> ## Documentation Index
> Fetch the complete documentation index at: https://docs.memanto.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Requirements

> Hardware, OS, and software prerequisites for running Memanto on-prem.

# On-Prem Requirements

The on-prem stack runs Memanto plus the Moorcheh server (and optionally Ollama) in Docker on a single host. Everything is automated by the `memanto` CLI — you only need to make sure the prerequisites are in place before you start.

## Operating Systems

The on-prem stack is supported on:

* **Windows 10/11** with Docker Desktop (WSL2 backend)
* **macOS 12+** (Apple Silicon and Intel) with Docker Desktop
* **Linux** (Ubuntu 20.04+, Debian 11+, RHEL 8+, Amazon Linux 2) with Docker Engine 20.10+

## Hardware

| Component | Minimum      | Recommended                | Notes                                                                                                                        |
| --------- | ------------ | -------------------------- | ---------------------------------------------------------------------------------------------------------------------------- |
| **CPU**   | 4 cores      | 8+ cores                   | More cores noticeably speed up Ollama inference.                                                                             |
| **RAM**   | 8 GB         | 16 GB+                     | Ollama with `qwen2.5` needs \~6 GB resident; embeddings add \~1 GB. With OpenAI/Cohere as providers, 4–6 GB total is enough. |
| **Disk**  | 10 GB free   | 30 GB+ free                | Ollama model images are 1–7 GB each. Moorcheh storage grows with your memory volume.                                         |
| **GPU**   | Not required | NVIDIA GPU with 8 GB+ VRAM | Optional — speeds up Ollama. CPU-only inference works on any modern machine.                                                 |

For air-gapped deployments using OpenAI or Cohere as the embedding/LLM provider, the hardware footprint is much smaller (no Ollama container needed).

## Software

### Required

* **Docker Engine 20.10+** or **Docker Desktop 4.0+** with the daemon running.
  * The Memanto onboarding wizard fails fast with a clear error if `docker info` does not succeed.
  * Verify with:
    ```bash theme={null}
    docker --version
    docker info
    ```
* **Python 3.10+** for the Memanto CLI itself.
  * Verify with:
    ```bash theme={null}
    python --version
    ```
* **`memanto`** Python package.
  ```bash theme={null}
  pip install memanto
  ```
* **`moorcheh-client>=0.1.3`** — the Python package that ships the `moorcheh up` command and exposes the on-prem SDK shape Memanto talks to. The onboarding wizard installs this automatically the first time you choose **On-Prem** at the prompt; you can also install it explicitly:
  ```bash theme={null}
  pip install "moorcheh-client>=0.1.3"
  ```

### Optional

* **`uvicorn[standard]`** if you plan to run `memanto serve` directly. Installed automatically as a dependency of `memanto` in most cases.
* **NVIDIA Container Toolkit** if you want Ollama to use a GPU inside Docker.

## Network & Ports

| Port      | Bound to                  | Used by                                                                                           |
| --------- | ------------------------- | ------------------------------------------------------------------------------------------------- |
| **8000**  | `localhost` (default)     | Memanto's REST API (`memanto serve`, `memanto ui`).                                               |
| **8080**  | `localhost` (default)     | Moorcheh on-prem server, started by `moorcheh up`.                                                |
| **11434** | inside the Docker network | Ollama, when used as the embedding/LLM provider. Started as a sibling container by `moorcheh up`. |

You do **not** need any inbound internet access for the runtime path. Internet is required only:

* Once, to `pip install memanto` and `moorcheh-client`.
* Once per Ollama model, to pull the image from the Ollama registry.
* For every `answer.generate` call if your LLM provider is OpenAI or Cohere.

## Provider Choices

You will be prompted to choose providers during onboarding. The choices and what they imply:

| Provider   | Embedding                | LLM (Answer)             | API key  | Cost      | Best for                                                          |
| ---------- | ------------------------ | ------------------------ | -------- | --------- | ----------------------------------------------------------------- |
| **Ollama** | `nomic-embed-text`       | `qwen2.5`                | None     | \$0       | True air-gap; local development; demos; cost-sensitive workloads. |
| **OpenAI** | `text-embedding-3-small` | `gpt-4o-mini`            | Required | Per-token | High-quality embeddings; existing OpenAI relationships.           |
| **Cohere** | `embed-english-v3.0`     | `command-r-plus-08-2024` | Required | Per-token | High-quality long-context answers; multilingual embeddings.       |

You can mix providers — e.g., **Ollama embeddings** with an **OpenAI LLM** for answers. The onboarding wizard prompts for each independently.

## Disk Layout

Once onboarding finishes, the on-prem stack uses these locations on the host:

| Path                            | Owner                      | Purpose                                                                                                     |
| ------------------------------- | -------------------------- | ----------------------------------------------------------------------------------------------------------- |
| `~/.memanto/`                   | Memanto CLI                | Top-level config dir. The cloud backend stores everything here.                                             |
| `~/.memanto/on-prem/`           | Memanto CLI (on-prem only) | Isolated data dir: agents, sessions, registry, and `state.json` for the on-prem backend.                    |
| `~/.memanto/on-prem/state.json` | Memanto CLI                | Source of truth for `url`, `embedding_provider`, `embedding_model`, `llm_provider`, `llm_model`.            |
| `~/.moorcheh/config.json`       | `moorcheh-client`          | Embedding and LLM provider config consumed by the on-prem server.                                           |
| `~/.moorcheh/uploads/`          | `moorcheh-client`          | Staging area for files uploaded via `memanto upload` — paths inside this dir are mapped into the container. |

Cloud and on-prem state are deliberately kept in separate directories so you can switch backends without one polluting the other.

## Verifying Prerequisites

Before running the on-prem setup, the wizard performs these checks for you:

1. `docker` is on `PATH`.
2. `docker info` returns successfully (daemon is up).
3. `moorcheh-client>=0.1.3` is importable (installs it if not).
4. Provider API keys (if you chose OpenAI or Cohere) are non-empty.

If any check fails, the wizard prints a one-line error with a hint and exits with a non-zero status — no partial state is left behind.

## Next Step

→ [On-Prem Quickstart](./quickstart) — install the stack end-to-end in 5–10 minutes.