How do Claude Code plugins work, and can a plugin add shared memory?

Short answer

A Claude Code plugin is a folder that bundles skills, subagents, hooks and MCP servers, described in .claude-plugin/plugin.json, and installed as one unit from a marketplace with /plugin. Claude Code's built-in memory is local files. A plugin can add memory that lives elsewhere: an MCP server gives Claude tools to read and write it, and hooks load it at the right moment.

What is inside a plugin

A plugin is a folder with a manifest at .claude-plugin/plugin.json and any mix of four building blocks:

Part What it does Role in a memory plugin
Skills Instructions and scripts Claude loads when a task matches Tell Claude when to look something up or save a lesson
Subagents Specialised agents Claude can hand work to Optional: an agent that tidies or summarises memory
Hooks Handlers on events, in hooks/hooks.json Load relevant memory at SessionStart, save before PreCompact
MCP servers Tools Claude can call, local or remote Search, read and write the memory store

Because these come as one unit, a team installs the whole behavior at once instead of copying scripts and settings by hand.

Install and share a plugin

  1. Add the marketplace, if it is not the official one: /plugin marketplace add owner/repo.
  2. Install: run /plugin install name@marketplace, which opens the plugin’s details so you can review it, choose a scope and install. Or browse with /plugin in the Discover tab.
  3. Pick the scope. User scope writes to ~/.claude/settings.json and applies to all your projects. Project scope writes to .claude/settings.json so the plugin is recorded in the repo. Local scope uses .claude/settings.local.json.
  4. Get teammates on it. Project scope records the plugin, but each colleague still installs it with claude plugin install name@marketplace --scope project. For a private marketplace, also list it under extraKnownMarketplaces and make sure teammates have access to its repo. To make a plugin mandatory, admins force the install through managed settings.
  5. Check the cost. claude plugin details <name> shows what the plugin puts in context on every turn.

How a plugin can add memory

Claude Code’s own memory is files: CLAUDE.md, which you write, and auto memory in ~/.claude/projects/<project>/memory/, which Claude writes. Auto memory is machine-local and not shared between machines or cloud environments.

A plugin can add a memory that is not tied to one machine. The pattern:

  1. An MCP server exposes tools such as search, read and save, backed by a store that lives on a server.
  2. A SessionStart hook pulls in what is relevant for this repo or this task, so Claude starts with it.
  3. A skill tells Claude when to consult the memory and when a correction is worth saving.
  4. Optionally a PreCompact hook saves notes before the context window is compressed.

Because the store is shared, what one developer’s session saves can reach the next developer’s session, within the access rights the store sets and if retrieval finds it. A simple test when you evaluate one: save a rule as developer A, ask for it as developer B, and check that the answer points back to where the rule came from.

The misunderstanding: “a memory plugin is just a bigger CLAUDE.md”

A memory plugin that dumps everything into every session would be worse than CLAUDE.md: it would cost tokens on every turn and bury what matters. The point of putting memory behind tools is retrieval. Claude asks for what fits the task and gets a few relevant pieces, not the whole archive. When you evaluate a memory plugin, ask how it decides what comes back, who can see what, and whether you can trace an answer to where it came from.

What this means for a team

The plugin system solves distribution: one install gives every developer the same skills, hooks and tools. It does not by itself make the team’s knowledge shared. That depends on where the plugin’s memory lives and what feeds it. If it only stores what happens inside Claude Code, it still misses the decisions in mail and documents, and the colleague who works in Cursor or ChatGPT. See how to share Claude Code context with your team and Claude Code hooks.

Where Bombos fits

Bombos is itself a plugin of this kind. It gives Claude Code, and the team’s other AI tools such as ChatGPT, Copilot, Cursor and Codex, one shared memory: team knowledge, rules the team has learned, and what already sits in mail, documents and repos, with a trail behind each answer. It runs in the cloud or on your own server, and one developer can install it for their own agents first.

Related questions

How do I install a plugin?

Type /plugin and browse the Discover tab, or run /plugin install name@marketplace. From the shell: claude plugin install name@marketplace, with --scope project to record it in the repo's settings.

How do I add a marketplace?

Run /plugin marketplace add owner/repo. A marketplace is a catalog described in .claude-plugin/marketplace.json. The official marketplace, claude-plugins-official, is added at your first interactive start.

Do plugins cost tokens?

Yes. An enabled plugin puts the names and descriptions of its skills into context on every turn. claude plugin details <name> shows what a plugin adds.

Can IT control which plugins developers use?

Yes. Through managed settings admins can allow or block marketplaces and force certain plugins to be installed.

Sources

  1. Claude Code docs: Plugins
  2. Claude Code docs: Discover and install plugins
  3. Claude Code docs: Hooks reference
  4. Claude Code docs: How Claude remembers your project
  5. Claude Code docs: Set up plugins for your organization
The first half hour · free

What does your AI forget?

  1. 01You tell us which AI you use and where it keeps starting over.
  2. 02We show what a shared memory changes there, on your own examples.
  3. 03You keep the plan, with or without Bombos.
Piet Baudoin
Piet BaudoinYou talk to Piet, from the Bombos teamDutch team from Utrecht and Nijmegen · data in Europe
We reply the same working day.