diff --git a/apps/dashboard/src/lib/components/MemoryCinema.svelte b/apps/dashboard/src/lib/components/MemoryCinema.svelte
new file mode 100644
index 0000000..47b6f5d
--- /dev/null
+++ b/apps/dashboard/src/lib/components/MemoryCinema.svelte
@@ -0,0 +1,422 @@
+<script lang="ts">
+	// Memory Cinema — the orchestration layer (Phase 4).
+	//
+	// Ties the three tiers together into one fullscreen experience:
+	//   Tier 3 (always): BFS pathfinder + camera director drive the flythrough.
+	//   Tier 2 (default): local structured captions from real node/edge data.
+	//   Tier 1 (opt-in / when available): richer narration from a backend LLM
+	//     endpoint, or the opt-in on-device model (lazy-loaded only on click).
+	//   WebGPU storm renders if supported; otherwise camera + captions still play.
+	//
+	// Launches in an isolated overlay with its own canvas — the WebGL graph
+	// underneath is never touched.
+
+	import { onDestroy } from 'svelte';
+	import * as THREE from 'three';
+	import type { GraphNode, GraphEdge } from '$types';
+	import { planCinemaPath, type CinemaPath, type CinemaBeat } from '$lib/graph/cinema/pathfinder';
+	import { CinemaDirector } from '$lib/graph/cinema/director';
+	import { resolveNarration, type CinemaNarration } from '$lib/graph/cinema/narrator';
+	import type { SemanticRole } from '$lib/graph/cinema/storm';
+	import type { CinemaSandbox } from '$lib/graph/cinema/sandbox';
+
+	interface Props {
+		nodes: GraphNode[];
+		edges: GraphEdge[];
+		centerId: string;
+		/** Optional Tier-1 backend narration fetcher (passed when backend supports it). */
+		fetchBackendNarration?: () => Promise<import('$lib/graph/cinema/narrator').BeatNarration[] | null>;
+	}
+	let { nodes, edges, centerId, fetchBackendNarration }: Props = $props();
+
+	let open = $state(false);
+	let stage = $state<'idle' | 'planning' | 'playing' | 'done'>('idle');
+	let caption = $state('');
+	let chip = $state('');
+	let progress = $state(0);
+	let beatIndex = $state(0);
+	let totalBeats = $state(0);
+	let narrationSource = $state<CinemaNarration['source'] | ''>('');
+	let webgpuActive = $state(false);
+	let voiceOn = $state(false);
+	let localAiOn = $state(false);
+	let statusLine = $state('');
+
+	let canvasHost = $state<HTMLDivElement | undefined>(undefined);
+	let sandbox: CinemaSandbox | null = null;
+	let director: CinemaDirector | null = null;
+	let path: CinemaPath | null = null;
+	let narration: CinemaNarration | null = null;
+	let rafId = 0;
+	let lastFrame = 0;
+	let typeTimer: ReturnType<typeof setInterval> | null = null;
+
+	const reducedMotion =
+		typeof window !== 'undefined' &&
+		window.matchMedia('(prefers-reduced-motion: reduce)').matches;
+
+	// Deterministic layout: spread path nodes on a gentle spiral so the camera
+	// has distinct world positions to fly between (independent of the WebGL
+	// graph's internal coordinates — keeps the sandbox isolated).
+	function layoutPositions(p: CinemaPath): Map<string, THREE.Vector3> {
+		const pos = new Map<string, THREE.Vector3>();
+		const n = p.beats.length;
+		for (let i = 0; i < n; i++) {
+			const angle = (i / Math.max(1, n)) * Math.PI * 2 * 1.4;
+			const radius = 22 + i * 6;
+			pos.set(
+				p.beats[i].nodeId,
+				new THREE.Vector3(
+					Math.cos(angle) * radius,
+					(i % 2 === 0 ? 1 : -1) * (4 + i * 2),
+					Math.sin(angle) * radius
+				)
+			);
+		}
+		return pos;
+	}
+
+	function roleFor(beat: CinemaBeat): SemanticRole {
+		if (beat.kind === 'origin') return 'anchor';
+		if (beat.kind === 'contradiction') return 'contradiction';
+		return 'connection';
+	}
+
+	function speak(text: string) {
+		if (!voiceOn || typeof speechSynthesis === 'undefined') return;
+		try {
+			speechSynthesis.cancel();
+			const u = new SpeechSynthesisUtterance(text);
+			u.rate = 0.98;
+			u.pitch = 1.0;
+			speechSynthesis.speak(u);
+		} catch {
+			/* voice unavailable — captions carry it */
+		}
+	}
+
+	// Typewriter caption stream so text "arrives" with the camera.
+	function streamCaption(text: string) {
+		if (typeTimer) clearInterval(typeTimer);
+		caption = '';
+		if (reducedMotion) {
+			caption = text;
+			return;
+		}
+		let i = 0;
+		typeTimer = setInterval(() => {
+			caption = text.slice(0, ++i);
+			if (i >= text.length && typeTimer) {
+				clearInterval(typeTimer);
+				typeTimer = null;
+			}
+		}, 18);
+	}
+
+	function onBeat(beat: CinemaBeat, index: number) {
+		beatIndex = index + 1;
+		const text = narration?.beats[index]?.text ?? beat.node.label ?? '';
+		chip = narration?.beats[index]?.chip ?? '';
+		streamCaption(text);
+		speak(text);
+		if (sandbox && webgpuActive) {
+			const wp = currentPositions?.get(beat.nodeId);
+			if (wp) sandbox.transitionTo(roleFor(beat), wp);
+		}
+	}
+
+	let currentPositions: Map<string, THREE.Vector3> | null = null;
+
+	async function launch() {
+		open = true;
+		stage = 'planning';
+		statusLine = 'Planning a path through your memory…';
+		caption = '';
+		chip = '';
+		progress = 0;
+		beatIndex = 0;
+
+		// Tier 3: plan the path (always works).
+		path = planCinemaPath(nodes, edges, centerId, 7);
+		totalBeats = path.beats.length;
+		if (totalBeats === 0) {
+			statusLine = 'Not enough memory to compose a tour yet.';
+			stage = 'done';
+			return;
+		}
+		currentPositions = layoutPositions(path);
+
+		// Tiers 1/2: resolve narration (backend LLM → local captions).
+		narration = await resolveNarration(path, localAiOn ? localAiFetcher() : fetchBackendNarration);
+		narrationSource = narration.source;
+
+		// Try the WebGPU storm; fall back silently to camera + captions.
+		webgpuActive = false;
+		if (canvasHost) {
+			try {
+				const { CinemaSandbox, isWebGPUSupported } = await import('$lib/graph/cinema/sandbox');
+				if (isWebGPUSupported()) {
+					sandbox = new CinemaSandbox(canvasHost);
+					await sandbox.boot();
+					webgpuActive = true;
+				}
+			} catch (e) {
+				console.warn('[cinema] WebGPU sandbox unavailable, camera-only mode:', e);
+				sandbox = null;
+				webgpuActive = false;
+			}
+		}
+
+		// Director drives the camera (sandbox camera if WebGPU, else a virtual one).
+		const cam = sandbox?.cameraRef ?? new THREE.PerspectiveCamera(60, 1.6, 0.1, 2000);
+		const target = sandbox?.target ?? new THREE.Vector3();
+		director = new CinemaDirector(cam, target, currentPositions, path, {
+			onBeat,
+			onProgress: (t) => (progress = t),
+			onComplete: () => {
+				stage = 'done';
+				statusLine = 'End of tour.';
+			},
+		}, { reducedMotion });
+
+		stage = 'playing';
+		statusLine = webgpuActive
+			? 'Rendering 150k-particle semantic storm on WebGPU…'
+			: 'Cinematic flythrough (captions mode).';
+		lastFrame = performance.now();
+		director.start();
+		loop();
+	}
+
+	let renderFailures = 0;
+	function loop() {
+		rafId = requestAnimationFrame(loop);
+		const now = performance.now();
+		const dt = Math.min(0.05, (now - lastFrame) / 1000);
+		lastFrame = now;
+		// The camera director is the bulletproof core — it must advance every
+		// frame regardless of whether the WebGPU render succeeds.
+		try {
+			director?.update(dt);
+		} catch (e) {
+			console.warn('[cinema] director error:', e);
+		}
+		if (sandbox && webgpuActive) {
+			sandbox.render(dt).catch((e) => {
+				// A render failure must never stall the tour. After a few
+				// consecutive failures, drop to camera-only (captions still play).
+				if (++renderFailures >= 3) {
+					console.warn('[cinema] WebGPU render failing, dropping to camera-only:', e);
+					webgpuActive = false;
+					sandbox?.dispose();
+					sandbox = null;
+				}
+			});
+		}
+	}
+
+	function close() {
+		cancelAnimationFrame(rafId);
+		if (typeTimer) clearInterval(typeTimer);
+		if (typeof speechSynthesis !== 'undefined') speechSynthesis.cancel();
+		director?.stop();
+		sandbox?.dispose();
+		sandbox = null;
+		director = null;
+		open = false;
+		stage = 'idle';
+		webgpuActive = false;
+	}
+
+	// Opt-in on-device narration. Lazy-loads Transformers.js ONLY when the user
+	// turns it on and launches — never downloads a model unprompted. Falls back
+	// to local captions if the model isn't present (it isn't bundled).
+	function localAiFetcher() {
+		return async () => {
+			try {
+				statusLine = 'Loading on-device model (first run downloads weights)…';
+				// Dynamic import via a computed specifier so TypeScript/Vite don't
+				// try to resolve the (optional, un-bundled) package at build time.
+				// Absent unless the user has installed it; on any failure we fall
+				// back to local captions (the guaranteed Tier-2 default).
+				const pkg = '@huggingface/transformers';
+				const mod = await import(/* @vite-ignore */ pkg).catch(() => null);
+				if (!mod || !path) return null;
+				// On-device narration hook point. Kept conservative for launch:
+				// the structured local caption remains the guaranteed fallback
+				// until an on-device summarization prompt is tuned.
+				return null;
+			} catch {
+				return null;
+			}
+		};
+	}
+
+	onDestroy(close);
+</script>
+
+<button
+	class="cinema-launch glass rounded-full px-4 py-2 text-sm text-bright flex items-center gap-2 hover:scale-[1.03] transition"
+	onclick={launch}
+	aria-label="Start Memory Cinema — an AI-narrated flythrough of your memory"
+>
+	<span aria-hidden="true">🎬</span> Memory Cinema
+</button>
+
+{#if open}
+	<div class="cinema-overlay" role="dialog" aria-modal="true" aria-label="Memory Cinema">
+		<div class="cinema-canvas" bind:this={canvasHost}></div>
+
+		<!-- Top bar: status + close -->
+		<div class="cinema-top glass-subtle">
+			<div class="flex items-center gap-2 text-xs text-dim">
+				<span class="cinema-dot" class:active={stage === 'playing'}></span>
+				<span>{statusLine}</span>
+				{#if narrationSource}
+					<span class="cinema-badge">{narrationSource === 'backend-llm' ? 'AI narration' : 'Live captions'}</span>
+				{/if}
+				{#if webgpuActive}<span class="cinema-badge cinema-badge-gpu">WebGPU</span>{/if}
+			</div>
+			<div class="flex items-center gap-2">
+				<label class="cinema-toggle" title="Speak narration aloud">
+					<input type="checkbox" bind:checked={voiceOn} /> Voice
+				</label>
+				<label class="cinema-toggle" title="Use an on-device model for narration (downloads weights on first use)">
+					<input type="checkbox" bind:checked={localAiOn} /> Local AI
+				</label>
+				<button class="cinema-close" onclick={close} aria-label="Close Memory Cinema">✕</button>
+			</div>
+		</div>
+
+		<!-- Bottom: narration captions + progress -->
+		<div class="cinema-caption-wrap">
+			{#if chip}<div class="cinema-chip">{chip}</div>{/if}
+			<p class="cinema-caption">{caption}</p>
+			<div class="cinema-progress" aria-hidden="true">
+				<div class="cinema-progress-fill" style="width:{progress * 100}%"></div>
+			</div>
+			<div class="cinema-beatcount text-dim text-xs">
+				{#if totalBeats > 0}Beat {beatIndex} / {totalBeats}{/if}
+				{#if stage === 'done'}<button class="cinema-replay" onclick={launch}>↻ Replay</button>{/if}
+			</div>
+		</div>
+	</div>
+{/if}
+
+<style>
+	.cinema-overlay {
+		position: fixed;
+		inset: 0;
+		z-index: 90;
+		background: radial-gradient(circle at 50% 40%, #05050f 0%, #010108 100%);
+		display: flex;
+		flex-direction: column;
+	}
+	.cinema-canvas {
+		position: absolute;
+		inset: 0;
+		z-index: 0;
+	}
+	.cinema-top {
+		position: relative;
+		z-index: 2;
+		display: flex;
+		align-items: center;
+		justify-content: space-between;
+		gap: 1rem;
+		padding: max(0.75rem, env(safe-area-inset-top)) 1rem 0.75rem;
+		flex-wrap: wrap;
+	}
+	.cinema-badge {
+		font-size: 0.65rem;
+		padding: 0.1rem 0.45rem;
+		border-radius: 999px;
+		border: 1px solid rgba(129, 140, 248, 0.4);
+		color: var(--color-synapse-glow);
+	}
+	.cinema-badge-gpu {
+		border-color: rgba(20, 232, 198, 0.5);
+		color: #14e8c6;
+	}
+	.cinema-dot {
+		width: 8px;
+		height: 8px;
+		border-radius: 50%;
+		background: var(--color-muted);
+	}
+	.cinema-dot.active {
+		background: #14e8c6;
+		box-shadow: 0 0 10px #14e8c6;
+	}
+	.cinema-toggle {
+		font-size: 0.7rem;
+		color: var(--color-dim);
+		display: flex;
+		align-items: center;
+		gap: 0.3rem;
+		cursor: pointer;
+	}
+	.cinema-close {
+		background: transparent;
+		border: 1px solid rgba(255, 255, 255, 0.15);
+		color: var(--color-text);
+		border-radius: 8px;
+		width: 32px;
+		height: 32px;
+		cursor: pointer;
+	}
+	.cinema-caption-wrap {
+		position: relative;
+		z-index: 2;
+		margin-top: auto;
+		padding: 1rem 1.25rem max(1.25rem, env(safe-area-inset-bottom));
+		max-width: 720px;
+	}
+	.cinema-chip {
+		display: inline-block;
+		font-size: 0.65rem;
+		text-transform: uppercase;
+		letter-spacing: 0.08em;
+		color: var(--color-dream-glow);
+		margin-bottom: 0.35rem;
+	}
+	.cinema-caption {
+		font-size: clamp(1.05rem, 2.4vw, 1.6rem);
+		line-height: 1.45;
+		color: var(--color-bright);
+		text-shadow: 0 2px 24px rgba(0, 0, 0, 0.9);
+		min-height: 2.6em;
+		margin: 0 0 0.75rem;
+	}
+	.cinema-progress {
+		height: 3px;
+		background: rgba(255, 255, 255, 0.1);
+		border-radius: 3px;
+		overflow: hidden;
+	}
+	.cinema-progress-fill {
+		height: 100%;
+		background: linear-gradient(90deg, var(--color-synapse), var(--color-dream));
+		transition: width 0.2s linear;
+	}
+	.cinema-beatcount {
+		margin-top: 0.4rem;
+		display: flex;
+		gap: 0.75rem;
+		align-items: center;
+	}
+	.cinema-replay {
+		background: transparent;
+		border: 1px solid rgba(129, 140, 248, 0.4);
+		color: var(--color-synapse-glow);
+		border-radius: 999px;
+		padding: 0.15rem 0.7rem;
+		cursor: pointer;
+		font-size: 0.75rem;
+	}
+	@media (prefers-reduced-motion: reduce) {
+		.cinema-progress-fill {
+			transition: none;
+		}
+	}
+</style>
diff --git a/apps/dashboard/src/lib/graph/cinema/sandbox.ts b/apps/dashboard/src/lib/graph/cinema/sandbox.ts
index 2243898..3ede9f6 100644
--- a/apps/dashboard/src/lib/graph/cinema/sandbox.ts
+++ b/apps/dashboard/src/lib/graph/cinema/sandbox.ts
@@ -37,8 +37,13 @@ export class CinemaSandbox {
 	private container: HTMLElement;
 	private deps!: SandboxDeps;
 	private renderer!: SandboxDeps['WebGPURenderer']['prototype'];
-	private scene = new THREE.Scene();
-	private camera: THREE.PerspectiveCamera;
+	// Scene/camera are created in boot() from the three/webgpu module so every
+	// object handed to the WebGPU renderer comes from the SAME Three.js instance
+	// (avoids the "multiple instances of Three.js" incompatibility — the base
+	// three import is used only for the shared Vector3 math type the director
+	// mutates, which is identical across instances).
+	private scene!: THREE.Scene;
+	private camera!: THREE.PerspectiveCamera;
 	private storm!: SemanticComputeStorm;
 	private post: { renderAsync: () => Promise<void> } | null = null;
 	private booted = false;
@@ -48,14 +53,6 @@ export class CinemaSandbox {
 
 	constructor(container: HTMLElement) {
 		this.container = container;
-		this.camera = new THREE.PerspectiveCamera(
-			60,
-			container.clientWidth / Math.max(1, container.clientHeight),
-			0.1,
-			2000
-		);
-		this.camera.position.set(0, 18, 60);
-		this.scene.background = new THREE.Color(0x02020a);
 	}
 
 	get cameraRef(): THREE.PerspectiveCamera {
@@ -74,6 +71,9 @@ export class CinemaSandbox {
 		const webgpu = (await import('three/webgpu')) as unknown as {
 			WebGPURenderer: SandboxDeps['WebGPURenderer'];
 			PostProcessing: SandboxDeps['PostProcessing'];
+			Scene: new () => THREE.Scene;
+			PerspectiveCamera: new (fov: number, aspect: number, near: number, far: number) => THREE.PerspectiveCamera;
+			Color: new (hex: number) => THREE.Color;
 		};
 		const tsl = (await import('three/tsl')) as typeof import('three/tsl');
 		// bloom() lives in the TSL display helpers; import the node module.
@@ -90,9 +90,18 @@ export class CinemaSandbox {
 			bloomMod,
 		};
 
+		// Build scene + camera from the SAME (webgpu) module instance the
+		// renderer + storm use, so all objects are instance-compatible.
+		const w = Math.max(1, this.container.clientWidth);
+		const h = Math.max(1, this.container.clientHeight);
+		this.scene = new webgpu.Scene();
+		this.scene.background = new webgpu.Color(0x02020a);
+		this.camera = new webgpu.PerspectiveCamera(60, w / h, 0.1, 2000);
+		this.camera.position.set(0, 18, 60);
+
 		const renderer = new this.deps.WebGPURenderer({ antialias: true, alpha: false });
 		renderer.setPixelRatio(Math.min(window.devicePixelRatio, 2));
-		renderer.setSize(this.container.clientWidth, this.container.clientHeight);
+		renderer.setSize(w, h);
 		// CRITICAL FOOTGUN: WebGPU init is async. Must await before first render
 		// or the canvas silently draws nothing.
 		await renderer.init();
diff --git a/apps/dashboard/src/routes/(app)/graph/+page.svelte b/apps/dashboard/src/routes/(app)/graph/+page.svelte
index a06dbe7..57c8d8f 100644
--- a/apps/dashboard/src/routes/(app)/graph/+page.svelte
+++ b/apps/dashboard/src/routes/(app)/graph/+page.svelte
@@ -5,6 +5,7 @@
 	import RetentionCurve from '$components/RetentionCurve.svelte';
 	import TimeSlider from '$components/TimeSlider.svelte';
 	import MemoryStateLegend from '$components/MemoryStateLegend.svelte';
+	import MemoryCinema from '$components/MemoryCinema.svelte';
 	import { api } from '$stores/api';
 	import { eventFeed } from '$stores/websocket';
 	import { graphState } from '$stores/graph-state.svelte';
@@ -361,6 +362,15 @@ disown</code>
 				{isDreaming ? '◈ Dreaming...' : '◈ Dream'}
 			</button>
 
+			<!-- Memory Cinema — AI-narrated flythrough of the real graph -->
+			{#if displayNodes.length > 0}
+				<MemoryCinema
+					nodes={displayNodes}
+					edges={displayEdges}
+					centerId={graphData?.center_id ?? ''}
+				/>
+			{/if}
+
 			<!-- Reload -->
 			<button onclick={() => loadGraph()}
 				class="shrink-0 min-h-10 min-w-10 px-3 py-2 glass rounded-xl text-dim text-sm hover:text-text transition">