For centuries, a Jewish name was more than an identifier. It was a portable memory. It spoke of origin—Toledo, Worms, Baghdad, Salonika—of function—Cohen, Levi, Sofer, Shamash—of trade, sometimes even of the river where a grandfather was born.
When a name shifts—Cohen becomes Kahn, then Kogan, then Cahen—it is not a betrayal; it is the trace of a displacement, of a border crossed, of an Ottoman or Russian scribe who transcribed according to his own rules. Each variant is a milestone on the journey.
Our work begins here: to refuse that these deformations be forgotten. To render them legible as one renders legible the effaced letters on a stele.לְהַשִׁיב לָאוֹתִיּוֹת אֶת קְרִיאָתָן
Five centuries of exile left archives in tatters. The pinkasim of Vilna burned in 1941. The ketubot of Salonika survived in cellars. Ottoman registers were dispersed among four successor states. French colonial records speak of 'Israelites' without noting Hebrew. Jewish names in them are sometimes mistreated, sometimes ignored, sometimes piously preserved by hands that were not Jewish.
To work from these archives is to accept that they say also the gazes that produced them. Our method does not claim to erase these gazes. It makes them explicit, confronts them, and sometimes undoes them.
zakhor.ai uses learning models to cross-reference, not to fabricate. When the engine proposes a missing link—an inferred mother, a probable brother—it does not create a person; it identifies a plausible configuration from real items, and marks it as such.
We hold to the strict distinction between attestation, inference, and hypothesis. Three statuses, three colors, three levels of proof. No link is silently promoted. No ancestor is added without being able to trace back to the chain of reasoning that made it appear.
Better an incomplete and true lineage than a full and fictional one.טוֹב יִחוּס חָסֵר וֶאֱמֶת מִיִּחוּס מָלֵא וְשֶׁקֶר
zakhor.ai reveals no kinship without explicit consent from both parties. A family separated by three centuries of exile will not be reunited by surprise in a window; it will be reunited, if it wishes, after seeing the degree, the common lineage, and having chosen to be revealed.
This rule is not a legal convenience. It is a principle: no one has to carry the memory of an entire diaspora without knowing it. Consent, here, is the first form of respect.
Each year, zakhor.ai publishes a public report of errors. How many inferences proved false, in which communities, in which eras. No memory tool is built without errors; an honest memory tool is one that exposes them and learns from them.
If you find an error in a lineage—a date, a place, a parent—you can contest it. The engine recalculates, the scientific committee arbitrates disputed cases, and the correction leaves a trace: we will always know what was known, and when.
zakhor.ai is not a private service. It is a collective instrument. The names it indexes belong to the families who bear them and to the communities who preserved them. The algorithms it uses are auditable. The sources it cites remain the property of their keepers.
What we are building is, ultimately, what our grandparents would have wanted to be able to build: the possibility, for anyone who wishes, to trace back the thread of their name—even if it means discovering, sometimes, that a strand is missing, and that is precisely what needed to be learned.
"Eleh toldot"—"these are the generations." Genesis repeats this formula eleven times. Eleven new beginnings. At each generation, the lineage renews itself.בְּכָל דּוֹר וָדוֹר חַיָּב אָדָם לִרְאוֹת אֶת עַצְמוֹ...
This manifesto is open. If you share these principles—the name as witness, the archive as gaze, AI in service of proof and not fable, the decision left to the living—you can attach your name to it and carry it with us.
