Claude Code and Codex have a nasty habit of treating your

The "infinite context window" is a lie if the agent still gets confused by stale decisions or buries the actual tool output under a mountain of transcript noise. I got tired of paying the "context tax" and explaining the same architectural choices over and over, so I started using Qarinah. It's a local project-memory compiler that actually tracks provenance instead of just guessing based on vector similarity.

Why standard RAG fails coding agents

Most people think a bigger window or basic embeddings solve memory. They don't. Embeddings find "similar" text, but similarity isn't the same as authority. A snippet from a deprecated README is "similar" to your current task, but it's functionally a lie. You need to know if a newer decision superseded an old one, and you need a citation to prove it.

Qarinah treats project knowledge as "typed events" (decisions, approvals, tool outcomes) with explicit relations like supersedes or blocks. Instead of dumping the whole history into the prompt, it builds a compact, evidence-linked pack specifically for the task at hand.

Real-world performance

The math on this is actually wild. In an evaluator covering things like TS refactoring and production debugging, the input context dropped from over 442k tokens to about 5.6k. That's a ~98% reduction in repeated context. While your actual bill won't drop by exactly 98% (thanks to output tokens and caching), it's a massive win for speed and accuracy.

AI workflow: Getting it running

If you want to stop the token bleed, here is the practical tutorial for setting it up.

1. Install and init in your project:

npm install --save-dev qarinah
npx qarinah init . --capture content
npx qarinah scan
npx qarinah doctor

2. Record a decision so the agent doesn't ask you for the 50th time:

npx qarinah record --kind decision \
 --title "Use additive database migrations" \
 --body "Add, backfill, switch, then remove. Never rename a production column in place."

Integrating with your tools

You can plug this directly into your LLM agent setup. It stays local, so you aren't uploading your whole repo to some random backend.

For Codex:

codex plugin marketplace add AjnasNB/qarinah --ref v0.1.1
codex plugin add qarinah@qarinah

For Claude Code:

claude plugin marketplace add AjnasNB/[email protected] --scope user
claude plugin install qarinah@qarinah --scope user

This turns your project memory into a structured database rather than a giant, expensive scroll of text. Your wallet (and your agent's sanity) will thank you.