Your whole life does not fit on a phone. Mandaire is the private server that holds it: it ingests every source, resolves who is who, and hands any AI you use, Claude, ChatGPT, or Gemini, only what it is allowed to see. The full picture can only be assembled server-side. Only Mandaire does it.
The problem
Three structural gaps. Not scale problems that a larger context window fixes.
Your AI accumulates references without resolving them. "Alex," "A. Rivera," and the person in your iMessage threads are the same individual. LLM memory cannot know that without being told, and being told once does not propagate across every source and every session. Statistical entity resolution does, automatically, across every source, continuously. Stateless retrieval has a harder version of this problem: when a name matches a widely-indexed public figure, it anchors on the public record. A personal-contact Jordan Chen with 40 emails in your inbox becomes the public Jordan Chen. A persistent entity model built from your corpus resolves the personal entity first, with source provenance and communication history attached. The distinction holds across every query and every AI you connect.
LLM memory is a derived projection: weights, not rows. It cannot answer "what claims do I hold about this person with low confidence and no cited evidence?" That query is the calibration loop; without it you cannot measure your own false-positive rate, correct what drifted, or know when a belief went stale. Mandaire's inference layer stores every claim as a typed row with evidence, confidence, timestamp, and a supersession chain linking every correction to the prior it replaced. LLM memory remembers. Inference proves.
Your AI cannot enforce deterministic disclosure rules. "Always surface what I owe before a meeting with this person." "Do not pass this topic to that recipient in this context." These require a rules engine that runs upstream of any model, not a model that might get the instruction right most of the time. LLM memory answers what it knows. Judgment enforces what you have decided it may say. The disclosure engine runs before the query reaches your AI: fail-closed, logged, same answer every time for the same recipient, structurally identical response when it holds nothing and when it is holding something it has decided not to share.
The knowledge graph: 360-degree ingestion from your sources plus AI write-back. Statistical entity and relationship resolution across every source. Calibrated inference on what your relationship is with each contact, what context your history reveals, what is reliable enough to act on. Deterministic rules-gated retrieval via MCP. Every AI you already use queries it.
One private data layer. Two ways in.
Both products build on the same three-layer knowledge graph: 360-degree ingestion and AI write-back, statistical entity resolution, and deterministic rules-gated retrieval via MCP. Connect once; every AI you already use reads from it.
mandaire.app
"I have two minutes before my next call. What do I need to know right now?"
Briefs you in 90 seconds before any meeting: what you discussed last time, what you committed to, what the other person is waiting on. Maintained across Gmail, iMessage, WhatsApp, Calendar, and your AI conversations. Your AI drafts in your voice, with the full history already loaded.
Founders, investors, advisors, senior operators. People whose professional value compounds with relationship depth: finance, consulting, law, media. Anyone whose context spans years and a dozen channels, not just this week's emails.
mandaire.app →
mandaire.dev
"I shipped fast and built the wrong thing. I need better judgment before the next build."
Mandaire gives your AI a persistent log of your intent, your decisions, and what you chose not to build. Before any new feature, it reads back what you actually agreed to. Every tradeoff on record, across sessions. The same context layer as .app, applied to building software.
Non-technical founders with a clear vision. Operators without a technical lead. Anyone who has shipped something they had to rip out.
mandaire.dev →
Before you connect anything
Every response Mandaire produces is calibrated for the person asking: who they are, what your relationship warrants, what the context permits, and what serves your interests in this moment. The same question from your investor and your spouse returns structurally different context, not because Mandaire withholds, but because it calculates on your behalf. This is not a privacy constraint bolted on after the fact. It is the product.
Your data lives on a dedicated server, not shared with any other user, never used to train any model. The deterministic disclosure gate runs before any query reaches your AI: scoped by person and topic, with every decision logged to the audit record. The Privacy page explains what runs locally, what metadata we can see, and what is passed to Claude, ChatGPT, or Gemini when you ask a question through MCP.
Common questions
Does Mandaire send my Gmail to OpenAI or Google? No. Your data is stored on your own dedicated Mandaire instance. The AI you use (Claude, ChatGPT, Gemini) receives only the specific context you choose to pass via MCP tools. Nothing is sent in bulk to any provider.
Is my data portable? Yes. Export runs today: request a full export of your graph, inferences, and raw corpus from the privacy request page. We cannot lock you in because the architecture does not depend on us holding your data.
Does Mandaire send emails or take actions on my behalf? No. Mandaire is read-only with respect to your sources: it reads from Gmail, iMessage, and Calendar but never writes to them. Your AI reads context from Mandaire via MCP, then takes actions through action connectors you configure (such as Gmail MCP). Mandaire holds state; your AI acts. Decisions, commitments, and context updates your AI records get written back to Mandaire so every future query is better informed.
What does Mandaire say when it does not have much data on someone? It tells you. Confidence is part of the output: the system surfaces what it knows, how many sources it has seen, and where the gap is. An answer on thin data comes flagged as thin. This is calibration working correctly. A confident-sounding answer on two data points is a different kind of problem.
Architecture and trust model at mandaire.org.
What this produces
Names and employers changed for privacy. The structure is real: relationship depth, open commitments on both sides, what to avoid, resolved before you ask.
Your AI reads what you give it permission to share. The data stays yours. The reasoning model is your choice, and when you switch, the context travels with you.
The frontier AI companies are building reasoning layers. They are not building the knowledge graph that makes those layers useful for your actual history. What none of them can do: ingest every source without hoarding it, resolve the same person across six apps into one record, score and calibrate every inference against the full corpus, or return a different answer to the same question for two different recipients. That last point is not a scale gap. ChatGPT, Claude, and Gemini cannot do it, ever: the disclosure primitive does not exist in the generative paradigm. Mandaire is that graph.
Real-time cross-platform access is becoming a commodity. MCP connectors and native app integrations mean any AI can reach any source on demand. The moat has moved. What no AI can reconstruct: a corpus continuously ingested, resolved, and joined across every source for years; the disclosure norms learned from observing who told what to whom across thousands of actual interactions; deterministic enforcement that holds the same way every time, regardless of what the model is told. An on-demand fetch gives an AI today's content. It cannot give it the resolved arc, the dormancy chapters, the corrections history, the relationship graph built from a full longitudinal record. Identity is the join key. The corpus is the evidence. Disclosure is the export license. Every assistant, Siri included, becomes a client of Mandaire rather than a competitor.
The charter, the architecture, and the six bets at mandaire.org.