What you save when project context stops repeating

Qarinah reduces the amount of repeated context in coding projects by compiling a compact project-memory pack. This analysis compares the full-history baseline to the Qarinah pack across six software-task fixtures.

The full-history baseline contained 442,113 portable estimated input-context tokens, while the Qarinah pack used just 5,682 tokens. This means 436,431 fewer estimated input-context tokens, a 98.71% reduction in repeated context, and a 77.81:1 baseline-to-pack ratio. This ratio refers to the compared input-context volume, not a claim about cost or session length.

To estimate the cost savings, the table applies four flat, uncached input-token rates to the same two token estimates. At $1 million tokens, the full-history baseline costs $0.442113, while the Qarinah pack costs just $0.005682, saving $0.436431. At $3 million tokens, the savings increase to $1.309293. The pattern continues for $5 and $15 million tokens.

The calculation is simple: estimated tokens divided by 1 million, multiplied by the flat input rate. This calculation excludes provider-native tokenization, caching, output tokens, reasoning tokens, tool calls, retrieval, hosting, and fixed fees. Actual costs depend on the provider, model, cache behavior, context composition, and repetition frequency.

Despite compression, the pack remains useful because it preserves source event IDs and content hashes for selected context. This allows later agents to perform bounded handoffs that can be inspected instead of receiving an opaque story. Qarinah also passed all deterministic file-specific exact and typo-tolerant queries, verifying its retrieval behavior without establishing universal task quality.

The methodology, fixture records, costs, and exclusions are publicly available. Qarinah is an open-source, Apache-2.0 licensed tool compatible with various coding platforms. Project capture is opt-in and metadata-only by default, ensuring privacy.