A personal model for every user

The new wave of personal agents all share the same parts: persistent memory, connected tools, background work, a personality file that describes you to a general model. Wrapping an agent harness in a consumer interface doesn't make AI personalized. Providing you the means to fine tune a model of yourself does. A personality file can describe a person to a generalized frontier model. A heartbeat can keep the agent working between requests. Retrieval can return the relevant history. None of these operations modify model weights or better align a model's reasoning with the person whose life supplies its context.

When I started at Sentience, we approached "write like me" and "think like me" as prompt-engineering and retrieval problems too. But the description of a voice is not the learning of that voice. A style guide may inventory my habits without learning the relations between them, or the intentions that underlie them: the compression of a casual invitation, the gathering force of a technical disagreement, the meandering cadence which imbues an apparently ordinary sentence with an ethos that is unmistakably mine. What we are demonstrating moves personalization into learned parameters: a personal model post-trained on my writing, coupled to the memory and preference harness developed through my use of Sentience.

Each user gets low-rank adapters on a frozen base model (Hu et al., 2021), continued from a shared voice model trained on an opt-in cohort. Training pairs map a situation and an intent to the person's real reply, with loss on the reply only. A preference stage (Rafailov et al., 2023) refines it further. We evaluate voice and meaning separately on held-out replies, because sounding like you is worthless if the reply changes what you meant. An Individual Turing Test built on one volunteer's decade of messages found that LoRA fine-tuning did better on conversational style, retrieval and memory did better on personal opinions, and hybrids beat either alone (Guo et al., 2026). Personalizing in context is the industry default (Zhang et al., 2025). Training a model for each person is not.

Model sovereignty

At Sentience we propose something different: model sovereignty. What if the model and memory built from your use of Sentience belonged to you? We host your model, harness, and memory as a custodian providing a service, not an owner. The model trained on your writing is yours. You shouldn't have to rent it back.

The goal is that just using Sentience trains your model. With explicit, revocable permission, your writing, choices and corrections can furnish examples for further training.

Voice across registers

A person's voice isn't a single register waiting to be bottled. I write differently to a colleague, a collaborator, a friend. Even with the same friend I write differently if I'm excited or depressed. An invitation to collaborate cannot become a commitment you didn't ask for merely because "yes and" are the next most probable tokens out of the decoder. Encoding judgement and relational awareness is essential in order to nail the nuanced pragmatics of personalized expression.

The game

The first personal model we're publishing is mine. Below is a little thought experiment in the form of a game: my real writing next to my Sentience, ChatGPT and Claude. Five rounds. A text, a Slack reply, an email, a dream, an opinion. Read them and guess who wrote what.

Loading a recorded round…