Agent Memory plugin

Shared project history and cooperative edit ownership for Codex, Claude Code, Grok Build, Cursor, and other Agent Skills or Agent Plugins hosts. One skill teaches focused retrieval, full-record expansion, evidence verification, claims, and durable knowledge maintenance across repositories.

Prerequisite

Run memd 0.6.3 or later on the machine running the agent. The plugin connects over HTTP to http://127.0.0.1:8787/mcp; it does not start another daemon or own a database. Check curl http://127.0.0.1:8787/health before connecting. Remote or cloud agents need an endpoint reachable from their own runtime; their loopback address does not refer to your laptop.

Plugin versions are independent of daemon versions. Updating this plugin updates guidance and connection configuration. It does not install or restart memd.

Codex

codex plugin marketplace add rhymiz/agent-memory
codex plugin add agent-memory@agent-memory

Start a new session and select Agent Memory. The repository marketplace is .agents/plugins/marketplace.json. The package includes both a portable plugin.json and the .codex-plugin/plugin.json compatibility manifest. For an exact release, add the marketplace with --ref plugin-v0.1.1.

Claude Code

claude plugin marketplace add rhymiz/agent-memory
claude plugin install agent-memory@agent-memory

The catalog is .claude-plugin/marketplace.json. To test a downloaded or cloned bundle for one session, use claude --plugin-dir /path/to/agent-memory.

Grok Build

grok plugin install 'rhymiz/agent-memory#plugins/agent-memory'
grok plugin details agent-memory

Grok supports the Claude plugin layout. If it already discovers this plugin from your Claude installation, use that installation instead of adding a second copy. For a local bundle, pass its directory to grok plugin install. Local installations follow the source directory through a symlink; update that source to update the installed plugin. Before changing installation routes, inspect grok plugin list --json and grok plugin details agent-memory for the selected source. Different sources with the same plugin name can coexist while discovery shows only one; uninstalling it can reveal the other. After removing an obsolete source, inspect again before installing a replacement. Preserve the installation you intend to use and verify its manifest version in a fresh session.

Cursor

Load a local bundle with the CLI:

cursor-agent --plugin-dir /path/to/agent-memory

Sign in with cursor-agent login and approve the plugin's MCP connection and requested tools. Cursor names this plugin's server plugin-agent-memory-agent-memory. Headless tool permissions must use that name, not the standalone MCP configuration key agent-memory. For retrieval and batch coordination, merge these entries into .cursor/cli.json:

{
  "permissions": {
    "allow": [
      "Mcp(plugin-agent-memory-agent-memory:project_briefing)",
      "Mcp(plugin-agent-memory-agent-memory:memory_get)",
      "Mcp(plugin-agent-memory-agent-memory:claims_acquire)",
      "Mcp(plugin-agent-memory-agent-memory:claims_renew)",
      "Mcp(plugin-agent-memory-agent-memory:claims_release)"
    ]
  }
}

Include other named tools needed by your workflow, such as project_context_get or memory_update, using the same prefix. A separately configured standalone server named agent-memory uses Mcp(agent-memory:tool) instead. The --approve-mcps flag approves connections; it does not grant every tool permission. See Cursor CLI permissions.

For the IDE, place the plugin directory at ~/.cursor/plugins/local/agent-memory and start a new agent session. For example, from a checkout of this repository, on a first installation:

mkdir -p ~/.cursor/plugins/local
cp -R plugins/agent-memory ~/.cursor/plugins/local/agent-memory

Cursor also supports repository marketplaces through its Customize/team marketplace UI; use https://github.com/rhymiz/agent-memory. The catalog is .cursor-plugin/marketplace.json. Team marketplace availability depends on your Cursor plan. This repository provides an installable package; it does not imply listing or approval in Cursor's public marketplace.

Standalone skills and other agents

Use the portable plugin.json and mcp.json with hosts that implement Agent Plugins. Hosts without plugin support can install the shared-agent-memory directory from skills/, or the standalone skill release archive. Copy the complete directory, including references/.

Common user-level skill locations:

Host	Skill directory
Codex and compatible agents	`~/.agents/skills/shared-agent-memory`
Claude Code	`~/.claude/skills/shared-agent-memory`
Grok Build	`~/.grok/skills/shared-agent-memory`
Cursor	`~/.cursor/skills/shared-agent-memory`

Configure an HTTP MCP connection separately at http://127.0.0.1:8787/mcp when using only the skill. Use the same approach for a custom endpoint: configure the connection in the host, rather than editing an installed plugin cache. A host must support both skill discovery and MCP to load both parts; arbitrary harnesses may require their own instruction injection and MCP configuration.

Use one installation route per host to avoid duplicate skills or connections. Choose the destination supported by the installed client, preserve unrelated configuration, and refresh or restart its session after installation. Check both the selected skill path and MCP source. User-level, repository-level, plugin, and compatibility discovery can expose different copies of the same name.

Using the guidance

Ask the agent to use the shared-agent-memory skill for the current task. For consistent project use, add a short instruction to the project's normal agent guidance file:

Before significant project work, use the shared-agent-memory skill to retrieve relevant context and coordinate shared edits. Skip trivial standalone tasks.

The skill derives identity from the current repository unless the project supplies an explicit ID. Consumer project IDs, credentials, local paths, and evaluation questions do not belong in this package. A plugin does not grant permission to edit a project or store knowledge beyond the user's task.

Verify the installation in a fresh session: the agent should identify the bundled skill, retrieve a focused briefing, expand a relevant full record, and distinguish historical evidence from current verification. Discovery alone does not establish correct use. No particular memory needs to exist in a new project.

Releases and updates

Plugin releases use tags such as plugin-v0.1.1, with agent-memory-plugin-0.1.1.tar.gz, agent-memory-skill-0.1.1.tar.gz, and SHA256SUMS. Verify the checksum and unpack the plugin archive; its agent-memory/ directory is the complete plugin. The skill archive contains shared-agent-memory/ for standalone installation. Neither archive requires Bun or Node to load its guidance.

Marketplace installations update through the host's plugin manager; pinned Git refs remain pinned until changed. Local copies and standalone skills need an explicit update from the new release. Unpack into a new directory, verify it, then replace the selected old bundle; copying over it can leave removed files behind. For symlink-based installations, update the source directory the link targets. Keep the plugin's manifests at the same version and test discovery after an update. Plugin releases do not replace the latest daemon release used by the Linux installer.

For release verification, start a fresh headless session with the documented permissions. Have the agent identify its skill file and selected plugin version, retrieve a focused briefing, and expand a relevant full record when one exists. Exercise claim acquisition and release in an isolated project, including an invalid resource that should produce a readable error before correction. Inspect actual tool results and remaining claims. Missing knowledge and client/provider failures must remain visible in the verification result.