Key Takeaways

  • Quinton Farmer raised a $30 million Series B for Tolen from Keith Rabois at Khosla Ventures and Mike Mcdano at Union Square Ventures to build personal AI companions.
  • Model intelligence has turned into an accessible baseline; Farmer directs the bulk of Tolen's engineering effort toward custom memory architectures developed alongside OpenAI.
  • Companion recall requires deliberate pruning systems, because deciding which user interactions to erase prevents context rot and conversational creepiness.
  • Tolen enforces a strict rule against unsolicited proactivity: an agent may only initiate prompts if the topic is downstream of something the user brought up first.

Superintelligence Turned Out to Be the Easy Part

Three years ago, most software builders assumed that generating raw cognitive power would take a decade while personal connection would be simple. Farmer found the exact opposite while building Tolen. Foundational models can now solve hard reasoning tasks out of the box, but turning that baseline capability into a companion that fits into a user's private life is where consumer software currently struggles.

As Farmer explained: “three years ago we started thinking about personal super intelligence, and I don't think we particularly expected that the super intelligence would end up being the easy part. And the personal and sort of how to connect to you and really understand you and see you. That proves to be much trickier.”

Tolen spends the bulk of its internal engineering resources on memory systems rather than model fine-tuning. Working with OpenAI, Farmer's team treated human recall as an editorial problem rather than raw storage. If a personal agent logs every casual tangent, shopping list item, and bad mood, its context window becomes noisy and its personality feels robotic. True recall requires clearing outdated states, dropping irrelevant noise, and surfacing old details only when they match the emotional context of a new conversation.

Never Speak Until the User Mentions It First

Consumer AI products often self-destruct by trying to be helpful too early. When an agent interrupts your day with unsolicited advice based on third-party data or aggressive background parsing, users feel surveilled instead of supported. Farmer protects user trust by placing strict technical constraints around proactive reachouts.

Farmer outlined the operating boundary directly: “when we think about how we want to sort of proactively reach out to you, it should always be downstream of something that you have first brought to your tooling, or some inference we've made from conversation.”

This constraint shifts the companion's posture from an invasive monitor to an attentive listener. If a user complains about terrible sleep on Tuesday, the companion can check in on Wednesday. But if the agent discovers a calendar conflict or reads external device telemetry without prior user permission, it stays silent. Farmer pairs this behavioral guardrail with approachable, non-threatening character design. Making the assistant look like a humble, simple helper rather than an omniscient entity disarms natural consumer suspicion, giving the system room to build credibility over months.

What to Do With This

Audit your agent's notification triggers tomorrow morning. Delete any background prompt that initiates a topic the user did not manually introduce during their previous three sessions, and add an automated expiration date to conversational memories so your system clears temporary context after seven days.