GPS

GPS is an AI memory layer for coding agents that persists your repository's knowledge and surfaces it contextually during edits, built for developers using Claude Code, Codex, and Cursor. Instead of dumping conventions into a bloated CLAUDE.md, GPS anchors notes, invariants, and decisions to specific functions and components, then builds a queryable graph of code structure — calls, callers, tests, and provenance — alongside human intent like repo conventions, past corrections, and test commands. When an agent prepares an edit, GPS returns a decision brief containing the relevant structure, tests, invariants, notes, and prior decisions, so the agent stops repeating mistakes like logging PII in the wrong place. It captures hard-won lessons automatically, and recurring notes graduate into standing rules over time. GPS is local-first and CLI-first, integrating via CLI or MCP, with one-command setup via npx. It is open source under the MIT license and currently in alpha (v0.3.0), with a single-repo proof base and monorepo support planned.

Developers use GPS to give their AI coding agents persistent, repo-specific memory so the agents recall rules, decisions, and test commands across sessions.