# Memcrate

what it does

memcrate gives your ai tools a single place to read and write your personal context (the projects you’re working on, the decisions you’ve made, the sessions you’ve already had) so you stop re-explaining yourself to every new chat.

it’s three pieces:

• a vault format: a defined Core/Context/ + Core/Sessions/ directory shape in plain markdown. human-readable, git-able, no database
• three verbs: /save writes a session log, /pin promotes a fact to permanent context, /load pulls the right context back at the start of a new session
• a rust cli: running memcrate creates the vault and installs the skills; memcrate profile is there when you want to fill in your identity and project context

install one-liner on linux/mac:

curl -fsSL https://memcrate.dev/install.sh | sh
memcrate

on windows:

irm https://memcrate.dev/install.ps1 | iex
memcrate

or use npm install -g memcrate / cargo install memcrate if those fit your machine better.

why i built it

i use claude code, cowork, and a handful of other ai tools across three machines. every one of them started a session knowing nothing about me, so i kept pasting the same “here’s who i am, here’s what i’m working on, here’s where the project lives” preamble at the top of every conversation.

the /save /pin /load pattern was something i’d been running by hand inside my own obsidian vault for a while. it worked. but it was claude-code-only, glued to my specific folder layout, with skills i’d written for my own machine. nothing portable, nothing another person could adopt.

memcrate is that pattern generalized: a defined vault shape that any tool can read, a cli that sets it up in one command, and skill packs you install per-tool. local files. your context. nothing leaves your machine.

the verb trio generalizes prior art from EliaAlberti/cpr-compress-preserve-resume, a claude-code-only session-memory skill pack, and adds /pin for the bridge from session memory to permanent memory.

how it’s built

a quick tour:

• rust cli: single static binary. ships the reference vault and skill packs embedded via include_dir!, so the first run does not need to download templates
• github actions release pipeline: tag push triggers a matrix build across linux x86_64, mac arm64, and windows x86_64; uploads tarballs + zip + sha256 to github releases
• install scripts: install.sh (posix) and install.ps1 (windows) detect the platform, download the right archive from releases/latest/download/, and drop the binary on PATH. both honor MEMCRATE_VERSION and MEMCRATE_INSTALL_DIR env overrides
• npm wrapper: npm install -g memcrate downloads the same prebuilt binary for people who live in Node tooling
• cargo publish: cargo install memcrate works as the universal fallback (intel mac, anything weirder, anything without a release asset)
• skill packs: Claude Code, Claude Desktop, and Codex skills ship today. Future tools should be thin adapters over the same local vault format
• license split: code is MIT, the format spec under docs/spec/ is CC0. the format is the load-bearing part; anyone should be able to implement against it

what’s next

• more tool surfaces: Cursor, Aider, and an MCP server are the next adapter candidates. format-first means every new tool is a thin pack, not a fork
• more live-agent checks: the v0.4.2 test proved that real agent behavior catches issues static checks miss
• a launch piece: an article walking through the verb trio and why local-markdown is the right shape for personal context

the next thing is making the “every ai tool already knows me” experience work the same way on a brand-new machine as it does on my main one. install the cli, run one command, and the context is portable infrastructure.