mempalace/mempalace: A Local Memory Palace for Claude Code

What is the mempalace/mempalace plugin?

mempalace/mempalace is the repository behind MemPalace, a local-first memory system for AI coding agents. Per its own summary, it stores memories verbatim rather than compressing them into a lossy summary, keeps its storage backend pluggable, reports 96.6% R@5 raw on the LongMemEval benchmark, and makes zero calls to an external API to do its own recall. Flowy indexes that claim as written, without independently verifying the benchmark figure. Flowy files the plugin under the data category and indexes 9 components from the repository: 5 commands, 3 skills, and 1 MCP server.

The name is the organizing idea behind the plugin. MemPalace's own command descriptions talk about a memory store divided into wings, rooms, and drawers, a nested structure that mempalace-search can filter by and mempalace-status reports counts for. Flowy's indexed descriptions do not spell out exactly what separates a wing from a room from a drawer beyond that nesting, so treat it as a folder-like hierarchy worth looking at directly rather than something this guide can define precisely.

How do memories get into the palace?

mempalace-mine is the ingestion command. It mines project files and conversation exports into the palace, with auto-classification support, so existing history does not have to be retyped as new memories by hand. The broader mempalace skill covers the same ground at setup time: it is described for installing, configuring, and operating MemPalace, including first-time setup, MCP wiring, mining, status, and the wings, rooms, and drawers structure itself.

How do you get memories back out?

Two components cover retrieval, and they are not interchangeable. mempalace-search is a command you run yourself: semantic search across the palace, with wing and room filtering. mempalace-recall is a skill rather than a command, described to search the palace on its own before Claude answers a question about past work, people, projects, or prior decisions, instead of waiting to be asked. In practice, mempalace-search is the tool for a deliberate lookup, and mempalace-recall is meant to stop an agent from answering a "what did we decide" question from scratch.

Can it hand work between agents?

mempalace-task is the multi-agent piece. Its description covers creating, handing off, claiming, executing, and closing tasks through what it calls the MemPalace logstream: delegating work, preparing a ready-to-paste task for another agent, getting back a MemPalace task id, or launching a supported headless agent in controlled mode. Reach for this when several agents, or several sessions, work the same backlog rather than one agent alone.

What do the setup and status commands cover?

mempalace-init installs the package, initializes a palace, registers the MCP server, and verifies the setup. Its own description names Cursor specifically as the editor it registers against, so check that step matches your own setup rather than assuming it is Claude Code end to end. mempalace-status shows the current state of the palace: wings, rooms, drawer counts, and suggestions. mempalace-help is the in-plugin reference, listing the skills, MCP tools, CLI commands, hooks, and architecture that ship with MemPalace, which is often faster to run than to re-derive from a guide like this one.

What does the MCP server do?

Flowy's index carries no description of its own for the mempalace MCP server. Every other component that mentions MCP treats it as wiring rather than a feature in its own right: something to register during mempalace-init, something the mempalace skill sets up. Read it as the transport the commands and skills above use to talk to a running palace, not as a separate capability to learn on its own.

Private palace or shared brain?

The mempalace skill's own description lists three modes: a private local palace, a shared-brain hub, or a client joining an existing hub. Paired with mempalace-task's logstream, that points to MemPalace being built for more than a single solo agent: a team, or a set of agents working in parallel, can point at one shared palace and one task log instead of each keeping a separate one. The plainer fit is anyone who wants an agent to stop re-explaining the same project history every session, on a backend they control rather than a third-party API.

How do I install it

Flowy indexes mempalace/mempalace from its public GitHub repository. It does not host, bundle, or modify any of these 9 components, so what you get is whatever MemPalace ships. Install from the listing page linked at the top of this guide, and note MemPalace's own summary: it states it has no official website other than that repository, so treat lookalike sites accordingly.