Garry 开源了 GBrain，这是一个利用混合搜索和多策略分块技术来提升 AI 代理上下文检索准确性的知识管理系统。

garry open-sourced his own openclaw/hermes knowledge brain. 

example: your agent has 1,000 days of logs. you ask about something from last march. it pulls the wrong thing. gbrain fixes that. plenty of tools do this. gbrain's difference:

three chunking strategies:
→ recursive: fast, deterministic baseline
→ semantic: cosine similarity + Savitzky-Golay smoothing to find topic boundaries
→ llm-guided: haiku judges topic shifts in sliding windows, highest cost, strongest semantic boundary detection

hybrid search via RRF fusion. contract-first tool generation: operations.ts → CLI + MCP + tool definitions, change one place, all agents sync.

GBRAIN_SKILLPACK.md: tells your agent when to read, when to write, how to detect entities, how to back-link. an operating manual for the knowledge layer.

storing 133 days of markdown was never hard. the hard part is when something you wrote on day 87 is exactly what's needed on day 134, and your agent pulls it up on its own.