Teach one robot. Every robot learns. Undo it in one click.

A robot fleet can learn from people safely when every lesson carries its teacher, passes each home's own rules before it spreads, and can be traced and taken back across every home at once. One Lesson is that memory layer, built on MongoDB Atlas, with ElevenLabs as its ears and voices.

Built today, 14 Nov 2026, at Agents in Context Toronto. First commit 10:06 AM.

Two kitchens side by side: Home 2's robot carries the knife down to the low drawer, Home 5, a kids' home, carries it up to the high cabinet — Same words, "tidy the kitchen": Home 2 puts the knife **LOW**, Home 5 has kids and puts it *HIGH*.

The 60-second demo

Recorded from a real run against the event's Atlas sandbox. Robot voices, the spoken lesson and the live Home 5 agent are ElevenLabs.

What the memory does

Teach

Say a lesson out loud. ElevenLabs Scribe hears it, a parser in code turns it into steps, and change streams carry it to six homes in under a second.

change streams · resume tokens

Refuse

Homes with kids refuse any lesson that puts a sharp item within a child's reach, and say why, out loud. The rule is code, not a model.

deterministic tags · per-home rules

Find

Ask any robot in English, French, Spanish or Mandarin. Atlas embeds every lesson itself and reranks, then the home's rules decide what it may use.

Automated Embedding · $vectorSearch · $rerank

Trace

"Who taught you that?" The robot answers from its lesson's family tree and the event log: teacher, parent lesson, homes that took it or refused it.

$graphLookup

Undo

One click rolls a bad lesson back, and everything built on it, in one transaction with a receipt. "Forget me" erases a teacher's identity everywhere.

transactions · receipts · TTL · $jsonSchema

Talk to it

Home 5 is a live ElevenLabs agent. Its only tools call our rule check, the home and the lineage. It has no tool to teach, roll back or forget.

ElevenLabs Agents · client tools

What the judges see on screen

ROLLED BACK banner naming the lessons, 4 homes back on the safe version, measured milliseconds and the receipt — The undo: what was rolled back, how many homes, the measured milliseconds, the receipt.

Talking to Home 5: the ElevenLabs agent calls ask_home and answers aloud — Talking to Home 5: every tool call is on screen, so you can see the agent never decides safety itself.

Built with

Change streams + resume tokensAutomated Embedding (voyage-4)$vectorSearch$rerank (rerank-3)$graphLookupTransactionsTTL index$jsonSchema ElevenLabs ScribeSix ElevenLabs voicesElevenLabs Agents

Measured today

0.435 → 0.639"put the dishes away": best match before the lesson exists, then after (floor 0.60)

382 · 523 msa lesson reaching all six homes, two saved runs, Toronto to us-west-1

22 / 22demo checks pass against the sandbox; 14 unit tests, safety guards proven by breaking them

4 / 4live agent round trips answered aloud; unsafe instructions refused by the rule check

Honest limits: the fleet is simulated, and One Lesson governs which skills a fleet shares, not how motors move. Six homes on one laptop server is what we measured. Teacher names are typed and marked unverified in this build; Auth0 is in the code but off.