A developer has built a memory agent for Obsidian that eliminates the need for vector databases by leveraging large context windows. The system uses SQLite and direct LLM reasoning instead of embeddings, Pinecone, or Redis. Claude Haiku 4.5's 200K context window makes this approach feasible for personal knowledge management.

Traditional memory systems rely on embeddings and similarity search to overcome small context windows, requiring complex infrastructure including Redis, Pinecone, or Chroma plus embedding APIs. This new approach stores structured memory records directly in SQLite and lets the LLM reason over them. The system extracts summaries, entities, topics, and importance scores from raw text, then uses a consolidation agent to find patterns across memories.

By using Haiku 4.5's 200K context window, the system can store approximately 650 structured memories before hitting limits, though practical usage targets around 50 recent memories. The architecture includes three agents: IngestAgent for extracting structured data, ConsolidateAgent for finding cross-cutting connections, and QueryAgent for synthesizing answers. This two-table design with raw memories and consolidation insights provides both factual recall and pattern recognition without the complexity of traditional vector database setups.