Anyone who has used Claude Code on a real codebase knows the drill: you write a claude.md file full of project rules, it works for a week, then the codebase evolves and your rules go stale. You either spend time hand-editing them or let Claude rewrite them, which tends to produce bloated, unfocused instructions.
CodeYam, a small dev tools company, just shipped CodeYam Memory as a direct answer to this maintenance problem. The CLI tool runs a background agent that reads your Claude Code session transcripts, identifies patterns where the model got confused or made repeated mistakes, and generates targeted rules scoped to specific parts of your repo.
The interesting technical detail: rather than dumping everything into a single monolithic rules file, CodeYam Memory uses Claude Code's native rules system with path matching. That means a rule about your database layer only activates when Claude is working on database files, not when it's editing your frontend components. This keeps context windows (the amount of text the model can process at once) cleaner and rules more relevant.
The tool is positioned as a "set it and forget it" layer on top of Claude Code's existing configuration. You run your normal coding sessions, and CodeYam periodically reviews what happened and updates your rules accordingly.
It's a narrow but real pain point. Managing AI coding assistant context is one of those unsexy problems that burns small amounts of time repeatedly. Whether CodeYam's automated approach actually produces better rules than a developer spending five minutes on manual updates is the open question. The path-scoped targeting is a smart design choice, though, and it's the kind of tooling that tends to emerge once a platform gets enough daily users hitting the same friction.