ModelBrain Adds Local Reasoning and a Linux Package to Its Free Local Memory Layer for AI Assistants

A model sized to your computer reads your notes and cites its sources. ModelBrain now runs on macOS, Windows and Linux and works with Claude Desktop, Claude Code, Cursor and ChatGPT over the Model Context Protocol.

Atlanta, GA · October 6, 2026 · For immediate release

NexSpark LLC today announced the release of ModelBrain, a free memory layer for AI assistants that runs on the user's own computer. Assistants that support the Model Context Protocol (MCP) can remember, recall, expand and forget information through the same local store, so something saved in one assistant is available in the others.

This release adds local reasoning. Instead of only returning matching notes, ModelBrain can read the relevant notes with a language model that runs on the user's own hardware, answer the question, and cite which notes the answer came from. A model sized to the computer downloads once, at roughly 2.5 to 5 GB, and then runs locally with no API key and no per-question cost.

ModelBrain stores memories as ordinary files on the user's machine. For local use there is no account to create and nothing is sent to a server. Memories are never used as training data.

"People switch between Claude Desktop, Claude Code, Cursor and ChatGPT during a single day, and each one starts from zero. ModelBrain gives them one shared memory over MCP, so a decision you record in Cursor is there when you open Claude Code an hour later."

Matthew Firth, founder of ModelBrain

ModelBrain registers itself with Claude Desktop automatically, offers one-button setup for Claude Code and Cursor, and connects to ChatGPT through an OAuth relay. The macOS build for Apple Silicon is notarized and signed. The Windows installers for x64 and ARM64 are not yet code-signed, so Windows SmartScreen shows a warning on first run. The Linux package is a command-line install for x86_64 and ARM64 on glibc 2.28 or newer (Debian 10+, Ubuntu 20.04+, RHEL 8+) that registers with Claude Code and Cursor.

ModelBrain publishes how it tests its own recall. In a 285-row adversarial evaluation run against the production prompt path, with pass and fail criteria written in advance, it scored an 86% right-answer rate and 100% on prompt-injection resistance. The same evaluation reports a 70% strict accuracy figure and a 79% score for its weakest category. The method and results are on the fact sheet and in What "accurate" should mean for an AI's memory.

ModelBrain is available now on the download page. It is also listed in the official MCP Registry as net.modelbrain/mcp, and the launcher is on npm as modelbrain-mcp.