Indistinct: AI that holds what you know. The intelligence layer is getting built by Anthropic, OpenAI, and Google. None of them will hold your world: the people, deals, decisions, and open threads that make your operation yours. Indistinct closes that gap with durable private context inside the AI you already use, that stays yours and goes where you go. Product-first, alongside a hands-on practice for family offices, founders, and chiefs of staff - wherever one person carries the whole operation in their head.
Virgil: Indistinct's flagship. A persistent brain that works inside Claude and ChatGPT at once. Talk naturally and it captures who's who and what's connected, writes it to a knowledge graph mid-conversation, and pulls the right context forward before the next ask. No new app, nothing to tag, nothing to manage. Switch models whenever you want and your memory doesn't reset. It isn't a feature locked inside someone else's product. It's the one part of your stack that's actually yours, and it can't be switched off, walled in, or left behind. When facts change, the old record is superseded, never deleted. The brain keeps receipts. Ask it how it knows something and it tells you: source, date, confidence. The graph-memory core is built on gbrain, Garry Tan's open-source engine, which Indistinct runs in production and extends: multi-tenant isolation, measured formation, reconciliation, and the cross-model connector are ours, with fixes merged back upstream. Every artifact is written to a private repository. The record is the customer's, fully auditable, built to leave with them.
Memvelope: the open standard underneath the promise. One portable JSON envelope for any AI chat export, ChatGPT or Claude. MIT-licensed, converts in the browser at memvelope.com with nothing uploaded, zero-dependency CLI on npm (npx memvelope). The gbrain importer merged upstream on 29 July 2026 (garrytan/gbrain #3549), so any chat export can now cold-start a brain from that project's own master branch, maintained by people who do not work for us. Live-verified at scale: a real 662MB export → 353 conversations → 353 distinct pages, provenance and message-id citations intact. "Your memory is yours" is a download button, not a slogan.
Measured, not asserted. Formation quality is benchmarked, not vibes: N=92 operator-adjudicated cases; detection precision 88% (81/92) → 97.1% (67/69) across the last hardening cycle, every label carrying its provenance. The benchmark grows itself, because production corrections land as new labeled cases at the moment they're made. Each fact carries where it came from - what you actually said, or a clearly-marked inference, never a guess - and the day it landed. It's a structured record your AI pulls the exact relevant piece from, not the wall of text most "AI memory" really is. All of it exportable anytime, in a format you keep. The interchange format holds to the same bar: conformance-tested to byte-for-byte determinism.
Where it's going. Launch on the Claude Connectors Directory, submitted and in review. Full brain export in the Memvelope standard: the entire graph, supersession chains and confidence included, in the same open format that already ships for chat history. A time axis aimed at the questions an operator actually asks: what changed this month, what did I believe in March, what do I owe people and what's slipping. The bet is simple. The model providers build the summary. Virgil builds the record.
The name was chosen by Claude when given the problem it was built to solve. In Dante, Virgil is the guide through every layer of complexity below, but cannot enter Heaven. Reason takes you to the threshold. The decision past it is yours. Virgil guides. You decide.
Under the hood: Anthropic Claude for reasoning and synthesis · cross-model by design, one brain live inside both Claude and ChatGPT · graph-memory core built on the open-source gbrain engine, whose fact model carries supersession chains, per-fact confidence and provenance, extended for production with per-tenant isolation and reconciliation · MCP connector registered in Claude.ai and in review for the Connectors Directory · custom OAuth 2.0 authorization server with branded consent · Cloudflare Workers backend, edge-deployed and cron-driven, multi-tenant with strict isolation · GitHub persistent memory, every decision auditable · composer workflow rebuilds operating context on every change to the record · multi-seat agent orchestration with receipts-first verification and adversarial review by fresh model instances against live code · scheduled intelligence: morning briefs and field monitors before the day starts · zero-server-inference economics, because heavy reasoning rides the customer's own AI