works.
The AI monitors your location, device state, and health data to proactively check in. Heartbeats are sent every N seconds with context; the AI thinks every few minutes.
Record a voice note from anywhere — Action Button, Lock Screen widget, or "Hey Siri" — and it's automatically transcribed, saved to memory, and turned into a TickTick task if it sounds like a reminder.
Tap either button on the iPhone. Your server URL + key are pre-filled — nothing to type.
The server runs on http:// — not https. If your Shortcut shows a "certificate invalid" error, the URL you entered starts with https://. Change it to .
1. Shortcuts app → + New → Record Audio (stop: On Tap)
2. Get Contents of URL — POST to the endpoint above
Body: Form · field file = [Recorded Audio]
Header: X-Api-Key = [your key]
3. Get Dictionary Value key summary
4. Speak Text → [dictionary value]
Action button: iOS Settings → Action Button → Shortcut → Quick Note
Full-system backups include both vector stores, chat sessions, settings, and profile. Automatic hourly backups are enabled by default.
This will replace ALL current data. A safety backup will be created first.
Notifications are blocked in your browser. Go to browser site settings and allow notifications for this page, then reload.
No backups yet. Create one or wait for the first automatic backup.
The assistant automatically receives your current location. Define named places (office, home, etc.) so it knows the context — useful for reminders and office-mode awareness. Coordinates come from your browser; address lookup uses OpenStreetMap (no account needed).
Sessions idle for 5+ minutes are automatically compacted into structured digests and injected into the AI's context — so it knows what you discussed recently. Long sessions use incremental updates (only new turns are processed).