Yogev
Ben-Tov
Building the world's most powerful AI Chief of Staff at Team0 — and helping companies transition to an AI-first world.
Felt understanding
An agent should hold a model of the people, projects, and commitments in your life rather than a pile of past messages it can search. That part is roughly solved. The hard part comes after it.
Your agent believes two contradictory things right now. Its record knows you left that job. Its next sentence congratulates you on it. Storing the truth and speaking the truth are separate disciplines, and nearly everyone builds only the first — so an agent states that it sent the email when all it ever had was the intention to send it.
Most systems are tuned for total recall. I think the intelligence is in suppression: knowing a loop is already closed, the thing was already handled, and the right move is to say nothing. Silence is a success state, not a failure.
And once an agent acts in your name, writing to people who never agreed to talk to software, the question stops being what it knows and becomes what it is allowed to do. Someone should be able to get a meeting booked with you without ever seeing your calendar. Trust widens on your evidence, never because the other side asked for it.
The bar I build toward
- Being accurate about someone's world
- Knowing what actually matters to them
- Showing continuity across time
- Being honest about the limits of what it knows
- Knowing when to stay quiet
Writing
On SubstackI write about how to build AI that understands people. New essay every week or two.
- Aug 2026Who is this person, really?Getting a fact right means nothing if it lands on the wrong person. Here is how to decide who a name is about, and when to admit you cannot tell.
- Jul 2026Your agent needs two memories, not oneThe usual advice is to extract the facts and throw the conversation away. Keep both.
- Jul 2026Your agent shouldn't have a memory. It should have a model.The difference between recording what someone said and actually knowing them.