hermes - 💡(How to fix) Fix Feature Request: Semantic / Per-Message Skill Retrieval

Current behavior: Hermes injects the full skill index into every system prompt. With 20+ skills, this costs ~800 tokens per API call. The experimental semantic_retrieval flag exists in the codebase but matches skills only against the first message of a session — making it impractical for long, multi-topic sessions.

Problem: If a session starts with a casual message ("good morning"), the wrong skills get loaded and there's no mid-session reload mechanism. For users with diverse skill sets (fitness tracking, infra, coding, research, etc.), this means either all skills are always loaded, or the wrong ones are.

Proposed solution: Match skills per message against the current user input (not session-start, not SOUL/USER keywords). Simple keyword or embedding-based cosine similarity against skill descriptions would suffice. Load only Top-N matches per turn.

Expected impact:

• ~80% reduction in skill-index tokens (~800 -> ~150 tokens per call)
• More relevant context per message
• Especially valuable for users on cost-sensitive APIs (DeepSeek, OpenRouter)

Setup context: Running Hermes on WSL2 + DeepSeek V4 Flash. Tested the current semantic_retrieval implementation -- deactivated due to the first-message-only limitation.

Happy to test any implementation. Thanks for the great work on Hermes! 🙏