Claude Code, Claude Desktop, and Codex
All three have a Memcrate integration today, and all three read the same SKILL.md format, so one canonical set of skills serves them. Cursor, Aider, and MCP are on the roadmap.
Install
From any directory:
memcrate
One command installs into both tools:
~/.claude/skills/ ~/.codex/skills/
├── save/SKILL.md ├── save/SKILL.md
├── load/SKILL.md ├── load/SKILL.md
└── pin/SKILL.md └── pin/SKILL.md
On Windows those are %USERPROFILE%\.claude\skills\ and %USERPROFILE%\.codex\skills\.
Both tools pick the skills up on the next prompt. No restart, no permission setup.
Use
In any Claude Code or Codex session, the three verbs are typed commands. The prefix differs between the two tools:
| Verb | Claude Code | Codex |
|---|---|---|
| load | /load | $load |
| save | /save | $save |
| pin | /pin | $pin |
Claude Code uses a leading slash; Codex invokes skills with a leading dollar sign. Same skill file, same behavior, different prefix.
Each one runs the matching skill. The skills handle:
- Finding your vault (default
~/memcrate-vault; falls back to a$HOMEscan) - Reading the canonical context files
- Writing session logs and context updates in the right format
What each skill does
/load
Reads Core/Context/Profile.md, Core/Context/Projects.md, Core/Context/Current State.md, and the three most recent session logs. Gives you a short oriented summary. Accepts an optional argument to scope by topic or number of sessions.
/save
Reviews the current session and writes a structured log to Core/Sessions/YYYY-MM-DD-slug.md with frontmatter and the standard sections.
/pin
Promotes a fact, decision, or preference from the current session into one of the canonical context files. Figures out which file (Profile / Projects / Current State) and updates it in place.
Vault discovery
The skills probe in this order:
- Read
~/memcrate-vault/Core/Context/Profile.md - Read the current working directory’s
Core/Context/Profile.md - Scan one level deep in
$HOMEfor*/Core/Context/Profile.md(catches custom locations like~/myvault)
Step three names the concrete call for each tool, since an explicit instruction is followed more reliably than an abstract one: Claude Code uses its Glob tool, Codex runs ls -d.
So if you ran memcrate --vault ~/myvault, the skills will still find it without configuration.
Skills you already own
Memcrate marks every skill it installs. It will never replace a skill it did not install, so if you already have your own load, save, or pin, it leaves yours alone, sets up everything else, and tells you which one it skipped.
That means a conflict in one tool does not cost you the other. If your Claude Code setup already has a load skill, Codex still gets the full set.
Upgrading
Each Memcrate release ships the latest skill files. To upgrade:
memcrate
This refreshes the skill files in place with the current versions.
Claude Desktop
Nothing extra to install. Claude Desktop’s local agent mode reads the same ~/.claude/skills/ directory Claude Code does, so the memcrate run above already set it up. Type /load, /save, and /pin exactly as you would in the terminal.
Skills you upload through the Claude web UI are a separate thing. Those live in your account and sync down into a cache Claude Desktop manages, and they can’t reach a vault on your disk. For local work you want the filesystem install above, which is what memcrate does.
Other tools
The roadmap (in priority order):
- Cursor -
.cursorrulestemplate that maps natural-language patterns to the verbs - Aider -
.aider.conf.ymlrewriter that adds the canonical context files - MCP -
mcp-memcrateserver exposingvault_save,vault_pin,vault_load,vault_search
Until those ship, the vault format works with any tool that can read files. Point your tool at the four canonical files and you get most of the value.