I Ditched Vector Search for My Coding Agent's Memory. FTS5 Won.

Short summary

A developer built a coding agent that searches tool output and logs using SQLite's FTS5 full-text search instead of vector embeddings. FTS5 excels at keyword-dense content where queries are literal (e.g., 'HTTP 500 errors'), not semantic, and costs far less. Teams should evaluate whether their content genuinely needs semantic matching or if BM25 ranking is sufficient before adopting embedding infrastructure.

• •FTS5 is better than vectors for keyword-dense, structured data like logs, stack traces, and JSON
• •BM25 ranking handles agent queries better when both query and content use literal terms
• •Vector search should be reserved for prose with vocabulary mismatch; most agent tooling doesn't need it

Generated with AI, which can make mistakes.