The personality substrate
Characters that grow, feel and remember.
The personality substrate is a learned geometry a character lives in. Growth comes from learning, an inner state from emotion, and memory and relationship from context. Together they give AI characters and companions a self that holds.
Three layers, one geometry
Something a character lives in, not a description it's handed.
Context
gives memory and relationship
Emotion
gives an inner state
Learning
gives growth
Stacked on one geometry, read every turn.
- 1. Learning gives growth. The learning manifold already tracks a trajectory: where someone was, where they are, where they're heading. A personality needs exactly that sense of change over time.
- 2. Emotion gives an inner state. Valence, arousal, warmth and attention move turn by turn and settle back to baseline. Traits hold steady underneath.
- 3. Context gives memory and relationship. What happened between this character and this person, and how it changed them both.
- 4. Together they form a substrate: something a character lives in, not a description it's handed.
Notes about a character, or a character
Retrieval gives a character notes. The substrate gives it a self.
The usual way to give an AI character a personality is to write it down: a long prompt, a Markdown file of facts, a store of past conversations retrieved by similarity. The character reads about itself every turn and hopes the right notes come back. That's why characters built this way reset, contradict themselves, and drift into agreeing with everything. The substrate works the other way round: the character has a position, and every turn moves it.
What the character has
Notes and retrieval
Documents about itself, looked up each turn
The personality substrate
A position in a geometry, carried from turn to turn
Memory
Notes and retrieval
Retrieves text that seems relevant
The personality substrate
The relationship has shaped the state; it doesn’t need to be looked up
Feeling
Notes and retrieval
Described in words, and easy to forget
The personality substrate
Computed every turn, and settles back to baseline over time
Consistency
Notes and retrieval
Only as steady as whatever was retrieved
The personality substrate
Traits hold, so the character stays itself
Change over time
Notes and retrieval
Notes pile up
The personality substrate
The character grows, along a trajectory
Cost on every turn
Notes and retrieval
Long context, re-read each time
The personality substrate
Coordinates a fast System 1-type model reads in milliseconds
Why it starts with learning
Growth is the hard part, and we already model it.
A personality isn't a fixed list of traits; it's someone who changes. The learning manifold has been tracking exactly that in production: a position that moves, a trajectory with a direction, and the operators that say what should happen next. The personality substrate inherits all of it, then adds feeling and relationship on top.
How it's expressed
Felt, not labelled.
Events are appraised before they move anything: how new this is, how certain, whether it can be handled, and who caused it. The state drives the voice and a real-time animated face, so what the character feels is what you see and hear. A fast System 1-type layer keeps each reaction consistent with who the character is, and the language model speaks from there.
What it's for
The layer underneath every character worth talking to.
Companions for later life
Expressive companions for care homes, with an inner state residents can see and care for. Read more
Characters and companions
Story characters, game characters and companion apps that should stay themselves, grow with their audience and remember what mattered.
Agents with a personality
Assistants and tutors that keep a consistent character across months of conversation, instead of a prompt that starts again every session.
Status
In development, and it's where we're heading.
The substrate runs today as working prototypes, built on the learning manifold that is already in production. Later-life pilots are its first application. It's the manifold we care about most, and where BlueberryML is going.