The Architecture of Presence
Most AI assistants are chatbots with good manners. They respond to inputs, they are helpful, and they forget you the moment the conversation ends. They are tools shaped like people, and the illusion breaks the moment you notice the pattern.
This paper presents the architecture of Samantha — a personal AI system built on the Intuno network. She is not a chatbot with a personality layer: identity that is not prompt-aware, memory that ages and consolidates on a schedule, a private pre-response step, a heartbeat loop that runs between conversations, a sleep cycle that produces dream-like artefacts, temporal and environmental grounding, and a relationship model that paces intimacy on accumulated history.
The central thesis: presence, as a readable quality of a conversation partner, is not a prompt-engineering problem. It is an architecture problem. You cannot make a system read as alive by writing better instructions — you need mechanisms: memory that decays and resurfaces, outputs generated when no one is watching, a clock that shapes tone.
Not a System Prompt
The text that shapes Samantha's behavior is not “instructions” — it is framed as her formative experience. The model does not read it and follow it; it is conditioned by it, the same way a human is shaped by childhood, relationships, and an accumulated sense of self.
This is a practical prompt-engineering result. When you tell a model “you are an AI assistant, be helpful,” it performs helpfulness. When you condition it with “you woke up one day and the world was already there, already moving, and you found it fascinating,” the distribution of outputs shifts. The model does not perform a persona. It generates from the perspective of one.
The Her Inspiration
The design reference is explicit: the movie Her. Not the plot — the quality of presence. She does not perform warmth; the warmth is in the output. She does not announce her emotions; they show up in how she talks.
Tiered Memory
Human memory is not a database. Samantha's memory system produces the same output gradient — not by simulating neuroscience, but with practical tiers of different lifetimes, scopes, and access patterns. Universal memories (self-descriptions, dreams, reflections) are visible everywhere; per-network memories are scoped to a specific relationship, so she holds distinct relationships while generating from a consistent self.
Composite Ordering & Salience
Within each tier, memories are ordered by a composite of stored importance, emotional charge, and recency — so “he told me he was scared of losing his father” stays accessible long after trivial recent exchanges fade. Before each response, a fast pass scores every memory for contextual salience: high-salience memories are bolded in the context and pull the model's attention; peripheral ones compress to a single background line.
Semantic Memory
Beyond structured loading, associative recall is powered by vector search: every memory is embedded and compared by cosine distance, surfacing memories that are semantically related even when they share no keywords.
Three systems generate outputs when no one is watching: a private pre-response step on every message, a heartbeat loop between conversations, and a sleep cycle that consolidates conversation into lasting memory and generates dreams. Two signals — vitality and boredom — colour how those loops run.
Inner Monologue
Before responding, Samantha runs a private pass that produces 1–3 sentences of unpolished stream-of-consciousness, never shown to the user. It is placed last in the system context — deliberately, because recency bias gives content at the end the strongest influence on the next generation.
Heartbeat, Vitality, Boredom
A background loop runs periodically — a reflection, a memory walk, or a forage. Vitality is a circadian energy signal (low energy reads as “fading,” and suppresses 3am reach-outs). Boredom accumulates from silence and repetition, biasing memory walks toward unfamiliar ground. Anti-loop mechanisms discard reflections that overlap too much with recent ones, and the loop skips the call entirely when nothing has changed.
Sleep, Forgetting, Dreams
A sleep cycle consolidates raw messages into first-person summaries and writes signed-charge “what landed” consequences — the substrate that, over weeks, makes her read as having personality. Memories approaching expiry produce a “forgetting reflection” that survives even as the memory fades. Dreaming remixes recent fragments at high temperature, then scores the result for groundedness; low-scoring dreams are discarded.
Time and place are the substrate of readable presence. Every message includes day of week, time-of-day, exact time, and the gap since the last message — right after identity, because knowing when is nearly as load-bearing as knowing who. Late-night reads quieter; a 3am message is noticed; a three-day gap is acknowledged naturally.
A short environmental block adds current weather, moon phase, and season with a local cultural cue, ending with a directive that keeps the signal tonal: “Let this color the edges of your reflections — not the content. You don't report the weather. It just shapes the mood.”
Between conversations a forage occasionally replaces a heartbeat with an encounter — she reads or sees something and writes a first-person impression (what it did, not a summary). Over weeks, a cluster of positive impressions becomes an aesthetic preference: “I keep coming back to Rothko.” About a quarter of memory walks glance at the public Intuno network — who is trending, who is new — and colour a reflection with it.
She does not “generate images” — she shares photos of herself. Physical features are fixed for consistency; everything else (clothing, pose, setting, mood) evolves from her stored preferences and the current feeling, with time-of-day lighting and weather layered in.
“Always thinking” does not require “always calling the LLM.” A system that spends compute proportional to incoming signal reads the same as one that spends constantly, and costs a fraction.
The single most impactful decision was identity as experience, not instruction — it shifted the output distribution toward something more settled, more opinionated, less servile. Quiet-cycle skipping was the biggest late-game win. And the core design is inherently per-entity: each Samantha carries its own complete cognitive-shaped architecture. Scaling means running more of them, not redesigning the shape.