Seven-layer memory architecture for a local AI agent: lessons from building it

The author describes building a seven-layer memory system for a local AI agent called Hermes Sterling, aimed at retaining facts and decisions without loading the full history into every prompt. The design separates always-available Markdown files from on-demand retrieval using SQLite, FTS5, local embeddings and related components. The team framed memory as five distinct problems: persistence, retrieval, provenance, cost and integrity. Its most useful findings came from negative results, including a metric that was mathematically impossible and a timeout shorter than the model's cold load. The article is explicitly a single-environment case study, not a general framework comparison.