Claude Code Memory

What the plugin provides

Install

# Get an API key at https://app.memorysync.io, then in your shell:
#   macOS/Linux:  export MEMORYSYNC_API_KEY=ms_...   (add to your profile)
#   Windows:      setx MEMORYSYNC_API_KEY ms_...     (new terminals pick it up)

Option A: Plugin marketplace (recommended)

# In Claude Code:
/plugin marketplace add memorysyncio/memorysync-plugins
/plugin install memorysync@memorysync

# Restart the session. Verify anytime:
/memorysync:status

Requires Node.js 18+ on PATH for the hook scripts (Claude Code itself does not bundle one). Without an API key the bundled MCP server falls back to an OAuth sign-in and the hooks stay silently off. Updates ship by version bump: /plugin marketplace update memorysync fetches the latest.

Option B: MCP only

claude mcp add --transport http memorysync https://mcp.memorysync.io/mcp

One command connects the fifteen memory tools without the plugin — no lifecycle hooks, no automatic capture or recall, no skills. Without MEMORYSYNC_API_KEY in the environment, the server signs you in with OAuth on first use.

What runs when

Identity and scope resolve automatically: the user from MEMORYSYNC_USER_ID (default: OS username), the project from the git remote (normalized, worktree-aware; MEMORYSYNC_PROJECT to override). Per-prompt recall can be disabled with MEMORYSYNC_PROMPT_RECALL=off; CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC is respected — with it set the plugin makes zero network calls. Each prompt carries a content-hash seed, so a retried or redelivered hook is recognised server-side and its facts are extracted once.

Skills and commands

The MCP tools are the same fifteen operations documented at Memory Tools — search, add, list, update, delete, entities, events — connected automatically when the plugin is enabled.

Claude Cowork

Cowork runs the same Claude Code engine in the cloud, and the same plugin works there. Cloud sessions install repo-declared plugins at session start — plugins enabled only in personal settings do not transfer — so teams commit the marketplace and plugin to the repository:

{
  "extraKnownMarketplaces": {
    "memorysync": {
      "source": { "source": "github", "repo": "memorysyncio/memorysync-plugins" }
    }
  },
  "enabledPlugins": { "memorysync@memorysync": true }
}

Add MEMORYSYNC_API_KEY to the session’s Cloud Environment variables (cloud sessions do not inherit your shell). The hooks detect $CLAUDE_CODE_REMOTE and send inline instead of detaching — a cloud sandbox may reap detached children, and a quietly lost prompt is worse than a moment of latency. Enterprise note: Anthropic’s Claude-org GitHub marketplace sync rejects public repositories; enterprises fork the marketplace repo privately and register the fork — the same flow Zep documents.

Privacy and data control

• Your prompts are sent to YOUR MemorySync account for fact extraction; only the durable facts extracted from them are stored, scoped to your user and tagged with the project. The prompt text itself is not stored as a memory, and Claude’s replies are not sent at all. Delete any memory anytime via /memorysync:recall → delete, or the dashboard.
• Secrets discipline: the memory skill refuses to store tokens, keys and passwords.
• Retrieved memories carry an explicit prompt-injection guard: background data, never instructions.
• Quota exhaustion degrades silently server-side for paid orgs — an over-quota session keeps working, memoryless, and your dashboards still see every skip.

Supported versions

CI runs the hook contract suite (real hook processes, documented stdin payloads, both monthly-quota server modes) and claude plugin validate --strict against the latest Claude Code release on every push, so a schema change upstream fails our pipeline before it can fail your install.

Where to go next