Why my AI assistant's memory stays on my machine
AI assistants have a memory problem, and it is not that they forget. It is that they hide the context they picked, and they scatter your history across chats. The durable layer — everything you have told an assistant over months — ends up owned by whichever product you happened to use. Singularity is my attempt to move that layer back onto my own machine.
What it does
Singularity is a local memory and permission layer for AI assistants. It imports your chat history and project files and answers from them with citations you can open to the exact source range — not "based on your documents," but this file, these lines.
The redaction gate
The part I care about most: every outbound model call is held behind a redaction gate that I approve before anything leaves the machine. The assistant can propose what to send; it cannot decide on its own. Every import, call, approval, denial, and answer lands on a single auditable SQLite timeline, so "what did it share, and when?" always has an answer.
Boring technology on purpose
It is built in TypeScript on Fastify and React, with SQLite FTS5 for retrieval. It is model-agnostic: Ollama, any OpenAI-compatible endpoint, or a fully offline mock. Nothing in the stack requires an account or a cloud service to keep the memory itself working.
The lesson
Keeping memory local makes two things inspectable at once: what the assistant knows, and what it is about to share. That is the same instinct behind my PII detection work — privacy holds when the architecture enforces it, not when a setting promises it.