Mempalace

MemPalace Setup

A guided, skill-first setup for a searchable memory palace. The user may have installed this skill with npx skills add before the MemPalace Python package or MCP server exists; that is the normal bootstrap path.

Setup protocol

1. Inspect before changing anything

• Detect the OS and current agent harness.
• Run mempalace --version, uv --version, and an appropriate Python version check. Do not assume that an installed Python package is reachable on PATH.
• Check for an existing palace and MCP registration. Never reinitialize or rebuild an existing palace just to make setup simpler.

2. Install the CLI when necessary

Prefer an isolated uv tool installation:

uv tool install mempalace

If uv is unavailable, use the PATH-visible Python installation:

python -m pip install mempalace

After installation, run mempalace --version. If it still is not reachable, fix PATH or use the matching uv tool run invocation before continuing.

3. Choose the topology with the user

Ask which outcome they want unless it is already clear:

1. private local palace — one machine, local stdio MCP;
2. shared-brain hub — this machine owns the palace and serves the fleet;
3. client joining an existing hub — this machine connects to a hub owned elsewhere.

Also ask which project or conversation corpus should be initialized, offering the current working directory as the default. A shared-brain client does not initialize a second copy of the owner's palace.

4. Run version-correct initialization

MemPalace provides dynamic, version-correct instructions via the CLI. To get instructions for any operation:

mempalace instructions <command>

Where <command> is one of: help, init, mine, search, status.

Run the appropriate instructions command, then follow the returned instructions step by step.

For a new local palace or hub, follow mempalace instructions init, configure the selected corpus, then verify with mempalace status. For a client, skip local initialization and obtain the hub URL and bearer token from the user.

5. Configure MCP

For local stdio integrations, use the command printed by mempalace mcp. Typical registrations are:

claude mcp add mempalace -- mempalace-mcpcodex mcp add mempalace -- mempalace-mcp

For a shared-brain hub, guide the user through mempalace serve and the official shared-brain guide. Do not expose a non-loopback server without authentication. For a client joining an existing hub, configure the harness's HTTP MCP transport with the supplied bearer token; never print or store that token in project instructions, drawers, or logstream events.

Restart or reconnect the harness when required, then verify that the live MCP tool list includes MemPalace tools. Package installation alone is not proof that MCP is connected.

6. Configure shared-brain identity and coordination

When shared-brain mode is selected:

• Agree on a stable host:harness:project identity: lowercase host label (machine), harness family (claude, codex, grok, antigravity, …), and the current workspace as project. Two windows in the same project are one actor.

• Render the canonical rules with:

  bash

  mempalace rules --host <host> --harness <harness> --project <example>

  Default --mcp full matches the 47-tool mempalace-mcp server this skill registers. If the user opted into mempalace-light-mcp, re-render with --mcp light instead. Replace an existing <!-- mempalace-shared-brain --> block instead of appending a duplicate.

• Install the rendered marker-delimited block in the harness's durable agent instructions (Claude ~/.claude/CLAUDE.md, Codex ~/.codex/AGENTS.md, Grok ~/.grok/AGENTS.md, Antigravity ~/.gemini/config/GEMINI.md).

• Check coordination access with a read-only mempalace logstream list or the equivalent MCP event-list call.

• Interactive sessions are declared-idle: they sweep the inbox on collab / before long tasks and do not arm a watcher at session start. Arm mempalace logstream watch --agent <host>:<harness>:<project> (the CLI defaults a sanitized --state-file) only when the user asked to listen, the agent claimed a task, or it delegated. A remote-only MCP client must instead loop on mempalace_event_wait, preserving the last event id as since_event_id; never point it at a local SQLite watcher. Explain any permission allowlisting needed. If it cannot maintain either loop, record that the agent is turn-based and must sweep its MCP inbox with mempalace_event_list on wake-up.

Do not post a test event without telling the user: logstream events are immutable. If the user approves a smoke event, address it narrowly and close the loop with an acknowledgement.

7. Report readiness

Summarize the installed version, palace location or hub URL (without secrets), MCP connection, stable agent identity, watcher mode, and the first safe next action. For active delegation, hand off to the mempalace-task skill.

Ask whether the user wants weekly stable-release checks. The default is no. Explain that enabling them contacts PyPI but sends no palace content, identity, or telemetry. When enabling, record the installer actually used with mempalace update configure --enable --installer uv-tool (or pipx / pip); use --disable to opt out. Checks never install anything. In mempalace_status, treat updates.server as the palace-serving runtime and updates.client (when present) as the local proxy runtime; do not conflate their versions or installers. For a client update, use the local mempalace update plan. A remote server update is informational on the client: surface it naturally and ask the hub operator to prepare and authorize the plan on the palace-serving machine. Never use a client-generated plan to upgrade the server, and never execute any plan without explicit approval.

Palace health: audit and repair session

When the user asks how well organized the palace is, whether memory is "messy", why a scoped search or wake-up misses things, or invokes /mempalace:audit, run the audit and then offer a repair session:

mempalace instructions audit

Follow the returned instructions. In short: run mempalace audit --json (read-only, safe while the MCP server is running), present the five layer scores and findings, then walk the user through repairs one structured question at a time with a recommended option first: merging wings and rooms spelled two ways, folding stub wings, deleting tunnels on generic tokens and self-link hallways, agreeing a knowledge-graph predicate vocabulary, and giving flat wings a closed room set with mempalace rooms propose / apply. Moves over deletions, numbers before actions, verbatim content always. Re-run the audit at the end and write a diary entry with the before and after scores and every decision made.

Recalling past work

This skill covers setup, mining, and status. For questions about past work, prior decisions, or people that may already be filed in the palace, prefer the mempalace-recall skill — it enforces search-before-answer so the agent reads the palace instead of guessing.

Cursor-specific notes

• The Cursor plugin auto-registers mempalace-mcp; a standalone npx skills add installation does not. Always verify the live tool list.
• For automatic background saving every N agent turns plus session-start memory recall, also install the Cursor hooks separately by running hooks/cursor/install.sh --scope user from a cloned MemPalace repo. See the Cursor hooks guide for the full walkthrough.
• The recommended agent_name when calling mempalace_diary_write from a Cursor session is cursor-ide (matches the precedent of claude-code and codex).

Canonical references

• Shared brain
• Coordination protocol
• Recall protocol