Bring your own AI brain. Gemini, Claude, ChatGPT, your choice. Mandaire is the memory any brain calls: contacts resolved across every source (iMessage, WhatsApp, Apple Notes, Gmail, and every conversation you have had with ChatGPT, Claude, and Gemini), commitments tracked across years, inferences scored by evidence, disclosure rules enforced before any query reaches your AI. On a server you control. Every AI you already use queries it via MCP.
We reply within a few days. No drip sequence.
The problem
You mention Jordan and your AI asks who that is. You told it last week. You switch from ChatGPT to Claude and start over. This is not a memory problem; it is a knowledge graph problem, and the difference is structural.
LLM memory accumulates references without resolving them. "Jordan Chen," "J. [email protected]," and the person in 34 email 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. It also weights a single message the same as a three-year relationship: it cannot score an inference by evidence, or surface a stale belief, or say "this commitment is eight weeks overdue." And it cannot enforce deterministic rules: "always surface what I owe before a meeting with this person" requires a rules engine upstream of any model, not a prompt.
These are architectural failures, not scale failures. A million-token context window still fails on all three. Mandaire solves them: every source resolved into one structured record, every contact understood: who they are, what they care about, what your relationship with them is, what you should share, and which claims are reliable enough to believe. Every query gated by deterministic disclosure rules you set. Switch AI providers whenever. The graph stays yours.
What this looks like in practice
Before a meeting, a call, or a message you need to get right: this is the primary use case. Two minutes of context from every channel you use, already resolved before you ask.
Pre-meeting brief
You ask your AI
"I have two minutes before my 2pm with Maya Patel. What do I need to know?"
Your AI answers
Last contact: a WhatsApp thread three days ago; she flagged a deliverable outstanding from your March call. You said you would send it by Friday. You did not. Before that: an April email exchange where she agreed to intro you to Sam Nakamura at her firm. That intro has not happened. Active commitments: two open, both on your side.
This is not a search result. Mandaire already knew the WhatsApp thread and the April email were about the same relationship and the same outstanding items. It had resolved the cross-channel context before you asked.
Judgment, not just retrieval
You ask your AI
"I have a call with Alex Rivera at 3pm; he's been trying to move the Meridian partnership forward. Brief me."
Your AI answers, including what was withheld and why
Alex Rivera, Director of Strategy, Meridian Capital. 23 interactions across Gmail and iMessage since November 2021.
Where you left it: Your March 15 call ended with a commitment to send revised term sheet language by end of month. You sent a draft April 2nd; he replied positively but flagged one clause. That clause is unresolved. He followed up twice: iMessage April 18, email May 1. No reply from you to either.
Withheld from this brief: a recent signal suggesting Meridian may be evaluating a competing product. You haven't confirmed this with Alex. I'm leaving it out because it changes how you should listen, not what you should say. Available if you want it.
Sources: Gmail (11 threads), iMessage (12 conversations). The "what I'm leaving out" answer is not a feature. It is the system deciding what is relevant to this conversation, for this recipient, in this context. A search tool finds everything. Mandaire decides what to surface.
Before you connect anything
Your data lives on a dedicated server, not shared with any other user, never used to train any model. Privacy here is calibrated and demonstrable: what each AI sees is scoped by per-person, per-topic disclosure rules, verifiable from the audit log.
Exportable in full at any time. Deleted on request if you stop. The Privacy page explains what metadata we can see and what is passed to Claude, ChatGPT, or Gemini when you ask a question.
What Mandaire does
Gmail, iMessage, WhatsApp, Calendar, photos, AI conversations. Statistical entity resolution joins the same person across every source automatically, not a list of references; a single structured record. Who has gone quiet. What you last committed to. Already resolved before you ask.
Via MCP, structured answers from the knowledge graph are available to Claude, ChatGPT, Gemini, and whatever ships next year. You are not locked to any provider. The graph lives outside all of them. Switch tools and bring your history with you.
After every AI session, your AI writes decisions, commitments, and confirmed facts back to Mandaire. The graph also ingests your AI conversations automatically. Every future query starts richer, with calibrated inferences, not just accumulated text.
A deterministic disclosure engine runs upstream of every query. Per-recipient, per-topic, per-context rules decide what each AI gets. The same query returns a different answer for a different recipient. ChatGPT, Claude, and Gemini cannot do this. The disclosure primitive does not exist in generative output. Enforced by code, not by a model that might drift.
On your dedicated server, not a shared cloud. Exportable in full at any time. Deleted on request, not archived. Portable across providers; the graph is yours, not ours.
Who it is for
If you brief yourself before every call. If you use Claude or ChatGPT most days and keep re-explaining who people are and where things stand. If your network intelligence is a professional asset. This is built for you first.
The first users are people for whom one missed context point costs real money or a real relationship. Mandaire is not for casual AI users. It compounds with depth of use and depth of relationships. The more of your life it can read, the more useful it gets.
Getting started
Your AI connects to Mandaire once and can read from it whenever it needs context. No data import required to start. Connect your sources when you are ready: Gmail, iMessage, WhatsApp, Calendar. The further back your archive goes, the more it knows from the start.
Invited by someone already on Mandaire? When you sign up, your AI can query their permissioned context immediately. No server setup on your side, no data to import.
The 7-day context test
Day 0: connect your sources. Your AI starts reading from them. Week one, Mandaire knows what you have told it. Month three, it knows the patterns you have not noticed: commitments tracked against what actually happened, stale inferences corrected by what you have said since, contacts resolved across sources you had not thought to connect. By the end of the first week, it should surface at least one thing you would otherwise have missed. Day 7: you decide if it is worth it. Most people stay.
That is the promise we are built on. Not seven years of use. Seven days.
Go deeper
Three pages worth reading before granting access to your full archive.
Proof
Not promise
Real artifacts from daily use. The daily brief, the drafts, the pre-meeting brief, the decision ledger. Verbatim. Names redacted.
Privacy
Private by design.
What data lives where, what we can and cannot see, what is passed to Claude, ChatGPT, or Gemini. Which sources are live and which are next.
Beta plan
90 days, committed
What we will prove with you. What counts as success. What counts as failure. Named, not hedged.
The underlying architecture and the bets behind this product are at mandaire.org.
Your data lives on a dedicated server, private to you. We do not sell it, we do not train on it, and we are building toward an architecture where access is cryptographically controlled by you, not us.
Private beta, built with AI-native founders and operators. We reply within a few days and confirm the setup works for you before either of us commits to anything.