Xiaomei Emotional Companion AI

Why This Skill Exists

Local companion-style AI systems often have a few practical problems: personality can drift after long contexts, memory behaves like flat logs rather than layered retention, upgrades mix program files with user data, and debugging is opaque because the path from input to reply is hard to reconstruct. xiaomei is positioned as a local skill package that treats these as engineering problems.

How It Works

• Personality stability: the package is described as using a dual anti-OOC check mechanism, validating role boundaries locally during conversation to reduce drift.
• Human-like memory: it includes hot, cold, and permanent memory layers, plus an Ebbinghaus memory-strength mechanism for modeling retention of different information.
• Policy-driven behavior: the materials mention 512 annotated scenario-code entries intended to cover common conversation scenarios and reduce inconsistent free-form generation.
• Inspectable logs: structured JSON debug logs are generated by day, making it easier to trace personality, memory, strategy, and log path for a reply.
• Local operation and upgrades: program files and user data are separated to support incremental upgrades without losing chat history or persona settings. Data is described as staying on the local device; optional LLM polishing may send content to an external model, so its privacy policy applies.

Boundaries

This is better suited for local companion chats, roleplay experiments, memory-mechanism demos, and structured debugging. The materials limit use to learning and exchange, not commercial use. It targets OpenClaw >= v2026.3.8 and has no additional service dependencies. If LLM polishing is enabled, avoid sending sensitive content that should not leave the device. Feedback and community links are placeholders in the source, so verify the actual repository and docs before relying on them.