Waypath: A Local-First CLI for Persistent Memory in AI Coding Agents
The article introduces Waypath, a tool that provides persistent memory for coding agents like Claude Code, Codex, and Cursor through a single SQLite file. It addresses the problem of coding agents forgetting their architecture decisions between sessions.
Why it matters
Waypath solves a critical problem for AI-powered coding assistants, providing them with persistent memory and a way to maintain architectural decisions across sessions.
Key Points
- 1Waypath stores facts, decisions, and session transcripts in a SQLite file with four independent kernels: truth, archive, ontology, and promotion
- 2The archive kernel uses FTS5 and reciprocal rank fusion (RRF) to provide fast, accurate keyword search and semantic-like relevance without GPU or network dependencies
- 3Waypath has a human-in-the-loop review gate to prevent agents from silently polluting memory with hallucinated decisions
- 4The modular architecture allows swapping out individual kernels without affecting the others
Details
Waypath is a local-first CLI and MCP server that provides persistent memory for coding agents through a single SQLite file. It addresses the problem of agents like Claude Code, Codex, and Cursor forgetting their architecture decisions between sessions. The tool stores everything in the SQLite file and sits on top of four independent kernels: truth (structured facts), archive (FTS5 + RRF for transcript search), ontology (graph traversal), and promotion (human review gate). This modular architecture allows swapping out individual kernels without affecting the others. The archive kernel in particular uses a hybrid approach of SQLite FTS5 for keyword search and recency-weighted reciprocal rank fusion (RRF) for relevance ranking, achieving sub-40ms retrieval times without GPU or network dependencies.
No comments yet
Be the first to comment