Core concepts
The SHV stack is built around a few simple ideas. They're not original to us, but they keep us honest when systems get complicated.
Brains, not endpoints
We treat Eunoia as a kind of brain, not a single product. It powers different surfaces: research tools, internal agents, mental health systems, and more. A chat interface is just one thin layer on top.
That means we design everything around sessions and agents, not around "send one prompt, get one reply".
Agents
An agent is a structured bundle of:
- a base model (Eunoia or something else),
- a tool set (browser, code runner, internal APIs),
- a memory configuration,
- a safety profile.
You can think of it as: model + tools + memory + rules.
Sessions
A session is a long-lived interaction between an agent and a user (or system). It keeps track of:
- conversation history,
- tool calls and results,
- emotional signals and risk scores (from Velin),
- state that matters to the problem at hand.
Sessions are where the work happens. The same agent can be involved in many sessions with different users or tasks.
Emotional layer vs. pretence
Velin is an emotional layer around the base model. It produces signals (mood, risk level, attachment strength), not role-play. We care about:
- spotting when users are fragile,
- slowing down or redirecting the model when needed,
- avoiding manipulative behaviour as a feature.
Plasticity and adaptation
Models change over time. We want that change to be:
- slow and controlled, not chaotic,
- traceable — you can see what update did what,
- reversible — you can roll back when an idea fails.
MG-SGLoRA is our current attempt to get there: sparse, targeted updates instead of retraining whole brains for every new idea.
Honest surface area
We try to be explicit about what's real:
- If something is internal-only, we say so instead of implying it's public.
- If a behaviour is brittle, we write down the failure modes instead of pretending it's solved.
- If we don't know yet, we label it as a question, not a feature.