Qarinah cuts coding agent input tokens by 98.7% using compressed project memory
Tool · Dev.to · stat: 98.7% ↓ Qarinah, an open-source project-memory tool, reduces repeated context for coding agents by compiling compact, cited packs instead of replaying full histories. Across six…
Tool · Dev.to · stat: 98.7% ↓
Qarinah, an open-source project-memory tool, reduces repeated context for coding agents by compiling compact, cited packs instead of replaying full histories. Across six software-task fixtures, the tool cut input-context volume from 442,113 tokens to 5,682 tokens. The local-first system also passed 380 of 380 deterministic queries across multi-file projects.
Granular context compression is replacing brute-force needle-in-a-haystack LLM prompting Startups building coding agents can slash API bills and improve retrieval accuracy by adopting local-first context pruning over raw history dumps.
Every claim ties to a primary source. See our methodology.