Implemented scaffolding in the conversation layer.
We introduced a director-style scaffold that structures every turn into Detect → Intent → Act → Evaluate → Correct. Instead of treating replies as one-off generations, the scaffold gives each message a job: analyze the user’s line, decide the character’s intent, perform the response, sanity-check canon/consistency, and course-correct if needed. It keeps the flow modular and debuggable, so we can iterate on any stage without breaking the rest.

Pattern recognition, topic tracking, pivoting, and memory summarization.
On top of the scaffold, we wired in topic tracking and a lightweight pattern detector that spots repeats, pivots, and contradictions. A compact memory summarizer distills user-only snippets into crisp facts, so the character can recall what matters (“birthday is Jan 12”) without dragging in noise. Together they keep the conversation anchored—when the user changes lanes, we notice; when they circle back, we don’t start from zero.

Context-aware prompt curation.
We redesigned the prompt the character sees so it always reflects the live context: the current topic summary, any pinned memories worth recalling, and a short “director’s note” that guides how to speak (not a script to read). The result is tighter outputs, less rambling, and responses that feel both in-character and situationally aware.

Tiny LLM helpers for speed and structure.
To keep latency down, we added small JSON-only helpers that label patterns and summarize memory fast. They’re cache-friendly, deterministic, and cheap—perfect for the prework phase—so the heavy model can focus on the final line while the helpers keep state and structure clean.

JSON profile for Seraphine and Alaric.
Seraphine and Alaric now are included with a structured profile (situation, relations, traits, and speech style) that acts as their canon backbone. The conversation layer pulls from this profile to steer tone, word choice, and boundaries, so they stays recognizably herself across scenes and topics.