Empirical

empirical-memory

Cross-tool, cross-project memory for Claude Code, Codex CLI, and GitHub Copilot CLI, powered by Empirical.

Start an idea in ChatGPT, pick it up in your coding agent with full context carried over. Track leads and relationships as durable memory instead of a separate CRM. Organize memory into separate workspaces for personal, team, or client contexts. Recalls prior decisions, preferences, and conventions across sessions and projects, and persists new durable ones as you work — so you don't have to re-explain context you've already given elsewhere.

What this marketplace provides

• empirical-memory: the required core memory Skill and MCP server connection.
• empirical-work-history: optional concise completed-work summaries.
• empirical-scar-tissue: optional recurring failure patterns and prevention guidance.
• empirical-todo: optional project-based todos with status and next actions.
• empirical-voice: optional writing in your own saved Empirical writing voice (needs a voice saved with the paid Writing Voice interview).

Each capability is a separate marketplace plugin, so users can install or remove optional capabilities without changing the core memory connection.

Package documentation:

• Core Empirical Memory
• Work History
• Scar Tissue
• Todo
• Voice

Install

Claude Code

claude plugin marketplace add mattgauzza/empirical-memory-plugin
claude plugin install empirical-memory --scope user
claude plugin install empirical-voice@empirical-memory --scope user   # optional, writing voice

Codex CLI

codex marketplace add mattgauzza/empirical-memory-plugin
codex plugin add empirical-memory@empirical-memory

or browse/install via the /plugins command.

GitHub Copilot CLI

copilot plugin marketplace add mattgauzza/empirical-memory-plugin
copilot plugin marketplace update empirical-memory
copilot plugin install empirical-memory@empirical-memory

or use the /plugin install slash command.

Recommended: install the latest empirical CLI

The MCP tools work standalone, but the Skills prefer the empirical CLI when it is available. The CLI provides the most reliable auth path, direct memory commands, and avoids an extra MCP tool-discovery round trip. It is not bundled with this plugin; install or update the latest version separately:

npm install -g @empirical-gauzza/cli

For a first-time setup, run the interactive installer to connect the CLI and supported coding agents:

empirical install

Installing or updating the latest CLI does not require rerunning empirical install. Use empirical plugin install all when you only want to refresh the native plugin, or let the CLI offer a one-time update when a newer marketplace plugin is detected during interactive use. You can check immediately with empirical plugin check --force.

For normal interactive CLI use without a permanent API key, run empirical auth login to sign in through the browser and save a local OAuth session. Headless bootstrap is still available for automation and external runtimes.

If you only want the plugin's MCP connection, you can skip the CLI installation.

If the CLI is already installed and you only need to install or refresh the native plugin, use:

empirical plugin install codex       # or copilot, claude, or all

Add optional capabilities explicitly with --work-history, --scar-tissue, --todo, --voice, or --all-optional. For example, empirical plugin install claude --voice installs empirical-voice, which writes in your saved Empirical writing voice (a paid feature; it needs a voice saved first). You can also use the native marketplace command: claude plugin install empirical-voice@empirical-memory --scope user.

This does not run the full MCP/auth questionnaire or modify global instruction files.

Optional capabilities selected by the empirical install questionnaire are installed through the same marketplace pathway. Neither records raw transcripts or secrets. Personalized answers are stored as the server-side Empirical policy and cached locally for native skills; they are not written into global AGENTS.md, CLAUDE.md, or Copilot instruction files.

Mode-aware skill content

skills/empirical-memory/ contains four files, not one:

• SKILL.mcp.md — for installs where the user chose MCP-only. Leads with MCP tools, mentions the CLI only as an aside.
• SKILL.cli.md — for installs where the user chose CLI-direct. Leads with empirical CLI commands as the confidently-recommended path, MCP mentioned only as a fallback.
• SKILL.both.md — for installs where the user chose both. CLI preferred where available (a latency/scripting preference, not a safety one now that both paths dedupe and auto-link the same way), MCP as fallback.
• SKILL.md — the file every host's native plugin mechanism actually reads (Claude Code, Codex CLI, and GitHub Copilot CLI all resolve a plugin's skill from the literal SKILL.md filename; none of their manifest schemas support selecting among sibling variant files at install time). Kept identical to SKILL.both.md, since "both" is the safest default for any consumer that hasn't been made mode-aware (an old CLI version, a direct git clone, a host with no CLI installed at all).

The empirical CLI is what actually makes the mode choice matter: after a native plugin install or update, it overwrites the on-disk SKILL.md for that host with the variant matching the user's chosen integration mode (see packages/cli/lib/installers/nativePlugin.js in the main empirical_v2 repo). If that overwrite step is missing, out of date, or fails, every install still gets a working "both" skill from this repo — never a missing or broken one.

When editing skill content: update SKILL.mcp.md, SKILL.cli.md, and SKILL.both.md for the mode-specific "how to query" / "how to record" mechanics; keep the shared policy sections (Before doing work, Follow the current user policy, Persist durable context, Authored Skills and Kits, Authentication and failures) consistent across all three by hand — there is no build step that enforces this. After editing SKILL.both.md, copy it over SKILL.md so the two never drift:

cp skills/empirical-memory/SKILL.both.md skills/empirical-memory/SKILL.md

Maintaining the manifests

This repo publishes to four hosts, each with its own manifest format. All of them are generated from a single source of truth, meta/manifests.json:

Generated file	Host
.claude-plugin/plugin.json, .claude-plugin/marketplace.json	Claude Code
.codex-plugin/plugin.json	Codex CLI
.github/plugin/marketplace.json	GitHub Copilot CLI
.agents/plugins/marketplace.json	generic AGENTS.md

plus a .claude-plugin/ and .codex-plugin/ pair under each directory in plugins/.

Do not hand-edit those files. Edit meta/manifests.json, then regenerate:

node scripts/build-manifests.mjs           # write the manifests
node scripts/build-manifests.mjs --check    # verify nothing drifted (CI runs this)
claude plugin validate . --strict           # same validator the plugin directory runs
npm install --no-save yaml && node scripts/check-skills.mjs   # every SKILL.md frontmatter must be valid YAML

To bump a version, change it in one place: the plugin's version in meta/manifests.json. The marketplace metadata.version follows the root plugin automatically.

Host-specific fields belong to their host. interface is a Codex/Copilot block; Claude Code silently ignores unknown top-level fields, so putting interface in a .claude-plugin/ file does nothing. Claude's equivalents are displayName, description, and homepage, and it has no counterpart for longDescription, category, capabilities, or defaultPrompt. The generator encodes this mapping.

License

MIT