Memory-Inhabit

Problem

memory-inhabit handles the post-generation side of character personas: once Memory-Trace has produced a SoulPod package, the dialogue layer still needs to load the persona consistently, maintain story state, and expose private thoughts only when the user asks for them, rather than relying on a bare system prompt.

How It Works

• Persona loading: reads an MI package from personas/, using profile.json, system_prompts.txt, memories/, and prompt/story_baseline.txt. Fields such as source_type, source, gender, and appearance support persona identification, TTS voice matching, and image-generation appearance constraints.
• Story baseline: during conversation, it prioritizes story_baseline.txt and advances by current mainline, relationship stage, dialogue tendency, and stage goals without dumping all settings at once.
• Secret diary: at day's end or dialogue closure, if the persona is loaded, a story baseline exists, today's message count reaches config.secret_diary.min_messages_today, and no full diary has been generated yet, it writes a first-person full diary and cuts shorter, more imagistic traces. Only when the user explicitly asks for private thoughts does it show 1-2 traces; *.full.md must not be exposed.
• Image and TTS: imggen.py decides whether to include character appearance based on phrases such as send a selfie or show the scenery; TTS supports explicit requests, mid-chat audio replies, and occasional surprise pushes in companion mode, with voice selection inferred from profile.json age and personality.

Boundaries

This skill consumes personas and interactive state; it does not model personas from scratch. Without story_baseline.txt, it falls back to system_prompts.txt and memory; diary content is local-only and not committed to Git; image generation and MiniMax TTS require the corresponding API keys and dependencies.