> ## Documentation Index > Fetch the complete documentation index at: https://arden.timganiev.com/llms.txt > Use this file to discover all available pages before exploring further. # Concepts > How facts, wiki pages, areas, automations, and agents fit together. Arden is five ideas stacked on each other. Read this once and the rest of the docs stop looking like unrelated features. ## The stack | Layer | What it is | Guide | | --------------- | ---------------------------------------------------------------------- | ---------------------------------- | | **Facts** | The correctness layer: durable, source-backed, event-sourced knowledge | [Memory](/guides/memory) | | **Wiki pages** | The readable layer: versioned Markdown that cites facts | [Memory](/guides/memory) | | **Areas** | Life domains, each optionally attached to one managed page | [Areas](/guides/areas) | | **Automations** | The engine: triggers that run an agent on their own | [Automations](/guides/automations) | | **Agents** | The workers: chat, subagents, workflows, background runs | [Workflows](/guides/workflows) | Each layer only depends on the one below it. Areas do not have their own storage – they attach a wiki page. Area agents are not a separate runtime – they are automations with a narrow tool scope. ## Facts vs wiki pages The split exists because two different jobs were fighting inside one store. **Facts** answer *"what is true, and how do we know?"* A fact has a stable ID, a lifecycle, provenance, and immutable event history. Nothing edits a fact in place: Arden prepares an exact change plan, then commits it only if the affected facts haven't moved underneath it. Plans can create, review, amend, supersede, expire, or retract. **Wiki pages** answer *"what should I read to understand this?"* They're Markdown, versioned, and organized in directories. A page cites the facts it rests on rather than duplicating them, so correcting a fact doesn't require losing its evidence trail. Memory Synthesis updates page prose when the facts behind it change. The rule of thumb: if you'd want to see the receipt, it's a fact. If you'd want to read it over coffee, it's a page. ## Where the proactivity comes from Nothing above acts on its own. **Automations** are the only thing that does. An automation is a trigger plus a prompt. Triggers can be scheduled, interval, event-driven, idle, message-count, or a Slack message. When one fires, it runs an agent – by default in a durable **channel session** whose transcript you can open, read, and reply to. Two things constrain what that agent may do: * A **tool scope**: an allowlist on top of a read-only floor. No scope means read-only, regardless of what the prompt asks for. * **Approval**: mutating or external actions stop and wait for you unless auto-approve is on for actions the scope already granted. ## What an area adds An automation on its own still reports into a transcript you have to go read. An **area** is what turns that into something you don't have to check. An area is a domain – health, research, a job search, or a side project. It may attach one managed wiki page and delegate a standing custodian. The custodian is a channel automation with an observe-only scope and one terminal action, `submit_area_report`, which atomically commits its work, evidence, at most three salient asks, and when to look again. Every run re-decides the area's requests against stable keys, so you get a current shortlist instead of a duplicate backlog. Home brings requests and agent activity together. **A quiet area is a good outcome, not a gap.** ## Which agent surface to use Arden spawns agents five ways, and picking the wrong one is the most common mistake: | Surface | Use when | | ------------------- | ------------------------------------------------------------- | | `research` | One investigation you want to wait for | | `background` | One detached investigation that delivers itself when done | | `workflow` | A curated multi-agent pipeline (find→verify, diverge→judge) | | `create_automation` | Recurring scheduled or event-driven work | | `create_loop` | Adaptive repeated work in this chat that stops on a condition | See [Workflows and subagents](/guides/workflows) for what each one actually does. ## Everything runs locally The server, the desktop client, the facts database, and the wiki all live on your machine. Arden talks to LLM providers and to the integrations you connect; it has no backend of its own. Facts, wiki pages, and how recall works. Life domains and standing custodians. Triggers, channels, and tool scoping. Subagents, presets, and the spawn surfaces.