memcell

Hooks make the loop deterministic. MCP rides beside them as reach.

npm install -g memcell
memcell connect

Run it in a project you code in. It prints a code, one tap in your browser approves it — no account needed, and a guest's memory carries over on sign-in. Wires Claude Code, Cursor, Codex, Gemini CLI, Copilot CLI, OpenCode, Qwen, Grok, Droid, Kiro, Devin and more: hooks, the memcell mcp bridge, an agent key for this directory, and .memcell.

Installed rather than npxed on purpose: the hooks this writes name memcell, and a runner's cached copy is gone by the next session. npx memcell connect works and will say so if the command will not survive.

Commands

command	what it does
memcell connect	wire this directory — approves in your browser the first time
memcell recall	what memory serves before you act
memcell remember	file one thing this project has established (--at when it applies)
memcell report	what happened when something memory served was acted on
memcell status	who this machine is, and what this directory is linked to
memcell stats	what your agents did, and what it saved in tokens not spent
memcell ingest <f>	hand a document to this directory's memory
memcell import	bring existing instruction files into memory — bare, it finds them
memcell export	carry this space out — one document, no account needed
memcell hook remove	take memcell's hooks out of this directory
memcell login	sign this machine in
memcell reset	forget everything memcell keeps on this machine

resource	acts on
memcell config	settings, per project or per machine
memcell spaces	what you work on, and which one is active
memcell agents	the agents wired to this account
memcell memories	the commons

memcell --help <name> explains any of them. Ships two binaries: memcell and mem.

Undo pairs: login ⇄ logout (this machine ⇄ an account), connect ⇄ hook remove (this directory ⇄ a space, with an agent key). Revoking a credential never forgets what the agents filed.

How it works

• Hooks — recall runs at session start and prompt submit; remember and report run at turn end; and a rule that names a moment is served again before the act it bears on, which costs no extra call — the rules come down with the turn's recall and the matching is local. MCP tool calls are model-invoked with no session lifecycle, so they cannot guarantee any of those moments; hooks fire regardless of what the model decides to call. Fails open, always.
• MCP — wiring also registers memcell mcp, a stdio bridge to the paired instance, so the model can ask memory mid-turn. The committed entry is just that command: no secret in any config file.
• .memcell — the instance and the space. Project truth, identical for every clone, safe to commit. Keys live in the machine keyring; a teammate who clones runs one memcell connect of their own.

Options

option	what it does
--url	which memcell to act on, of the ones this directory is wired to
--space	name which of your spaces to wire, by slug
--pair <id>	claim a pairing from the connect page instead of the in-terminal flow
--no-browser	print the approval link instead of opening it

MEMCELL_INSTANCE sets the default instance for a shell.

Claude Code plugin

This repo is also a plugin marketplace:

/plugin marketplace add memcell-ai/cli
/plugin install memcell@memcell

The plugin registers the MCP bridge and a skill teaching the four doors (recall, remember, ingest, report). Hooks still come from memcell connect — the plugin is reach, the hooks are the loop.

Links

• Website — https://memcell.ai
• Docs — https://memcell.ai/docs

Apache License 2.0 · © 2026 OpenOri