Permagent writes what it learns into a knowledge graph on your own disk. Nothing is discarded at the end of a session, so the agent you use in December still knows what you told it in March.
Every conversation, decision and document you give it lands in a durable memory graph on your disk. It stores typed entities and the real relationships between them, including which people work on which projects.
What that gets you. You stop re-explaining your work. The agent picks up where you left off instead of starting cold each time.
A small model running locally on your machine reviews new memories while you are idle, writes descriptions for them, and folds facts that keep recurring into durable entries keyed to the thing they are about.
What that gets you. Memory stays searchable as it grows, so recall a year in surfaces the right context rather than everything you ever said.
Point it at a project and it maps the directory structure and the symbols in each file, then stores that map in the Brain scoped to that project, the same way documents and notes are stored.
What that gets you. You can ask how a codebase is shaped without the agent re-reading every file, and without paying for the tokens to do it.
Drag any file into the app and it is read on your machine: OCR for images, text extraction for PDFs and documents. No upload step.
What that gets you. You can talk about a document without pasting its contents, and the reading itself costs no model tokens.
Anything you download in the built-in browser lands in a folder Permagent owns, with a record in your local database, instead of vanishing into your Downloads folder.
What that gets you. Files you fetch are already where the agent can find, read and file them, with no trip through Finder.
Every past conversation is listed with its title and when it was last active. Reopen one and it loads back into the chat exactly where it stopped, or rename and delete the ones you are done with.
What that gets you. Older work is something you can return to and tidy up, not a scroll-back you eventually lose.
Everything on this page describes behaviour that ships today. The facts come from the runtime’s own capability brief, the same one the agent reads about itself on every turn, so nothing here is a roadmap item wearing a present tense.
Can an AI agent run without sending your data to the cloud?
Can an AI agent work while you sleep?
What can an AI agent actually do on your computer?
How do you stay in control of an autonomous agent?
What stops an AI agent from going off the rails?
Which AI models can you use, and who pays for them?