Qarinah is local project memory for coding agents. It records permitted decisions, code structure, tool outcomes, approvals, and sources in an append-only, hash-linked ledger. When Codex, Claude Code, a CLI, or an MCP client needs context, Qarinah builds a small cited pack for the current task instead of replaying the whole project.

I built it because long-running agent work usually fails in one of two ways: the next model receives too much history, or it receives a short summary that cannot be verified. Qarinah keeps the durable record and the compact context pack separate. Every selected memory points back to its source.

In the published evaluator, full-history input changed from 442,113 estimated input-context tokens to 5,682 - 98.71% less repeated context - while every required target ranked in the top five. The benchmark and machine-readable evidence are public on qarinah.io.

What it includes:

- portable memory across Codex, Claude Code, CLI tools, and MCP clients

- cited Markdown and JSON context packs

- typed graph, hybrid index, and Google OKF export

- conflict, supersession, retention, authority, and provenance tracking

- local-first storage with content hashes and integrity checks

- project import/export for cross-editor handoffs

- Node.js 22, 24, and 26 support

- Apache-2.0 licensed source

Install it with: npm install --save-dev qarinah

I am looking for honest failure cases: large monorepos, conflicting decisions, multi-agent handoffs, and very small context budgets.