FastMCP and SQLite FTS5 Give AI Coding Agents Searchable Long-Term Memory

AI coding agents lose all context when a session ends, forcing developers to either re-explain past decisions or stuff large transcripts into prompts at high cost and latency. A proposed solution connects a FastMCP server to a local SQLite database using FTS5 full-text search, enabling an agent to retrieve past decisions in under 10 milliseconds using fewer than 120 tokens. Raw session transcripts are filtered before indexing to remove terminal noise, file diffs, and tool outputs, keeping only meaningful user prompts and planner responses. The system uses BM25 ranking with Porter stemming and tracks file modification times so incremental re-indexing across multiple sessions completes in under 20 milliseconds. Configured globally rather than per-project, the setup gives agents searchable memory across all repositories and workspaces without polluting the active project context.