Evidence
A number appears on this site only after it has a row in our research ledger. This page is that ledger, as far as the site quotes it: the figure, what it counts, how it was measured, when, and with which build.
Last updated 2026-10-25
The figures#
| Figure | What it counts | Method | Measured | Build |
|---|---|---|---|---|
| -27.4% session cost | $2.31 to $1.68 for the same 27 tasks, three sessions a side, 81 answers | Claude Code 2.1.292 on Claude Sonnet 5.5, the test memory before and after MinniMemoryMCP sorted it; cost at published API rates; a blind grader marked 74 of 81 correct before and 72 after | 6 October 2026 | MinniMemoryMCP 1.1.0 |
| -15.5% tokens | 3,574,009 to 3,019,536 tokens over the same run | The sums of every turn of each arm: fresh input, cache writes, cache reads and output | 6 October 2026 | MinniMemoryMCP 1.1.0 |
| -46.5% re-read every turn | 12,511 to 6,699 tokens loaded on every turn, 5,812 moved to files that load when a task asks | minnimemory init on a full-size test memory of 16 files, as the demo capture shows it | 10 October 2026 | MinniMemoryMCP 1.2.1 |
| 0 lines lost | 16 of 16 files kept every line; undo restored every file byte for byte | The same test memory and the example file that ships in the package, compiled and undone, then compared | 9 October 2026 | MinniMemoryMCP 1.2.1 |
| -58.1% tokens per answer | 1,006 to 422 output tokens per answer on average, 240 test answers | The same questions answered with and without the MinniMemory Interface answer rules, Claude Sonnet | 5 September 2026 | MinniMemory Interface rules run |
| -36% cache writes, MinniCode | 5,286 tokens written to the model's cache in a warm MinniCode session against 8,273 for Claude Code, same tasks | Three warm sessions per arm after one uncounted warm-up, Claude Sonnet through the claude provider, every task checked correct | 21 October 2026 | MinniCode, claude provider |
How to read the table#
Each row is one run. A run has a folder in our research repository with the raw token counts per turn, the grader’s verdicts and a summary; the ledger row points at it. The site quotes the run’s total, not an average of its sessions, and shows the rounded dollars with the exact percent from the run.
What is not here: any figure for MinniTuner. Its proof is background load before and after a plan runs, measured on the user’s own machine inside the app, and we do not quote a frame-rate number for a PC we have not read. When a MinniTuner run is in the ledger it joins this table.
The run behind the first two rows is read line by line in Why Claude Code sessions get expensive. The home page’s Evidence section shows the same rows as charts.