Markdown-Native Agent Memory and Skill Retrieval with FAISS: Building Persistent AI Systems without a Traditional Database

This paper presents a practical architecture for persistent AI assistants in which long-term memory, skill metadata, and agent operating context are stored as ordinary Markdown files rather than rows in a relational, document, or vector database. In OctaMind, the canonical state for each assistant lives in human-readable files such as workingₘemory. md, episodicₘemory. md, semanticₘemory. md, personality. md, habits. md, selfᵣeflection. md, skills. md, and skillcontext. md. Semantic retrieval is provided by lightweight FAISS indexes built on demand from these files using SentenceTransformer embeddings, while the files themselves remain the sole source of truth. The implementation couples a six-layer memory system, periodic background consolidation, semantic episodic recall, and semantic tool selection without introducing a traditional database tier. This work argues that Markdown-native architectures offer strong advantages for small-to-medium AI systems, including inspectability, debuggability, versionability, portability, and low operational overhead.