ami

Most personal automation tools either require you to hand credentials to a cloud service or demand you wire up every workflow by hand every time. Ami approaches this differently: it acts as a local agent harness that connects to your apps, repositories, and tools using your own personal tokens, then builds a live context graph — stored in SQLite and markdown under ~/.ami/ — that it updates continuously as you work. The core loop is: connect a tool, let Ami observe how you handle a task, watch it encode that pattern as memory, and then have it execute the same class of task autonomously going forward.

The standout architectural decision is the context graph memory model. Rather than storing flat conversation history, Ami tracks named entities, relationships between them, your decisions, feedback you’ve given, and even your writing style. The vendor describes this as making the agent ‘more autonomous the more you use it’ — meaning the agent’s behavior adapts to your specific patterns rather than starting from a generic baseline on every session.

Ami fits solo developers and power users who handle repetitive coordination work — triaging issues, drafting responses, managing to-dos across tools — and who are unwilling to route personal credentials through a third-party service. It does not fit teams that need shared agent state, role-based access, or a stable programmatic API: there is no API surface exposed, and the data model is explicitly personal and local. The project is early-stage by any measure, which means missing integrations and undocumented failure modes are the expected experience, not exceptions.

The project is MIT-licensed, built in TypeScript with a pnpm monorepo structure, and is self-hostable on any machine that can run Node. The docs describe tool integration via personal tokens, and the memory store lives entirely on-disk — nothing is transmitted externally by design.