agent-memory skips vector DB stack

This is a sharp rebuttal to the reflexive “add a vector DB” pattern in agent design, and it feels well matched to personal-scale agent memory. The tradeoff is that simplicity holds only as long as the model can still reliably discriminate what to load as memory volume and ambiguity grow.

• –The repo’s core idea is that agents already know how to read YAML-frontmatter skill files, so retrieval can ride on native text comprehension instead of an external search layer
• –Its 3+1 memory split—profile, procedures, directives, and episodes—makes the memory model legible for humans instead of burying behavior inside opaque retrieval code
• –Claude Code support via SessionEnd hooks is practical, because it gives users episodic logging without needing a separate database service
• –For small and medium memory scopes, this could be more robust than a fragile RAG stack; for larger scopes, token pressure, noisy recall, and attention degradation are the obvious limits