# MM3

> The memory and decision layer for coding agents. Every answer is citable. MM3 turns a short numbered yes/no checklist about code or a plan into a calibrated pass/fail/unsure verdict (evidence, never a command), reuses answers for unchanged code, and appends every run to a ledger in the project. Knows in ~20 ms (a free ledger lookup), learns in ~500 ms (a new verdict). Beta, Apache-2.0, Node 22.13+. Source: https://github.com/mvp-scale/mm3

Last updated 2026-10-06, synced with https://github.com/mvp-scale/mm3 main @ aff6af6 (machine-readable: https://mm3lab.dev/sync.json). Agent support and the pattern playbook change often, so treat the live pages below as the source of truth.

Using MM3 from an agent: the `mm3` command (or its MCP tool in Claude Code) has six verbs, `view`, `class`, `replay`, `scan`, `drill` and `loop`, plus `report`, `outcome`, `budget` and `template`. `mm3 agent` prints the full command card, `mm3 agent <verb>` prints one verb's rules, and `mm3 template <verb>` prints a filled-in request. Every answer ends with a `next:` line. MM3 reports odds and a human makes the call. Install: `npm install -g @mvpscale/mm3` then `mm3 init --agents` for any agent that reads AGENTS.md, or `/plugin marketplace add mvp-scale/mm3` and `/plugin install mm3@mvp-scale` in Claude Code.

## Docs

- [Overview](https://mm3lab.dev/index.md): what MM3 is, install, a real run, how it works, architecture, limits
- [Pattern playbook](https://mm3lab.dev/patterns.md): 30 recipes, one per verb at five levels (use, standardize, chain, integrate, anticipate), plus how loop, scan and drill decompose
- [Agent support](https://mm3lab.dev/agents.md): which coding agents work, which are in testing, and what is known to work or not
- [Agent skill](https://raw.githubusercontent.com/mvp-scale/mm3/main/skills/mm3/SKILL.md): when to reach for MM3 and how to write a request
- [The contract](https://raw.githubusercontent.com/mvp-scale/mm3/main/docs/contract.md): every claim the answer format makes
- [Contributing and rules](https://raw.githubusercontent.com/mvp-scale/mm3/main/AGENTS.md): commands, test tiers and rules for working in the repo

## Evidence

- [Evidence for each claim](https://raw.githubusercontent.com/mvp-scale/mm3/main/docs/evidence/README.md): the tests and runs behind each claim
- [Where the numbers come from](https://raw.githubusercontent.com/mvp-scale/mm3/main/docs/numbers.md): calibration and method
- [Journeys](https://github.com/mvp-scale/mm3-journeys): full public runs you can rerun and compare

## Optional

- [Full text of this site](https://mm3lab.dev/llms-full.txt): the README, the pattern playbook and the agent table in one plain-text file
- [README](https://raw.githubusercontent.com/mvp-scale/mm3/main/README.md): the repository front page
- [Issues](https://github.com/mvp-scale/mm3/issues): report a bug or an agent that does not work