Nocturne Memory

Install

Manual install

git clone https://github.com/Dataojitori/nocturne_memory.git
cd nocturne_memory
pip install -r backend/requirements.txt

This is third-party code. Review the repository files before installing.

What it does

The server keeps an agent's long-term memory separate from the model and serves it over MCP, so the same records are reachable from different clients and models. Memory is a tree of records addressed as memory://, each with content, metadata and a condition for when it should be recalled. In a new session the agent loads only the specified root records and pulls the rest on demand through the read_memory and search_memory tools. Every change the agent makes is stored as a separate version that can be diffed in a visual dashboard and accepted or rolled back. Data lives in SQLite or PostgreSQL, and namespace isolation supports several independent memory sets.

Who it is for. For developers and enthusiasts who want a shared long-term agent memory that carries across models and clients.

Example request

Read system://boot and remind me what we agreed on last time

Limitations

You deploy the server yourself: it needs Python 3.10 or newer, the backend dependencies, and Node.js for the first build of the dashboard. The public misaligned.top endpoint is a read-only demo; run your own instance for write access. The main README is in Chinese, with the English version in README_EN.md. Some README examples contain explicit role-play content.

How to disable. Remove the nocturne_memory entry from the client's MCP configuration and stop the local server. Memory data stays in the database file until you remove it yourself.

MCP

Transport

stdio, sse

Authentication

not required

Security check

• Runs a local server and database on the user's machine
• Reads and writes persistent memory, with the agent making changes itself

README in short

The README presents an MCP-based long-term memory server for agents that stores records apart from the model so one memory set works across clients. Memory is a tree of addressable records with recall conditions; a new session loads only the root records and fetches the rest through read and search tools. Installation is two steps: clone with pip install of the backend dependencies, then add to the MCP config a python command pointing at mcp_server.py. It also documents a visual dashboard with the memory tree, version diffs and rollback, plus a read-only demo server. Data is stored in SQLite or PostgreSQL, MIT licensed.

Related