We built CodeYam Memory because Claude Code kept repeating the same mistakes and our claude.md files got stale. CodeYam Memory uses a background agent to review your coding session transcripts, identifies confusion patterns, and generates targeted rules with proper scoping. This is a small first step towards our vision of exploring the ideal AI-native development experience, packaged as a lightweight CLI that you can use wherever you use Claude Code.
Hey PH community! I’m Nadia, one of the makers. Happy to answer anything.
We built CodeYam Memory because Claude Code kept making the same mistakes on our codebase. Our claude [dot] md files quickly got stale and maintaining by hand or with Claude wasn’t sufficient.
While digging into this we found that Claude has a native rules system that allowed us to target specific parts of our repo with path matching. This was ideal for our use case but trying to manage these rules by hand was already not working and would be even harder with more granular, targeted rules.
CodeYam Memory uses a background agent to review your coding session transcripts, identifies confusion patterns, and generates targeted rules with proper scoping. You review and approve everything. Dashboard for auditing, a background-agent review process so nothing goes stale as code changes, tracking of everything lives in a simple file in git.
How to Get Started:
Install:
npm install -g @codeyam/codeyam-cli@latest
Then from your project root run:
codeyam
This will launch a dashboard with further instructions for initializing CodeYam Memory.
Free, runs locally, no login required, and language agnostic. Would love feedback.
Some context on what “rules” are for people who haven’t seen them: Claude Code has a built-in system for structured context beyond claude.md files. Rules support path matching (apply context only to specific files/directories), scoped organization, and structured formatting.
We have been running CodeYam Memory on our own repo for the past few weeks. The main difference we see is fewer repeated mistakes and less manual context maintenance. It’s still early, but it has meaningfully improved how we work with Claude Code.
If discussion gets too long for HN threads, we also have a Discord for questions and feedback: https://discord.gg/eFPUs7CeFw
Have been using CodeYam for weeks; the "confusion archaeology" is SPOT ON, and frankly these insights were great for my team as well as the agents.
This is a real problem. We've burned a lot of time manually updating claude.md files that drift out of sync with how the codebase actually works. Background agent approach makes a lot of sense. Checking this out!
the stale claude.md problem is real and you've nailed the fix for project context. what i keep running into is the other half of the cold start - claude doesn't know who the developer is either. role, preferences, how they like to work. codeyam handles what the codebase needs claude to know. been working on the human side of that same gap with northr identity. feels like these two would stack well.
oh man this hits close. been building a desktop app with Claude Code for months and the CLAUDE.md staleness problem is probably our #1 source of wasted time. you write great instructions, they work for a week, then the codebase evolves and the agent starts doing weird things because the context is outdated.
we ended up with this whole system of CLAUDE.md + AGENTS.md + skills files and honestly keeping them in sync is like its own part-time job at this point.
curious about the "confusion pattern" detection — how does it figure out when Claude is confused vs when the instructions are just ambiguous? those feel like two different problems to me but maybe I'm wrong. also does it work with sub-agents / multi-agent setups or only single CLAUDE.md?
Oh this is solving a real pain point. I use Claude Code daily and the CLAUDE.md file situation gets messy fast, especially when you're juggling multiple features across different parts of the codebase. The background agent reviewing transcripts is a clever approach. Does it handle conflicting rules well? Like if one session teaches it something that contradicts a previous rule?
