What is braincell?

Braincell is a memory tool for AI coding assistants. It runs entirely on your own machine and stores everything in a local SQLite database file, so no data leaves your computer. The idea is to give each project its own "brain": a searchable store of documents, conversation transcripts, and notes that an AI assistant can consult when working on your code. You interact with it through a command line tool, a desktop application called the Memory Map, or directly through compatible AI clients like Claude Code or VS Code. The core operations are simple verbs: remember, recall, supersede, and forget. You can save a note to a project's brain, search for relevant notes later, replace an outdated note with a new one, or remove a note entirely. Search combines two methods: keyword matching (finding literal words) and vector similarity (finding conceptually related content, powered by a local model called Ollama). The system blends both approaches so results match both exact terms and broader meaning. Curated notes also factor in confidence and recency, so older notes gradually carry less weight but are never deleted on their own. Version 0.3.0 upgrades the Memory Map into a native desktop application built with Qt. A single command launches it, and closing the window cleanly shuts everything down. The update also fixes several stability issues: switching the underlying embedding model no longer leaves the database in a broken state, and sparse new projects can return useful results by pulling ranked excerpts from transcripts until enough curated notes exist. Under the hood, each project brain is one SQLite file containing document chunks, memory notes, a graph of links between related notes, and guard tables for schema versioning. Vector search is deliberately simple: it compares all stored vectors directly using NumPy, which is fast enough for the scale of data a single project produces. The project authors chose this approach over a more complex indexed search, and they include monitoring so you can see if your project ever grows large enough to need a faster method. Notes that have been replaced are never returned as current answers, but they still participate in matching because older wording often aligns with how someone might phrase a query. The full README is longer than what was shown.