Permagent: An Agent That Lives on Your Machine, Not in Someone Else's Cloud

There is a quiet assumption baked into almost every AI assistant you can buy today: that your data, your context, and your agent's memory belong on a server you don't own. You rent intelligence. The company holds the keys. When the connection drops, the subscription lapses, or the terms change, the assistant you came to rely on is gone — and everything it learned about you went with it.

Permagent is built on the opposite assumption. It is a local-first AI agent desktop OS for macOS Apple Silicon: a downloadable application where the agent, its memory, and your data live on your own hardware. The intelligence works for you because it runs on a machine you control. That single architectural choice — sovereignty by physics, not by policy — changes what an AI assistant can be.

This post is an honest account of where Permagent is today, what it can already do, and the road ahead.

Most "private" assistants offer privacy as a promise — credentials stored separately, data encrypted at rest, a privacy policy you're invited to trust. Permagent offers privacy as a fact of the architecture. Your documents, your memories, and the knowledge graph that represents everything your agent has learned never leave your machine. There is no server to trust because there is no server. Local models handle the always-on background work; the cloud is reached only when you explicitly allow it, for the specific tasks where a frontier model earns its cost.

This is not a limitation dressed up as a virtue. It is a different shape of product. An agent that runs locally can work twenty-four hours a day without metering every thought against an API bill. It can hold a rich, persistent memory of your work because that memory is a database on your disk, not a feature you rent. And it can be genuinely yours.

Permagent is not a concept. It is a working desktop application built on a foundation of components that have been designed, built, verified, and shipped. A few of the pieces that define the experience:

A persistent memory that understands, not just stores. At the heart of Permagent is the Brain — a temporal knowledge graph that captures what your agent learns over time. Unlike a chatbot that forgets you between sessions, Permagent's memory is curated and inspectable. A dedicated worker, the Librarian, runs locally to describe and organize memories so they can be recalled with precision. You can open the Brain and see what your agent knows. Memory is not a black box; it is a place you can visit.

An agent you can actually trust with authority. The hardest problem in autonomous AI is not capability — it is trust. An agent that can act on your behalf can also act wrongly on your behalf. Permagent's answer is a verified loop: work is governed by risk tiers, routine actions proceed automatically, consequential ones are surfaced for your approval, and every completed task carries machine-verifiable evidence and an append-only audit trail. When you wake up to a summary of what your agent handled overnight, you are not asked to take its word for it. You can see what was done, why, and that it was checked. Most autonomous assistants show you a log of actions taken. Permagent shows you a record you can verify.

A named orchestrator that delegates rather than does. Permagent's agent layer is organized around an orchestrator — a coordinator that breaks work into goals and routes each to the right specialist, rather than trying to do everything itself. Specialists like the Librarian handle memory; other workers handle code, documents, and research. The principle is simple: the orchestrator orchestrates, and rarely acts directly. This is how the system scales from a single task to a managed workflow without becoming a tangle.

A world you can watch your agents work in. Permagent is not a chat box. It includes a living three-dimensional environment — a cave that your agents carve from darkness toward light, built from the real events of your own work. As your Brain grows, the world grows with it. Memories become structure. Agent activity becomes motion you can see. It is an interface that makes the abstract work of an AI agent tangible, and it is unlike anything else in the category.

Tools that meet you where the work is. Permagent connects to a browser, spreadsheets, presentations, and a design canvas, and it can be driven from the desktop or reached remotely. The agent does not live in a silo; it operates in the surfaces where your work actually happens.

A theme runs through every part of Permagent: the right model for the right job. The local model — fast, free, private, tireless — handles the high-volume background work: describing memories, processing documents, the steady heartbeat of an agent that is always on. Frontier models, reached deliberately, handle the genuinely hard reasoning. The system escalates only when the stakes justify the cost, and it does so observably — with logs, status, and verifiable evidence rather than silent automation you have to hope is working.

This is what separates a tool that augments your judgment from one that asks you to surrender it. Permagent is built to be the former.

Permagent is on a path toward a public launch, and the work in front of us is concrete.

Sharpening the first experience. The strongest agent in the world fails if the first thirty seconds are confusing. We are building a guided onboarding that seeds your Brain from the start, an auto-updater that keeps the app and its daemon in lockstep, and the diagnostic tooling that makes a shipped product trustworthy in the field. The goal is an arc, not a setup screen: from the first run, the relationship between you and your agent should visibly deepen.

Lighting up the autonomous loop. The trust machinery — the verified decisions, the risk tiers, the audit trail — is built and waiting. The next milestone is bringing the full orchestration loop online, so that the agent can carry goals from start to finish under genuine supervision: deciding what it can do alone, what needs your sign-off, and proving its work either way.

Code quality as a competitive edge. AI can write code fast; it often writes code badly. We are building a quality pipeline that no single-model tool can match: a tireless local reviewer that sweeps projects overnight for bugs and surfaces candidates, and a cross-model "council" where independent frontier models review each other's work adversarially. The insight that drives it is counterintuitive but well-founded — diversity of models, not consensus, is what catches the errors a single model rubber-stamps. The aim is a product whose own code is continuously, demonstrably clean, and a tool that can bring that same standard to your work.

A bigger brain, on your terms. Your agent is only as capable as the model your hardware can run — and we have measured, precisely, where that ceiling sits and how to raise it. Near-term, speculative-decoding engines on Apple Silicon make substantially larger models run at interactive speed on a single dedicated machine. The path to the most capable local models runs through a bigger dedicated anchor in your own home, not through renting someone else's servers.

The Mesh. Furthest out, and most ambitious: a network where Permagent members pool the latency-tolerant capacity of their own machines so that everyone's background agents can transcend their individual hardware. Your agent's overnight work could run on a model far larger than any single member owns — privately, verifiably, and on infrastructure that belongs to the people using it rather than a corporation renting it back to them. We have studied the physics of this carefully and honestly: the interactive, real-time version waits on a specific technique maturing in the open-source engine layer, which we are tracking closely. But the collective-batch version — where your agent gets smarter while you sleep, on a network of trusted machines — is a real and defensible future.

The dominant model of AI assistance asks you to move your life onto someone else's computer and trust them to be good stewards of it. Permagent makes a different bet: that the most powerful, most personal, most trustworthy assistant is one that lives where you do, learns what you teach it, acts only with proof, and belongs to no one but you.

Your hardware determines what your agent can do — and with Permagent, your hardware is yours. That is not a constraint. With the Mesh, it becomes a collective advantage. The local-first promise, fully realized, is not a smaller version of the cloud. It is a better foundation for the thing the cloud was always supposed to deliver: an intelligence that works for you.

We are building toward that, deliberately, one verified step at a time.

Spectral, the memory engine that gives Permagent its persistent recall, has its own story—and a wager that honesty beats hype: Why We're Building Spectral.

Permagent is in active development toward a public release. It runs on macOS Apple Silicon as a free downloadable application, with optional paid services for cloud-model brokerage and, eventually, the Mesh.

Permagent and its memory engine, Spectral, are open source under the Apache 2.0 license. The code is on GitHub: permagent-runtime and spectral.