Summary

Claude Code's default memory is inadequate for long-term project management. By integrating the storage habits of Memarch with the targeted injection strategy of Hermes, you can build a superior, tiered memory system that reliably recalls context across months of work without bloated token costs.

Key Topics

• Vector Embeddings: Vector embeddings represent text as semantic points in high-dimensional space. In this context, it allows Claude to find relevant historical information even if the current query uses different vocabulary than the initial note.
• Context Injection: Instead of relying on the AI to remember everything, injection provides a consistent, curated starting point. The host argues for injecting a 'frozen snapshot' of critical project files to keep the agent grounded.
• Tiered Retrieval: This method optimizes efficiency and performance. By checking active session memory first, then project-level markdown, and finally the deep vector database, the agent avoids unnecessary compute costs.

Key Takeaways

• Implement the Memarch-style stop hook to log raw transcripts.
• Review the 'plan.md' free setup document.