An AI Assistant That Actually Learns How You Work Over Time

Ask any AI assistant to help you with a task today. Then ask it the same thing tomorrow. It will have no memory of yesterday's conversation, your preferences, or the decisions you already made. You start from zero every single time.

This is the biggest gap in current AI tools. They are capable but amnesiac. Every session is a first date where you re-introduce yourself, re-explain your preferences, and re-establish context.

What Persistent Learning Looks Like

Imagine an AI assistant that builds a knowledge graph of how you work. It notices that you always check Slack before email in the morning. It learns which contacts you respond to immediately versus which ones can wait. It figures out that when you open Figma on Tuesdays, you are working on the weekly design review.

After two weeks of observation, the assistant starts anticipating. It surfaces your Slack messages before you open the app. It drafts responses to routine emails using your writing style. It pre-loads the design files you will need for Tuesday's review.

This is not science fiction. The technical components already exist - knowledge graphs, embedding-based memory, local observation APIs. The challenge is assembling them into something that actually works without being creepy or intrusive.

The Privacy Requirement

For this kind of learning to be acceptable, it has to be local. Your work patterns, your contacts, your habits - this is deeply personal data. Sending it to a cloud server for processing is a non-starter for most people.

A local knowledge graph stored on your machine gives you the benefits of personalization without the privacy tradeoff. The agent learns from watching how you use your Mac, but that data never leaves your device. You can inspect it, edit it, or delete it at any time.

The Two-Week Threshold

Our experience suggests that two weeks is the threshold where a learning assistant becomes genuinely useful. Before that, it is still gathering patterns. After that, it has enough context to start making predictions that save real time. The assistant stops being a tool you use and starts being a colleague that knows your workflow.

Fazm is an open source macOS AI agent. Open source on GitHub.