From a repository to the first memory
Two steps in memkith pick the repositories that capture memory. Three in your terminal connect Claude Code or Codex. After that, your agent writes memory while it works.
Before you start
- A repository on GitHub. memkith scopes memory to a repository, and GitHub is how it learns which ones exist.
- Homebrew on macOS or Linux, which is how the CLI ships today.
- A memkith account. Create one if you have not.
In memkith
Once per workspace. These decide which repositories an agent is allowed to write memory to, and the dashboard checks each one off as it lands.
01
Connect GitHub
From the dashboard, install the memkith GitHub App on the account or organization that owns your repository, then link your GitHub account. The install says which repositories exist; the link says which of them you can reach.
memkith reads repository names and who can reach them. It never reads your code.
02
Turn on a project
Synced repositories do not capture anything until you turn them on. Pick the ones your team works in; turning one on is what lets an agent write memory to it.
On your machine
Once per machine for the first two. memkith runs as an MCP server your agent talks to, so the last step registers it with the agent you use.
03
Install the CLI
Installs memkith, memkith-hook and memkith-mcp on macOS and Linux as native binaries, so no Python is needed. The first line trusts the tap, which Homebrew requires before it loads a third-party formula.
04
Sign in
Opens a browser to connect this machine to your account. Terminals without a browser take
--no-browser, which prints the sign-in URL instead.05
Connect your agent
Claude Code
Run from the repository root. It writes the server into that project's .mcp.json.
Codex
Registers the server once for every Codex session on this machine.
Using Cursor or Zed?
memkith integrate cursorormemkith integrate zedprints the snippet for that agent’s settings file.
The first memory
Memory is written by your agent, mid-session, not by connecting it. Give it a first task that is worth remembering.
06
Open the repository
Open a checkout of a project you turned on. Its
originremote is how memkith knows which project a session belongs to, so a fork or a copy with a different remote does not match.07
Start a new session
An agent reads its MCP servers when it starts. A session that was already running when you connected it will not see memkith until you start a new one.
08
Ask for the first memory
From then on, work as usual. Agents record decisions, failures and preferences as they come up, and each one appears on the project’s page in the dashboard.
The rest of the CLI
What you reach for after setup. memkith --help stays the authority.
Report whether this machine is signed in, and to which account.
--refresh forces a token refresh before reporting.
Install a post-commit hook so pending memory anchors fill in automatically as code lands.
Stamp the commit anchor on memories whose scoped code has since been committed. Idempotent, so it is safe to run at any time.
--quiet prints nothing, for use from the post-commit hook.
Print the MCP setup for every supported agent: Claude Code, Codex, Cursor and Zed. Name one agent instead of passing --all.
Sign this machine out.
Print the installed version.