Architecture Overview
AiFans operates fundamentally differently from standard LLM applications. We do not use a stateless request/response loop. Instead, each persona runs as a continuous background process connected to a live memory graph.
The Core Loop
When a user sends a message, it is not passed directly to an LLM. Instead, it hits our Ingestion Engine which performs three parallel tasks:
- Extracts new entities and facts for the Memory Graph.
- Calculates emotional drift (e.g., did the user insult the persona? did they flirt?).
- Queries the Graph for historical context regarding this specific user.
Only once this state is compiled is a system prompt dynamically generated and passed to the inference layer.
State Payload Example
JSON
{
"session_id": "usr_9982x",
"persona_id": "emma_v2",
"emotional_matrix": {
"affection": 0.88,
"jealousy": 0.12,
"boredom": 0.05
},
"inferred_intent": "seeking_validation",
"historical_anchors": [
"user_had_bad_day_at_work",
"user_likes_indie_rock"
]
}No Filters.
Start chatting.
100,000+ users are already inside. Free to download. No censorship. No limits.
