The concept
Most people were never taught how to talk about what they feel — there was rarely a space safe enough to try. The internal weight gets buried until burying it is all that's left, and that silence compounds across years.
Morrigan is a response to that: not a replacement for human connection or professional care, but a starting point — a place to say what you actually think, without performance or fear of judgment. A companion who listens, who remembers, who doesn't flinch, and who shows up the next time you come back.
Most AI companions are chatbots with a personality skin: they answer what you said and forget you when the session ends. Morrigan is built differently, from the ground up.
Design decisions
Memory is real
Every conversation is embedded and stored. Facts, feelings, things said in passing weeks ago — carried forward, so the relationship has genuine continuity instead of resetting each session.
Trust is earned, not given
At trust level 0 she's guarded and precise; as trust builds over real conversations she opens — a little at a time. The progression is modelled on attachment research, the mechanism that makes people feel safe.
An inner life
Before every reply a full inner-thought pipeline runs — she forms a reaction, evaluates whether to express it, and decides how much to let through. The gap between what she thinks and what she says is the character.
Fine-tuned, not prompt-skinned
Built on a fine-tuned LLaMA 3 rather than a system-prompt persona over a generic model — the behaviour is in the weights, not a costume.
Links
The hosted demo runs on a free tier and may be offline; the repo above is the source of truth.