VastMemory Sphere — the 3D memory view, chat-choreographed

In one sentence: Retire the radial and tree views and replace them with one excellent 3D spherical memory space — easy to maneuver, beautiful at scale — that the AI plays like an instrument: as you search and talk, the sphere flies, re-clusters, filters, and opens previews on its own, so the conversation and the view are one experience.

Why (Daniel's verdict, 2026-06-11): the current radial and tree views are weak and neither is impressive — three mediocre views are worse than one great one. Memory is The Cloud's most differentiating organ (the Supermemory review: memory with a home), and its flagship view should feel like that claim. The deeper unlock isn't 3D for its own sake — it's chat-coupling: today the graph is something you poke at; the target is a view that responds to the conversation, which no memory product ships and which makes every demo of The Cloud unforgettable.

Reconcile-don't-duplicate: the Memory app's existing graph data, embeddings, and zoom/navigation work (memory graph search shipped #1209) are the substrate — this is a view replacement, not a data-model change. Context Design (nebula stacking) supplies the cluster semantics. The chat-coupling pattern generalizes the Session Mesh/congruency principle: surfaces tell the truth about what the AI is doing, live. Motion obeys the Hogwarts rule and the House Motion Language (Studio Motion Plugins doc) — choreography serves meaning, reduced-motion honored.

The Sphere

• Geometry: memories as instanced nodes on/within a navigable sphere — constellations, not a hairball. Semantic position (embedding-driven) so neighborhoods mean something; cluster halos with plain-language labels; edges shown sparingly on focus, never as permanent spaghetti.

• Maneuvering (the "easy to use" bar): orbit-drag, scroll/pinch zoom, click-to-focus with a smooth fly-to; inertia tuned calm; <kbd>Esc</kbd> always returns to the home orbit. One minute to feel at home, zero tutorial. 60fps floor via instancing + LOD; if a device can't hold the floor, it gets the dignified 2D fallback, never a slideshow.

• Search-to-flight: typing in search doesn't filter a list off to the side — the sphere answers: matching nodes brighten, the camera glides to the densest match region, non-matches recede (never vanish — context stays visible, dimmed).

Chat choreography (the headline feature)

A small view-ops vocabulary the AI emits alongside its chat replies — the same instrument a human plays with mouse and keys:

• focus(nodes) — fly to and brighten

• cluster(by: tag | time | space | person | project) — live re-arrangement, animated as one legible transition

• filter(query) / clear() — dim the rest, keep the world visible

• preview(node) — open the page/memory preview panel beside the sphere

• trail(nodes[]) — light a path through a sequence (a project's history, a person's thread)

So: you say "show me everything around Kindo this spring" — the sphere re-clusters by project, flies to the Kindo constellation, dims the rest, lights the timeline trail, and opens the tracker preview. The reply you read and the motion you watch are the same answer. Congruency law applies: the view never animates something the AI didn't actually do, and every AI-driven motion is also achievable by hand.

Phases

• Phase 0 — Retire radial + tree; Sphere v1. Remove the two weak views (one good default beats three mediocre options). Sphere with semantic placement, instanced rendering, orbit/zoom/focus, search-to-flight. Acceptance: 5k memories at 60fps on a MacBook; a first-time user finds a known memory in under 30 seconds unaided.

• Phase 1 — Clustering + filters. Cluster-by tag/time/Space/person with animated, legible transitions; filter chips; cluster labels. Acceptance: re-cluster of 5k nodes reads as one comprehensible motion, not an explosion.

• Phase 2 — Chat choreography. The view-ops vocabulary wired to the Memory chat; AI emits ops with replies; previews open in the side panel. Acceptance: the Kindo sentence above produces flight + cluster + trail + preview with zero clicks.

• Phase 3 — Previews + actions. Rich preview panel (page render, linked items, open-in-place); acting from the sphere (open, tag, link, file). Acceptance: a full memory-gardening session happens without leaving the view.

• Phase 4 — Scale + polish. LOD tiers for 50k+ nodes, ambient idle motion within the motion budget, theme-true light/dark, the demo-reel pass.

Rails

House Motion Language tokens for every transition; reduced-motion = crossfades, full function retained · accessibility fallback: the 2D list/search view remains a first-class citizen, not a punishment · performance floor is an acceptance check (60fps or LOD down, never jank) · choreography is honest — the view shows what the AI did, nothing it didn't · private memories render private in shared/screen-share contexts (presentation mode dims sensitive nodes).