Nex Mempalace Memory System

What it does

• Delivers scripts/setup-mempalace.sh: an idempotent Pi deployment that creates a venv, installs MemPalace (ChromaDB plus SQLite), writes the palace config, pre-warms the embedding model, registers the MCP server with your bot gateway, and restarts it.
• Structures memory as a palace: wings (projects and people) with keyword routing, halls (facts, events, discoveries, preferences, advice), rooms (topics), and drawers (stored exchanges), plus a knowledge graph and an agent diary.
• Ships assets/identity.template.txt, a Layer-0 identity of about 50 tokens loaded every session.
• Ships assets/wing_config.template.json and assets/config.template.json for the palace layout, collection name, and people map.
• Includes assets/memory-protocol.md, the skill that teaches the agent to search before answering, store after conversations, and keep content verbatim.

Why use this

A bot that forgets every conversation is a demo; one that remembers is an assistant. This is the deployment pattern Nex AI uses to give its own agents memory on a self-hosted Pi: one shared palace, multiple bots, no cloud dependency, and a memory protocol that makes the memory actually get used instead of just stored.

See it in action

Prompt:

I want my Telegram assistant to remember our past conversations across sessions.

Result:

python3 -m venv /home/pi/mempalace/venv   # backs up an existing venv first
venv/bin/pip install mempalace            # ChromaDB compiles on ARM64, 3-5 min
mempalace init --config config.json      # wings, halls, rooms, people map

mempalace mine ~/notes --wing projects
mempalace search what did we decide about the invoice flow

Plus a filled identity layer (who this agent serves, about 50 tokens) and the memory protocol wired into the bot.

Use cases

• Give a Telegram or OpenClaw bot memory that survives across sessions
• Run one shared memory palace that multiple bots read and write via MCP
• Self-host a ChromaDB-backed memory layer on ARM64 hardware
• Add an identity layer plus a search-before-answering protocol to an agent
• Mine existing note folders into structured, searchable agent memory

Known limitations

• Defaults target a Raspberry Pi (ARM64) with OpenClaw as the bot gateway; other Linux hosts and gateways work but need the MCP wiring adapted.
• ChromaDB compilation on ARM64 takes 3 to 5 minutes on first install.
• The shipped templates are placeholders; you fill in your own projects, people, and identity, no data ships with the skill.

FAQ

What exactly is in the download?

SKILL.md, README.md, scripts/setup-mempalace.sh, four assets (identity.template.txt, wing_config.template.json, config.template.json, memory-protocol.md), and references on palace structure, deployment, and the memory protocol.

Which agents does it work with?

Claude Code, Claude Cowork, Cursor, Codex CLI, and any agent that reads SKILL.md skills; the deployed memory itself is exposed over MCP to any MCP-capable bot.

Do I need anything else?

A Raspberry Pi or Linux box with Python 3, and a bot gateway that speaks MCP (OpenClaw is the default). No cloud services or API keys for the memory layer itself.

Why the memory protocol file?

Storage alone does nothing if the agent never looks. The protocol instructs it to search before answering when past context might exist and to store important things verbatim afterward; that is what makes the memory effective.