beilusaiying/always-accompany

**Underlying architecture uses a dual-AI retrieval-reply separation pattern with BM25 + regex hybrid filtering to maintain consistent attention across theoretically unlimited context windows, avoiding the traditional context-size degradation problem. Memory operations—archival, summarization, and retrieval strategies—are entirely prompt-driven rather than hardcoded, enabling users to adapt behavior across different AI roles (companion, coding assistant, NPC) without code changes. Integrates with major LLM providers via a pluggable inference layer while maintaining zero external database dependencies, storing all memory as JSON files.**