hi HN, I'm AJ.

Have been using Claude Code every day, and I was done re-typing the same project rules in every new chat. So I built this.

We try to follow some rules in our coding projects. For eg: "we use pnpm, not npm". The problem is, i tell claude about this rule in one chat, when i start another chat, i have to retype it. Also, when my team opens the same repo using another ai tool lets say cursor, they dont see the rule in their chat. Not cool.

So I built jevmem. It saves your project's rules in a file in your repo. and before Claude Code runs a command, it checks the command against those rules.

How it works:

1) You chat with Claude Code like normal.

2) jevmem saves your rules, decisions and everything that failed, one line each, in a file called JEVMEM.md in your repo.

3) You commit JEVMEM.md like any other file.

4) Claude Code, Cursor and Codex can all read it. now, your team also gets it when they pull.

for example: on Mon you tell Claude "we use pnpm, not npm". On Tuesday your teammate pulls and opens the repo in Cursor. They can see the pnpm rule.

oh and guard is my favourite part. So before Claude Code runs a command or edits a file, jevmem checks it against your rules. If it may break one of your rules, Claude Code stops and asks you first.

Eg: your rule is "don't hand-edit db/migrations/, those files are generated". 3weeks later, in a new session if claude tries to edit a migration, Claude Code stops and shows:

jevmem: this may break a saved rule: "Don't hand-edit db/migrations/: those files are generated." (JEVMEM.md)

You can also retrieve failed attempt. Like if you tried caching API responses in memory, and it broke the build, claude will remember that and some time later when you do the same thing, Claude tells you "Already tried: ..." with the reason.

Numbers from our tests:

-> Claude broke a project rule and left it in the code: 0 of 15 sessions with jevmem, 9 of 15 without it. We checked each one by hand.

-> Claude followed decisions made earlier in the project: 66 of 72 sessions with jevmem, 28 of 72 without.

Tbh, a CLAUDE.md file that you keep updating yourself did about as well here (67 of 72). The difference is that jevmem updates the file for you, likewise for cursor and codex.

jevmem also works alongside Claude Code's own auto memory. It doesn't replace it.

The catches:

- The guard works in Claude Code only. - You need a paid TypeSafe API key. jevmem's requests are billed to that key. - Your message text is sent to TypeSafe to be scored. Common secrets are removed first. - Nothing stored remotely. You turn it on per project. - We ran the tests on our own test projects.

(MIT license) Full results: https://avinash-jetwani.github.io/jevmem/results/

Install:

1. In claude: Plugins > Discover > jevmem > Add

2. npm install -g jevmem

3. In your project folder: jevmem setup

Repo: https://github.com/Avinash-jetwani/jevmem

I'd like to hear about your genuine feedback and how we could make it better!


Comments URL: https://news.ycombinator.com/item?id=50518308

Points: 1

# Comments: 0