Reduce your coding agent’s token usage with MinniMemoryMCP.
Now available in Claude Code
MinniMemoryMCP splits your agent’s memory into a small core that loads every turn and files that load only when a task needs them. Measured in Claude Code: 27.4% cheaper sessions, no lines lost.
$ npx -y minnimemory installtokens re-read every turn, before and after, test memory. Counted with MinniMemoryMCP 1.2.1, 10 October 2026.
Watch Minni AI at work.
Our research and four tools in one video: MinniMemoryMCP, MinniCode, MinniTuner, and MinniMemoryAgent, a memory upkeep agent that is not public yet. Every number on screen comes from a real run of our tools.
How we measured 27.4% cheaper Claude Code sessions.
Five results, measured, across MinniMemoryMCP and MinniCode.
- Cache reads2,931,482 to 2,609,556
- Cache writes626,130 to 391,988
- Output16,181 to 17,752, +9.7%
Cost falls further than tokens: the saving is mostly in cache writes, the most expensive input, while cache reads cost a tenth as much.
3,574,009 → 3,019,536 tokens$2.31 → $1.68 for the run72 of 81 correct after, 74 before
How we measured
Over 3.5 million tokens of measured work
After many rounds of testing and tweaking our interface, this is our final run. We gave Claude Sonnet 5.5 the same work in Claude Code, three sessions a side and 81 answers in all: once with the test memory as it was and once after MinniMemoryMCP sorted it. The run used 3,574,009 tokens before and 3,019,536 after, 15.5% fewer, and cost $2.31 before and $1.68 after, 27.4% less. A blind grader marked 74 of 81 answers correct before and 72 of 81 after.
MinniMemoryMCP 1.1.0, measured 6 October 2026
Less memory loaded on every turn
Our full-size test memory went from 12,511 tokens to 6,699 tokens loaded on every turn. The rest moves into files the agent opens only when a task needs them. No lines are lost.
Counted with MinniMemoryMCP 1.2.1 on a full-size test memory, 10 October 2026
MinniCode writes less to the cache than Claude Code
On the same coding tasks, a warm MinniCode session wrote 5,286 tokens to the model's cache against 8,273 for Claude Code, 36% less. Three sessions each on Claude Sonnet, after one uncounted warm-up, every task checked correct.
MinniCode on the claude provider, measured 21 October 2026
Every line kept, every file restored
We compiled the same 16-file test memory and the example file that ships with the package. Every line landed somewhere word for word, and undo put every file back byte for byte.
MinniMemoryMCP 1.2.1, measured 9 October 2026
Shorter answers with our answer-style rules
We asked an AI agent the same questions with and without the rules, 240 answers in all. With the rules, answers used 58.1% fewer tokens.
Measured 5 September 2026
Methods
Session cost: the tokens of every call, priced at Anthropic's published API rates, with the cache reads and writes each call made. Cost falls further than tokens because the saving is mostly in cache writes, the most expensive input, while cache reads, most of the tokens, cost a tenth as much. The grader saw each pair of answers blind, in both orders, and a verdict counted only when both orders agreed.
Memory counts: the token counter built into MinniMemoryMCP 1.2.1, run on a full-size test memory of 16 files loaded every turn. Anyone with the package and the same files gets the same numbers.
The 58.1%: the same questions answered with and without the rules, counting the tokens in each answer.
Every figure on this site
The evidence page lists each number the site quotes with what it counts, the method, the date and the build.
Read the run line by line
Why Claude Code sessions get expensive walks through this run's token table and what moved.
The MinniMemory Interface: rules for a smaller agent memory.
Our research, written down: how an agent’s memory is split, what it reads every turn, and the rules that keep its answers short. Every result above comes from it. You can apply it two ways today.
MinniMemoryMCP
initmaps the memory, previews the splitapplymakes the splitundorestores the originalsdoctorre-checks it later
We apply it to your agent
- 01Measurewhat your agent is handed each turn
- 02Applythe interface to its memory and instructions
- 03Reporttokens saved against answer quality
Three tools, one rule.
You see every change and approve it. Pick a tool and watch it work.
MinniMemoryMCP
An MCP server that shrinks what Claude Code loads every turn
- Less for Claude Code to re-read every turn
- Nothing in your memory is lost
- Plain Markdown files you own
$ npx -y minnimemory installHelp keep the research running.
MinniMemoryMCP and MinniCode are free, and so is BasicTuner. We pay for the test runs, the AI calls and the hosting behind them. If our tools saved you time or tokens, you can back the work.
Payments go through Buy Me a Coffee. Support is a gift: it does not buy a product or a license, and it is not tax deductible.
Frequently asked questions
01Do you see my code or my memory files?
MinniMemoryMCP never does. It runs entirely on your computer, sends nothing anywhere, needs no account or API key, and has no AI model inside it. MinniCode sends each request straight from your computer to the provider whose key you gave it (NVIDIA, OpenRouter or Claude) or to a model on your own computer, so our servers never see your code. The chat window on this page is answered by an outside AI provider, and we don't store those messages either.
02Why not just edit the file myself?
You can, and for a short file you should. It gets harder as the file grows: you have to decide which lines your agent needs on every turn, move the rest into separate files without dropping anything, and leave a pointer so the agent knows when to open each one. MinniMemoryMCP does that sorting for you, backs up your originals first, and can re-sort after you edit the file again.
03What if it wrecks my memory files?
It backs up your originals before it writes anything, and one command, undo, puts them back exactly as they were. Inside Claude Code the server itself never writes: apply and undo return the command to run and the list of files it touches, and your agent runs that command through your own approval prompt, so you see every change. Run from the terminal, apply writes the files itself, after the backup.
04What does it cost?
MinniMemoryMCP is free during early access. MinniCode is free. MinniTuner is in open beta: the free tier (checkup, chat, Basic plans, undo) is free forever, and MinniTuner Pro is $19.99 one time for 3 PCs with lifetime updates.
05Does MinniCode need a paid API key?
No. MinniCode runs on free models from NVIDIA or OpenRouter with your own free key, on your Claude subscription through Claude Code, or on a model on your own computer through Ollama or LM Studio. It asks for the key the first time you run it, and requests go straight from your computer to the provider.
06Can MinniTuner change something I cannot undo?
It tells you first. Before a plan runs it saves a restore point, then applies each step with its own undo record, and Undo replays that log. Anything that cannot be undone says so before you approve it. Windows Defender stays on, and anti-cheat and game launcher services are never turned off.
07Who makes Minni?
Minni was founded by Hamzah Muhammad, a former software engineer at Amazon, with two co-founders: a second engineer also from Amazon, and Reynardo, formerly at Microsoft. Four engineers build it in all. We ran into the same problem in our own work: coding agents re-reading their whole memory every turn, and token bills growing with it. Minni is what we built to fix that, a smaller always-loaded memory with the rest loaded only when needed, and we publish a number only after we have measured it.
Talk to us before it gets expensive.
Send us what your agent loads each turn. We apply the interface and report tokens saved against answer quality.