A practical implementation of Google's always-on-memory agent pattern adapted for Obsidian notes, replacing vector databases with SQLite and direct LLM reasoning. Using Claude Haiku 4.5's 250K context window, the system stores structured memory records (~300 tokens each) allowing ~650 memories to fit directly in context. Three agents handle the workflow: IngestAgent extracts structured metadata from raw text, ConsolidateAgent finds cross-cutting patterns across memories, and QueryAgent synthesizes answers from both raw memories and consolidation insights. The system includes file watching for automatic Obsidian vault ingestion, supports text/images/PDFs, and integrates with Claude Code and Kiro-CLI. The core argument is that vector databases are retrieval optimizations for large-scale problems — at personal note scale, large context windows make them unnecessary overhead.
