diff --git a/docs/.vitepress/config.ts b/docs/.vitepress/config.ts
index 79be407b944..7af3aa151dd 100644
--- a/docs/.vitepress/config.ts
+++ b/docs/.vitepress/config.ts
@@ -158,7 +158,7 @@ export default withMermaid({
{ text: "Iconography", link: "/icons" },
{ text: "Local echo", link: "/local-echo-dev" },
{ text: "Media", link: "/media-handling" },
- { text: "Room List Store", link: "/room-list-store" },
+ { text: "Room List", link: "/room-list" },
{ text: "Scrolling", link: "/scrolling" },
{ text: "Usercontent", link: "/usercontent" },
{ text: "Widget layouts", link: "/widget-layouts" },
diff --git a/docs/img/RoomListStore2.png b/docs/img/RoomListStore2.png
deleted file mode 100644
index 9952d1c910f..00000000000
Binary files a/docs/img/RoomListStore2.png and /dev/null differ
diff --git a/docs/room-list-store.md b/docs/room-list-store.md
deleted file mode 100644
index 4e131ee309d..00000000000
--- a/docs/room-list-store.md
+++ /dev/null
@@ -1,165 +0,0 @@
-# Room list sorting
-
-It's so complicated it needs its own README.
-
-
-
-Legend:
-
-- Orange = External event.
-- Purple = Deterministic flow.
-- Green = Algorithm definition.
-- Red = Exit condition/point.
-- Blue = Process definition.
-
-## Algorithms involved
-
-There's two main kinds of algorithms involved in the room list store: list ordering and tag sorting.
-Throughout the code an intentional decision has been made to call them the List Algorithm and Sorting
-Algorithm respectively. The list algorithm determines the primary ordering of a given tag whereas the
-tag sorting defines how rooms within that tag get sorted, at the discretion of the list ordering.
-
-Behaviour of the overall room list (sticky rooms, etc) are determined by the generically-named Algorithm
-class. Here is where much of the coordination from the room list store is done to figure out which list
-algorithm to call, instead of having all the logic in the room list store itself.
-
-Tag sorting is effectively the comparator supplied to the list algorithm. This gives the list algorithm
-the power to decide when and how to apply the tag sorting, if at all. For example, the importance algorithm,
-later described in this document, heavily uses the list ordering behaviour to break the tag into categories.
-Each category then gets sorted by the appropriate tag sorting algorithm.
-
-### Tag sorting algorithm: Alphabetical
-
-When used, rooms in a given tag will be sorted alphabetically, where the alphabet's order is a problem
-for the browser. All we do is a simple string comparison and expect the browser to return something
-useful.
-
-### Tag sorting algorithm: Manual
-
-Manual sorting makes use of the `order` property present on all tags for a room, per the
-[Matrix specification](https://matrix.org/docs/spec/client_server/r0.6.0#room-tagging). Smaller values
-of `order` cause rooms to appear closer to the top of the list.
-
-### Tag sorting algorithm: Recent
-
-Rooms get ordered by the timestamp of the most recent useful message. Usefulness is yet another algorithm
-in the room list system which determines whether an event type is capable of bubbling up in the room list.
-Normally events like room messages, stickers, and room security changes will be considered useful enough
-to cause a shift in time.
-
-Note that this is reliant on the event timestamps of the most recent message. Because Matrix is eventually
-consistent this means that from time to time a room might plummet or skyrocket across the tag due to the
-timestamp contained within the event (generated server-side by the sender's server).
-
-### List ordering algorithm: Natural
-
-This is the easiest of the algorithms to understand because it does essentially nothing. It imposes no
-behavioural changes over the tag sorting algorithm and is by far the simplest way to order a room list.
-Historically, it's been the only option in Element and extremely common in most chat applications due to
-its relative deterministic behaviour.
-
-### List ordering algorithm: Importance
-
-On the other end of the spectrum, this is the most complicated algorithm which exists. There's major
-behavioural changes, and the tag sorting algorithm gets selectively applied depending on circumstances.
-
-Each tag which is not manually ordered gets split into 4 sections or "categories". Manually ordered tags
-simply get the manual sorting algorithm applied to them with no further involvement from the importance
-algorithm. There are 4 categories: Red, Grey, Bold, and Idle. Each has their own definition based off
-relative (perceived) importance to the user:
-
-- **Red**: The room has unread mentions waiting for the user.
-- **Grey**: The room has unread notifications waiting for the user. Notifications are simply unread
- messages which cause a push notification or badge count. Typically, this is the default as rooms get
- set to 'All Messages'.
-- **Bold**: The room has unread messages waiting for the user. Essentially this is a grey room without
- a badge/notification count (or 'Mentions Only'/'Muted').
-- **Idle**: No useful (see definition of useful above) activity has occurred in the room since the user
- last read it.
-
-Conveniently, each tag gets ordered by those categories as presented: red rooms appear above grey, grey
-above bold, etc.
-
-Once the algorithm has determined which rooms belong in which categories, the tag sorting algorithm
-gets applied to each category in a sub-list fashion. This should result in the red rooms (for example)
-being sorted alphabetically amongst each other as well as the grey rooms sorted amongst each other, but
-collectively the tag will be sorted into categories with red being at the top.
-
-## Sticky rooms
-
-When the user visits a room, that room becomes 'sticky' in the list, regardless of ordering algorithm.
-From a code perspective, the underlying algorithm is not aware of a sticky room and instead the base class
-manages which room is sticky. This is to ensure that all algorithms handle it the same.
-
-The sticky flag is simply to say it will not move higher or lower down the list while it is active. For
-example, if using the importance algorithm, the room would naturally become idle once viewed and thus
-would normally fly down the list out of sight. The sticky room concept instead holds it in place, never
-letting it fly down until the user moves to another room.
-
-Only one room can be sticky at a time. Room updates around the sticky room will still hold the sticky
-room in place. The best example of this is the importance algorithm: if the user has 3 red rooms and
-selects the middle room, they will see exactly one room above their selection at all times. If they
-receive another notification which causes the room to move into the topmost position, the room that was
-above the sticky room will move underneath to allow for the new room to take the top slot, maintaining
-the sticky room's position.
-
-Though only applicable to the importance algorithm, the sticky room is not aware of category boundaries
-and thus the user can see a shift in what kinds of rooms move around their selection. An example would
-be the user having 4 red rooms, the user selecting the third room (leaving 2 above it), and then having
-the rooms above it read on another device. This would result in 1 red room and 1 other kind of room
-above the sticky room as it will try to maintain 2 rooms above the sticky room.
-
-An exception for the sticky room placement is when there's suddenly not enough rooms to maintain the placement
-exactly. This typically happens if the user selects a room and leaves enough rooms where it cannot maintain
-the N required rooms above the sticky room. In this case, the sticky room will simply decrease N as needed.
-The N value will never increase while selection remains unchanged: adding a bunch of rooms after having
-put the sticky room in a position where it's had to decrease N will not increase N.
-
-## Responsibilities of the store
-
-The store is responsible for the ordering, upkeep, and tracking of all rooms. The room list component simply gets
-an object containing the tags it needs to worry about and the rooms within. The room list component will
-decide which tags need rendering (as it commonly filters out empty tags in most cases), and will deal with
-all kinds of filtering.
-
-## Filtering
-
-Filters are provided to the store as condition classes and have two major kinds: Prefilters and Runtime.
-
-Prefilters flush out rooms which shouldn't appear to the algorithm implementations. Typically this is
-due to some higher order room list filtering (such as spaces or tags) deliberately exposing a subset of
-rooms to the user. The algorithm implementations will not see a room being prefiltered out.
-
-Runtime filters are used for more dynamic filtering, such as the user filtering by room name. These
-filters are passed along to the algorithm implementations where those implementations decide how and
-when to apply the filter. In practice, the base `Algorithm` class ends up doing the heavy lifting for
-optimization reasons.
-
-The results of runtime filters get cached to avoid needlessly iterating over potentially thousands of
-rooms, as the old room list store does. When a filter condition changes, it emits an update which (in this
-case) the `Algorithm` class will pick up and act accordingly. Typically, this also means filtering a
-minor subset where possible to avoid over-iterating rooms.
-
-All filter conditions are considered "stable" by the consumers, meaning that the consumer does not
-expect a change in the condition unless the condition says it has changed. This is intentional to
-maintain the caching behaviour described above.
-
-One might ask why we don't just use prefilter conditions for everything, and the answer is one of slight
-subtlety: in the cases of prefilters we are knowingly exposing the user to a workspace-style UX where
-room notifications are self-contained within that workspace. Runtime filters tend to not want to affect
-visible notification counts (as it doesn't want the room header to suddenly be confusing to the user as
-they type), and occasionally UX like "found 2/12 rooms" is desirable. If prefiltering were used instead,
-the notification counts would vary while the user was typing and "found 2/12" UX would not be possible.
-
-## Class breakdowns
-
-The `RoomListStore` is the major coordinator of various algorithm implementations, which take care
-of the various `ListAlgorithm` and `SortingAlgorithm` options. The `Algorithm` class is responsible
-for figuring out which tags get which rooms, as Matrix specifies them as a reverse map: tags get
-defined on rooms and are not defined as a collection of rooms (unlike how they are presented to the
-user). Various list-specific utilities are also included, though they are expected to move somewhere
-more general when needed. For example, the `membership` utilities could easily be moved elsewhere
-as needed.
-
-The various bits throughout the room list store should also have jsdoc of some kind to help describe
-what they do and how they work.
diff --git a/docs/room-list.md b/docs/room-list.md
new file mode 100644
index 00000000000..57e47755548
--- /dev/null
+++ b/docs/room-list.md
@@ -0,0 +1,198 @@
+# Room list
+
+The room list is the sorted, filtered list of rooms in the left-hand navigation panel, always
+scoped to the active space. It is a full [MVVM](MVVM.md) feature split across three layers:
+
+- **Model** — `RoomListStoreV3`, the store that keeps every room sorted and answers queries for
+ the rooms in the active space. It lives in
+ [`apps/web/src/stores/room-list-v3/`](https://github.com/element-hq/element-web/tree/develop/apps/web/src/stores/room-list-v3/).
+- **View models** — adapt the store (and other stores) into snapshots for the UI. They live in
+ [`apps/web/src/viewmodels/room-list/`](https://github.com/element-hq/element-web/tree/develop/apps/web/src/viewmodels/room-list/).
+- **Views** — the presentational React components, owned by `@element-hq/web-shared-components`.
+
+## Architecture
+
+```mermaid
+flowchart TD
+ events["Matrix events"] --> store
+ otherStores["Notification / preview /
call stores, room events"] --> children
+
+ subgraph model["Model"]
+ store["RoomListStoreV3
(owns the skip list)"]
+ end
+
+ subgraph vms["View models"]
+ rlvm["RoomListViewModel"]
+ children["RoomListItemViewModel
RoomListSectionHeaderViewModel"]
+ siblings["RoomListHeaderViewModel
RoomListSearchViewModel"]
+ end
+
+ subgraph views["Views"]
+ rlview["RoomListView
VirtualizedRoomListView"]
+ rowviews["RoomListItemView
RoomListSectionHeaderView"]
+ sibviews["RoomListHeaderView
RoomListSearchView"]
+ end
+
+ store -->|"events"| rlvm
+ rlvm -->|"creates & owns"| children
+ rlvm -->|"snapshot"| rlview
+ rlview -->|"renders"| rowviews
+ children -->|"snapshot"| rowviews
+ siblings -->|"snapshot"| sibviews
+```
+
+The key thing the diagram captures: view models subscribe to the **store**, never to the skip
+list — the skip list is an internal detail of `RoomListStoreV3`. And the per-room and
+per-section view models get their data from fine-grained domain stores directly, not from the
+room-list store's update events.
+
+## Model — the store
+
+`RoomListStoreV3` (the "V3" is the third implementation) is a singleton exposed as
+`RoomListStoreV3.instance`. Unlike the previous implementations, which re-computed lists on
+demand, V3 keeps every room permanently sorted so retrieval is cheap.
+
+**The skip list.** The core data structure is a
+[skip list](https://en.wikipedia.org/wiki/Skip_list): stacked linked lists that make insertion
+`O(log n)` on average. Rooms are wrapped in `RoomNode`s that cache whether the room is in the
+active space and which filters it matches, so reading the list is just an ordered walk that
+skips nodes outside the space or not matching the requested filters. When a room changes it is
+re-inserted (re-sorted into place) and the rest of the list is untouched.
+
+**Sorters.** A sorter defines the order of the list. There are three: **Alphabetic** (by room
+name), **Recency** (most recent meaningful activity, with low-priority and muted rooms forced to
+the bottom), and **Unread** (by unread importance). The user's choice is persisted and can be
+changed at runtime, which rebuilds the list.
+
+**Filters.** A filter answers a single yes/no question about a room, and each room's matching
+filters are precomputed on its node. Filters serve two roles: **capability filters** back the
+primary filter chips in the UI (Unread, People, Rooms, Favourites, Mentions, Invites, Low
+Priority), and **section-tag filters** decide which section a room belongs to.
+
+**Sections.** The store groups rooms into named sections (Favourites, Low Priority, Chats, and
+user-created custom sections) using the section-tag filters described above. Sections are a topic
+of their own — see [Sections](#sections) below.
+
+**Spaces and updates.** The list is always scoped to the active space; rather than re-filter on
+every read, each node caches its active-space membership and the store recomputes that flag when
+the space changes. The store listens for the Matrix events that affect ordering or membership
+(receipts, tag changes, account data, decryption, timeline events, membership) and responds by
+re-inserting or removing the single affected room. Emissions to consumers are coalesced with
+`requestAnimationFrame`, so a burst of updates within a frame collapses into one notification.
+
+**Public surface.** This is the seam the view models bind to:
+
+- Events: `ListsUpdate` (lists changed), `ListsLoaded` (initial load done), `SectionCreated`,
+ and `RoomTagged` (a locally-initiated tag change, which drives the "chat moved" toast).
+- `getSortedRoomsInActiveSpace(filterKeys?)` returns the sorted, filtered rooms grouped into
+ sections. Narrower helpers exist too (DM rooms, server-notice rooms, the full sorted list).
+- Section mutators (`createSection` / `editSection` / `removeSection` / `reorderSection`) and
+ `resort`.
+
+Note that **sticky-room behaviour lives in the view model, not the store** — the store itself is
+unaware of a selected room.
+
+## Sections
+
+Sections are the named groups the room list displays. Membership is driven entirely by the Matrix
+`m.tag` account data on each room, so moving a room between sections is just a matter of changing
+its tags:
+
+- **Favourites** and **Low Priority** use the default `m.favourite` and `m.lowpriority` tags and
+ are pinned to the top and bottom of the list respectively.
+- **Chats** is a synthetic catch-all section (tag `chats`) for every room that isn't in any other
+ explicit section.
+- **Custom sections** are user-created. Each is identified by a generated tag of the form
+ `element.io.section.`, and a room is placed in it by tagging the room with that tag.
+
+The store keeps an ordered list of section tags — Favourites first, Low Priority last, and the
+custom sections plus Chats reorderable in between (new custom sections are inserted just above
+Chats by default). Custom sections can be created, renamed, removed, and reordered through
+dialogs; that logic lives in [`section.ts`](https://github.com/element-hq/element-web/blob/develop/apps/web/src/stores/room-list-v3/section.ts).
+
+Each custom section also records the space it was created in, which controls the visibility of
+empty sections: an empty custom section is only shown in the space it belongs to, so sections
+created in one space don't clutter unrelated spaces. Legacy sections without a stored space (or
+whose space no longer exists) fall back to the Home meta-space.
+
+### Flat list mode
+
+When sectioning is turned off, the room list becomes a single flat list instead of a grouped one.
+This is triggered in two places:
+
+- If the `RoomList.showSections` setting is disabled, the store skips sectioning altogether and
+ `getSortedRoomsInActiveSpace()` returns a single `chats` section containing every room in the
+ active space.
+- Even with sectioning enabled, `RoomListViewModel` treats the result as flat whenever the only
+ section left is the Chats catch-all (`isFlatList`) — for example when the user has no favourite,
+ low-priority, or custom-tagged rooms.
+
+In flat mode the view renders a `FlatVirtualizedList` (no section headers); otherwise it renders a
+`GroupedVirtualizedList` with headers and section drag-and-drop.
+
+### Settings
+
+Sections are configured entirely through settings:
+
+- **`RoomList.showSections`** — whether sectioning is enabled at all. When off, the list is flat
+ (see above).
+- **`RoomList.CustomSectionData`** (account level) — the custom-section definitions, keyed by tag.
+ Each entry stores the tag, the user-chosen name, and the space the section was created in.
+ Malformed entries are dropped when the data is read.
+- **`RoomList.OrderedCustomSections`** (account level) — the display order of the reorderable
+ sections (the custom sections and Chats). Favourites and Low Priority are not stored here since
+ they are always pinned to the top and bottom.
+- **`RoomList.SectionExpansionState`** (device level) — the expanded/collapsed state of each
+ section, stored per space and then per section tag. Sections default to expanded when no state
+ has been persisted.
+
+The account-level settings sync across a user's devices, while the expansion state is
+device-local so that collapsing a section on one device doesn't affect the others.
+
+## View models
+
+The view models in [`apps/web/src/viewmodels/room-list/`](https://github.com/element-hq/element-web/tree/develop/apps/web/src/viewmodels/room-list/)
+extend `BaseViewModel` and produce immutable snapshots consumed by the views (see [MVVM](MVVM.md)
+for the base pattern).
+
+**`RoomListViewModel`** is the root orchestrator and the **only** view model that subscribes to
+the store's list events and calls `getSortedRoomsInActiveSpace()`. It builds the top-level
+snapshot — whose sections carry only **room IDs**, not room objects — and owns the sticky-room
+logic, the active filter, the "unread activity below the fold" toast, and section drag-and-drop.
+It also lazily creates and owns the child view models.
+
+**`RoomListItemViewModel`** (one per room) and **`RoomListSectionHeaderViewModel`** (one per
+section) are created on demand by the root view model. Crucially, they do **not** listen to the
+store's list events; instead they subscribe to fine-grained domain stores directly —
+`RoomNotificationStateStore`, `MessagePreviewStore`, `CallStore`, room events, and settings — so
+a single row can update independently of the rest of the list. The section header view model is
+_fed_ its set of rooms imperatively by the parent and aggregates their notification state.
+
+**`RoomListHeaderViewModel`** (the header bar: space title, sort menu, create actions) and
+**`RoomListSearchViewModel`** (the search/dial/explore row) are decoupled siblings of the root
+view model. They do not share state directly; the header view model receives things like the
+collapse-all-sections state from the root view model through the global dispatcher.
+
+Shared permission and create-room helpers used across these live in
+[`utils.ts`](https://github.com/element-hq/element-web/blob/develop/apps/web/src/viewmodels/room-list/utils.ts).
+
+## Views
+
+The presentational components are owned by `@element-hq/web-shared-components` (developed in
+Storybook); `apps/web` supplies the view models and a `renderAvatar` callback, and the shared
+package owns all rendering. The app-side entry point is
+[`RoomListPanel`](https://github.com/element-hq/element-web/blob/develop/apps/web/src/components/views/rooms/RoomListPanel/RoomListPanel.tsx), which
+composes the search row, the header view, and the room list itself.
+
+The room list proper is `RoomListView`, which delegates to `VirtualizedRoomListView` (built on
+[`react-virtuoso`](https://virtuoso.dev/)). Virtualization is what drives the lazy lifecycle of
+the child view models: as rows scroll into view the list calls back through
+`getRoomItemViewModel(id)` and `getSectionHeaderViewModel(tag)` to obtain the view model for each
+rendered room or header, and reports its visible range so off-screen item view models can be
+disposed. Because each row binds to its own view model, it re-renders on its own data without
+re-rendering the whole list.
+
+The list renders either as a flat list or a grouped list with section headers, depending on the
+`isFlatList` flag in the snapshot (see [Flat list mode](#flat-list-mode)); in the grouped case,
+drag-and-drop is wired back to the root view model's `changeSectionOrder` and `changeRoomSection`
+actions.