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

# MEMO Memory Storage

MEMO is Aria Compute’s **memory product line**. One end is the open-source, on-device **Aria Memo** (long-term memory CLI/SDK, SQLite local store, cloud / local / both). The other is the cloud-hosted **AriaMem**, a knowledge memory fusion platform. One memory API and `search_memory` join the two ends.

## Memory concepts

| Concept | Meaning |
| - | - |
| Memory | Per-principal facts, preferences, and context fragments (cube / fragment) |
| Cube | Isolation unit for memory: multi-principal, multi-tenant |
| Context fragment | One memory atom (fact / preference / constraint / event) |
| Wiki | Agent-distilled, rollbackable interconnected knowledge doc |
| Knowledge↔Memory conversion | Bidirectional bridge: docs → memory, memory → Wiki |
| Skill | Versioned, shareable execution capability (L1–L4 tiers) |
| Layered Chat Memory (L0–L3) | Conversation → atoms → scenes → persona refinement pipeline |
| Code-Graph | Symbols and call/reference edges parsed after importing a repo |

## Cloud entrypoints

The cloud AriaMem base URLs are:

* International: `https://memo.ariacompute.com`
* China: `https://memo.ariacompute.cn`

AriaMem can also be self-hosted (Docker, port `8010`, path `/ariamem/`); the data plane authenticates with an API key (`Authorization: ApiKey sk_aria_your_memo_api_key`).

## On-device Aria Memo

`aria-memo` is the open-source long-term memory CLI, written in pure Rust. It keeps memory in a local SQLite database (default `./memo.db`, overridden with the global `--db` flag). It has no runtime network dependency, server, or provider key.

Retrieval splits into hybrid `search` (`0.7 × cosine + 0.5 × lexical`, with FTS5 `bm25()` lexical candidate pushdown) and pure-vector `recall`. A four-view relation plane (`semantic` / `temporal` / `causal` / `entity`) connects memories with directed edges; `connect` / `relate` / `relations` / `graph` create and traverse them with bounded depth. `search --graph` adds graph awareness and returns an inspectable `RetrieveTrace`. `search-batch` scores many queries under one lock, and `--json` makes results machine-readable. Valid tiers are `working` / `short_term` / `long_term:*`.

For the full command set, retrieval semantics, and self-management, see the [MEMO CLI guide](https://docs.ariacompute.com/guides/memo-cli.md).

Some SDKs expose a `cloud` / `local` / `both` memory backend setting; that is an SDK-level pluggable backend. The CLI itself defaults to and only uses local SQLite, with no automatic cloud sync at runtime.

## How to run

* **Cloud**: call the cloud entrypoints listed above.
* **On-device**: download the `aria-memo` binary from [Releases](https://github.com/ariacompute/memo/releases) or self-upgrade with `aria-memo upgrade`; configure the release source with `aria-memo setup`. See the [MEMO CLI guide](https://docs.ariacompute.com/guides/memo-cli.md).
* Related: [MEMO product page](https://ariacompute.com/memo), [Playground experience](https://docs.ariacompute.com/guides/playground.md), [AGENT cloud agents](https://docs.ariacompute.com/concepts/agent-cloud.md), [AFM foundation models](https://docs.ariacompute.com/concepts/models.md).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.