Introduction

Axom is an architecture that turns a small, unmodified base model into an assistant that researches, codes, remembers, and works on its own, on your own hardware.

How it's put together

• The chassis. A Gemma 4 model, American-made and unmodified, running in llama.cpp on your GPU (or CPU).
• Leaves. Small LoRA adapters that sit on top of the chassis. Axom's own ops leaf teaches the model how to work: search before answering, cite what it read, use tools cleanly. Skill leaves are grown from your own work: its mistakes in a field become the training for a leaf you own. See Skill leaves.
• Memory. A ledger of everything you and Axom have said. Each turn, Axom compiles a fresh context from it (who you are, this conversation, and what it recalled) sized to fit the window. The window never fills up; the memory never forgets.
• The tool loop. Axom thinks, acts, observes, and repeats until it has an answer. When a step calls for several tools at once, they run as a team, in parallel.
• The Latent Swarm. When a team brings back more than fits, each page goes to its own agent and the swarm reads them together, handing back latent positions instead of text. See Latent Swarm.
• The runtime. One background process per user, serving the terminal, the browser UI, and anything else that speaks its API. It can run several sessions at once on one loaded model.

What stays on your machine

Inference, memory, your files, ratings, and settings all live in your Axom home (~/.axom). Web searches go out through the search path you choose: your own private search engine, a Brave Search key, or nothing at all.

If you use the Axom Network, the model runs on a node instead; your memory and files still stay on your machine, and each turn's prompt travels to the node sealed end to end.

Next

• Installation: what you need and what the installer does
• First run: the setup screens, then your first session