Where Ollama stores Modelfiles, models, and system prompts

Ollama is different from the other tools on this list: it doesn’t read rules or skills from your projects. Its one context artifact is the Modelfile — a recipe that bakes a base model, a system prompt, and parameters into a named local model. That makes a Modelfile the Ollama equivalent of an agent persona, and it has the strangest storage story of any tool here: after you build from it, the source file isn’t retained anywhere. ~ is your home directory.

At a glance

Artifact	Global	Per-project	Format
Modelfiles (persona source)	wherever you keep them — often a repo	Modelfile, *.modelfile	Modelfile (FROM/SYSTEM/PARAMETER)
Built models	~/.ollama/models/ (manifests/ + blobs/)	—	content-addressed layers
System prompts	baked into the built model’s layers	—	recover with ollama show
Models dir override	OLLAMA_MODELS env var (server-side)	—	path
Signing key	~/.ollama/id_ed25519(.pub)	—	SSH keypair — never index

Modelfiles as agent personas

A Modelfile is a plain-text recipe. FROM (required) names the base model; SYSTEM sets the persona; PARAMETER, TEMPLATE, ADAPTER, LICENSE, and MESSAGE round it out:

FROM llama3
SYSTEM """You are a terse code reviewer. Point at the bug, cite the line,
suggest the one-line fix. No essays."""
PARAMETER temperature 0.3

Build it into a named model with ollama create reviewer -f ./Modelfile, then run ollama run reviewer. Functionally this is the same artifact as a Claude Code subagent or a Cursor rule — a reusable system prompt bound to a model — which is why it belongs in the same inventory.

The vanishing Modelfile

ollama create does not keep your Modelfile

Building a model copies its content into content-addressed layers under ~/.ollama/models/ — the source Modelfile itself is not retained or referenced afterward. Delete the file (or run create from a temp dir) and the only copy of that persona lives inside a blob store nothing else can read. Recover it with ollama show <model> --modelfile, which reconstructs the Modelfile — including the SYSTEM prompt — from the stored layers.

The practical habit: keep Modelfiles in a git repo like any other prompt artifact, and treat the built model as a compiled output.

The models directory

Built models live under an OCI-registry-style layout:

~/.ollama/models/
  manifests/registry.ollama.ai/library/<model>/<tag>
  blobs/sha256-<digest>

Manifests are small JSON files naming a model’s layers; blobs are the sha256-named layers themselves (weights, system prompt, params — shared between models, so two personas built FROM llama3 store the weights once). Default locations differ by OS:

• macOS: ~/.ollama/models
• Windows: C:\Users\<you>\.ollama\models
• Linux (installer/systemd): /usr/share/ollama/.ollama/models

On Linux, it’s not in your home directory

The standard Linux installer runs Ollama as a dedicated ollama user, so models land in /usr/share/ollama/.ollama/models — a backup or scanner that only walks /home misses every model on the machine.

OLLAMA_MODELS — moving the store

Set OLLAMA_MODELS to relocate the models directory (common when a small SSD can’t hold 50 GB of weights). It’s read by the server, not the CLI client — exporting it in your shell does nothing if the background service didn’t get it. Set it where the service runs: launchctl setenv on macOS, systemctl edit ollama.service → Environment="OLLAMA_MODELS=/path" on Linux (the ollama user needs read/write on the new path), or user environment variables + restart on Windows.

Keys

~/.ollama/id_ed25519 / id_ed25519.pub is the keypair Ollama uses to authenticate pushes to a registry. It’s a real SSH-format private key sitting in a directory people casually back up and share — treat ~/.ollama as secret-adjacent, not just big.

Seeing all of it at once

Quartermaster’s scanner picks up Modelfile/*.modelfile wherever they live and indexes each as an agent persona (the FROM base model is recorded in its capabilities), alongside Claude Code, Cursor, Codex, and 8 more tools — so the persona you wrote as a Modelfile is findable next to the subagents and rules that do the same job elsewhere, and convertible between them. Key material is never indexed — the ~/.ollama keypair is hard-skipped, and env/header key names are kept while values are dropped. See the full discovery map or the quickstart.