Startup Optimizes Model For Phones By Making It Forget Everything Immediately

Neuralis, a three-year-old startup based in San Francisco, announced on Tuesday that its new on-device language model reaches production latency by discarding any information it processes more than sixty seconds after receiving it. The company is marketing the model as an autonomous agent capable of managing calendar invitations, filtering email, and responding to customer support requests, despite the fact that it cannot retain context between separate tasks or recall whether a user prefers their coffee with cream. Neuralis has raised eighteen million dollars from investors including Khosla Ventures and a fund operated by Nvidia's venture arm.

"We think of forgetting as a feature, not a limitation, because technically it is both," said Marcus Chen, the company's chief product officer, during a briefing attended by eight analysts from firms covering the AI infrastructure market. "Traditional agents carry their entire conversation history in memory. Our agent carries nothing. This is a different approach to the problem, and it scales infinitely because there is no state to store."

We think of forgetting as a feature, not a limitation, because technically it is both.

The startup's product sheet, distributed to prospective enterprise customers, describes the model as a stateless agent designed for resource-constrained environments. In practice, this means the system can be deployed on phones, smartwatches, and automotive systems, but each interaction begins from cognitive zero. When a user asks the model to reschedule a meeting, the model cannot remember whether the original meeting was at two o'clock in the afternoon or two o'clock in the morning. When asked a follow-up question about the same meeting, the model treats the follow-up as an unrelated inquiry from a stranger.

Neuralis has positioned this constraint as a competitive advantage in marketing materials sent to potential partners at three major automotive manufacturers and two large telecommunications companies. The company's website states that the model operates with ninety-seven percent less memory overhead than comparable systems, a figure that is mathematically accurate and does not address the question of whether the model can perform the tasks it claims to perform.

Enterprise sales director Keisha Williams, forty-two, based in Austin, Texas, reported during a call with prospective customers that early pilots at a mid-market customer service operation showed the system handling inbound calls at a cost of eighty-four cents per interaction. She did not mention that the same customer reported a forty-three percent increase in repeat calls from users who had to re-explain their issue to the system each time they contacted support, or that the customer eventually requested a refund.

At press time, Neuralis was preparing a Series B fundraising round and had begun describing the model's memory limitations in pitch decks as a feature that enabled responsible AI by preventing the system from forming long-term preferences or goals. The company is targeting a valuation of three hundred and twenty million dollars.