Qarinah
Your project remembers when your coding agent changes
11 followers
Your project remembers when your coding agent changes
11 followers
Qarinah 0.1.6 gives Codex, Claude Code, Cursor, CLI tools, and compatible MCP clients one local, cited project memory. Published six-fixture estimate: 442,113 full-history tokens vs 5,682 Qarinah tokens—98.71% less repeated context and a 77.81:1 ratio, with every required target top-five. At a flat $3/M uncached input rate, the compared slice is $1.326339 vs $0.017046, an estimated $1.309293 saving per repeat. Apache-2.0, local-first; not a provider bill or universal guarantee.
Hey Product Hunt - I’m Ajnas, the maker of Qarinah. Qarinah gives Codex, Claude Code, Cursor, CLI tools, and compatible MCP clients one project-owned memory. It keeps permitted decisions, outcomes, code relationships, and source evidence beside the repository, then compiles only the compact, cited context needed for the next task. Setup takes one command: npx qarinah setup . --codex --claude --cursor --capture content --allow-query Then use $qarinah in Codex, /qarinah in Claude Code, the project MCP configuration in Cursor, or npx qarinah query "" from any terminal. Latest verified Qarinah 0.1.6 results: - 442,113 to 5,682 estimated tokens across six software tasks - 98.7148% less repeated context and a 77.81:1 ratio, with every required target directly covered in the top five. - 9,489 to 119 estimated tokens for the model-facing continuation capsule - 98.7459% less. - 9,489 to 1,039 estimated tokens for the complete cited continuation pack - 89.0505% less. - 380/380 deterministic file retrievals, with 9/9 unsupported controls rejected. Every selected memory carries its event ID and content hash. Qarinah is local-first, Apache-2.0, and works without a hosted Qarinah account or API key.