---
post_type: "project"
title: "Serene"
blurb: "A local-first memory for therapists, and a small sealed device that holds it, so the person who carries everyone else does not have to carry it alone."
org: unbody-lab
tags:
    - human-AI interaction
    - adaptive memory
    - local-first
    - clinical AI
    - hardware
    - interaction design
    - trust
year:
    - 2026
current: true
order: 2028
stack:
    - Adapt
    - Local-first AI
    - NVIDIA Jetson
    - Human-AI interaction
---
Serene is a local-first memory for therapists: a desktop app and a sealed appliance that build a living, evolving understanding of each client across years. It exists for the solo therapist carrying a dozen inner lives at once - who said what, what shifted three sessions ago, what not to step on, what still has not been said. The goal is not to make therapy more automated. The goal is to give the therapist their attention back.

The first thing Serene says each day is: "Take a breath. Your clients are here when you are ready." That sentence sets the posture of the whole product. It is not a dashboard asking to be managed. It is a quiet room for someone whose job is to stay regulated while absorbing other people's dysregulation.

Open a client and Serene shows what it understands: the clinical trajectory as a narrative, a map of the people in their life, the patterns they steer around, where they sit in their window of tolerance, and how their personality structure reads over time. The in-app assistant, Seren, lets the therapist ask across a client's whole history and get an answer grounded in what the system has absorbed. A longitudinal report turns that memory into prose: the arc, not just the notes.

The important design problem is trust. Every conclusion has to come with receipts. Serene exposes a "Why & How" layer for each understanding: the specific evidence it used, the gaps it cannot yet support, and a stated confidence. It is also correctable. If the therapist says, "that is a neurotic-level defense, not borderline," the correction changes how future sessions are read. A memory that cannot be inspected is not a memory. It is an opinion wearing one.

Serene also holds a space called Yourself. No one in a therapist's life observes them at work. Supervisors hear about it secondhand. Serene is in the room, so it can notice how the therapist shows up across clients: what stirs them, what they are carrying, where their notes thin out, where certain sessions linger. But it does something most AI systems do not: it holds back.

The governing principle is asymmetric observation: behavior is observable; inner life is the person's to name. On the client side, Serene can help interpret patterns because that is the clinical object of the work. On the therapist side, it reflects only what can be witnessed. It may say, you wrote "couldn't shake it" twice. It will not say, you may be experiencing countertransference. Naming the therapist's inner life stays with the therapist.

That same idea appears in the hardware. Privacy is not treated as a policy page. It is made physical: a sealed appliance on the desk that a client can see. The architecture is local-first, with the product kept in the posture of scribe and memory rather than clinical decision support. The screenshots use seeded demo data - therapist Mira and client Lena Vasquez - but the surfaces are built: the appliance, the calm home, the Client Brain, Why & How, feedback, reports, command bar, and Yourself.

Serene is the clinical instance of the same thesis behind Adapt and my broader work: software should adapt to the person, not the reverse. Here, adaptation is partly memory and partly restraint - knowing when not to make the interpretive move.