Memory Lab

Cinevva's analyzer read this game's design doc and its source, then wrote the brief below. "Make your own" hands it to the game creator as your opening message, so you start from a real spec rather than a one-line idea.

Rebuild a browser-based 3D game called “Memory Lab,” a phone-friendly first-person/over-the-shoulder exploration lab with a dark navy sci-fi environment, glowing cyan grid floor, soft fog, and large open square room bounded by low dark walls. The camera starts behind/above the hidden player near the center, looking toward a floating AI core at the back of the lab. The player avatar is a small yellow capsule character with a dark visor, but it is hidden in the normal camera view; support keyboard WASD and arrow-key movement, plus mobile left touch pad for walking/strafe and right touch pad for looking. Keep the core game world interactive: the player can walk around, inspect clickable memory crystals and created objects, and interact with the AI through the HUD.

Place the AI in the lab as a hovering glowing cyan crystal/icosahedron surrounded by two thin orbiting cyan torus rings. The AI model can change among crystal, sphere/orb, cube, dodecahedron/star-like shape, or cone/needle, and its color/model changes with mood: curious cyan crystal, happy pale green orb, sad blue orb, angry red cube, nervous purple star, or creative gold star. Add an optional photo-derived decoration ring of small glowing color spheres around the AI. Surround the area with futuristic props loaded from Kenney-style space-station/factory assets: computer systems, displays, chairs, machines, pipes, robot arms, and similar lab equipment. Use realistic shadows, a dark blue floor, cyan emissive accents, and a clean low-poly sci-fi aesthetic.

The main objective is to teach and collaborate with the lab AI. Provide a persistent browser localStorage memory system. The player can type “teach <lesson>” or “teach key=value” / “learn key=value”; each lesson becomes a memory, updates the AI’s dictionary of known words, changes mood, and spawns a small purple glowing octahedral memory crystal in the lab. Store facts, goals, entities, learned questions, emotional statistics, feedback, voice settings, appearance settings, camera yaw, and memories persistently. Include built-in starter vocabulary and built-in knowledge about shape recipes, character recipes, Backrooms-style maps, and related helper specs. Let the AI answer questions using known facts, matching word overlap/partial matches to memories, and previously learned similar questions; unknown questions should be saved so future answers can improve. Support text commands: “memory,” “goals,” “stats,” “export,” “mood,” and “forget fact.”

Implement a prominent responsive HUD matching the screenshot: a rounded dark translucent instruction panel at upper left reading approximately “Memory Lab” and “Walk with WASD / arrows. Chat with the lab AI. Upload a photo, describe it, and the AI stores that lesson.”; a rounded upper-right status panel showing “Memories: [count]” and “Objects made: [count]”; and a bottom chat/control area with a scrolling AI message panel, a photo/file chooser, a text field labeled “Describe the photo,” a bright cyan-to-mint “Teach” button, an “AI build” button, “Quiz me” button, “Adventure” button, a chat input labeled “Talk or type to the AI…,” a microphone button, and a bright “Send” button. Use large rounded buttons, dark navy panels with cyan borders, white text, and cyan/mint gradients for touch accessibility. Include a reset control and a visible dictionary/AI status area if space allows.

Chat replies should be spoken automatically with browser SpeechSynthesis when available, with voice always enabled by default. Add a microphone button using browser SpeechRecognition/WebkitSpeechRecognition where supported; recognized speech fills/sends the chat input. Add a voice-settings panel with rate, pitch, and volume sliders; voice remains on. Add a model/character editor panel with sliders for AI size, orbit-ring scale, and glow, plus character styles/presets such as survivor, stalker, smiler, mimic, ringmaster, and robot; building a selected style spawns a smart character. The player can give natural-language build commands such as “make a blue dragon,” “make a house,” “make an NPC,” “make a smart robot character,” “make a torus,” “make a sphere,” “make a crystal,” “recreate the photo as a model,” and “make yourself a character.” Build grouped procedural geometry from boxes, spheres, cylinders, cones, torus rings, capsules, planes, crystals, columns, roofs, limbs, eyes, ears, and glowing detail chips; also spawn premade house, castle, fox, bunny, chicken, king, or pirate models. Created objects should be positioned around the map, increment the object count, and retain metadata describing which prompt/memory produced them. Add wandering smart NPC behavior and allow clicking memory crystals/created objects to inspect their source.

Support photo/file learning. Let the player choose any file. For image files, display a photo preview in the lab, scan the image in-browser for average/dominant colors, show several color swatches, use those colors to recolor/adapt the AI model, and require the player to type a label/description and press Teach to store a photo lesson. Photo memories store the palette and can drive later models. Directly import GLB, GLTF, and OBJ files into the map with Three.js loaders and animate imported models when animations exist. For other model-like formats, create an in-world file artifact plus a best-effort approximation based on the filename. For text/code files, read a preview, learn filename/type/content, detect systems such as player, enemy, NPC, map/level, score/pickup, UI, Three.js/scene/camera, and create prototype objects/adapt the AI model. For other files, make a labeled data artifact based on filename/type/size.

Make “AI build” create several objects from recent memories/photos at once: memory objects, a photo-inspired shape model, a building, and a smart walking NPC. Provide Yes/No feedback buttons: Yes teaches that the player liked the latest build; No stores a dislike and improves the latest build with additional detail and smarter NPC movement. “Quiz me” selects a stored memory, blanks a keyword in its prompt, and asks the player to type the missing word; a correct answer rewards the player by spawning an object, while an incorrect answer reveals the answer and lesson. “Adventure” opens a Memory Rift/Backrooms-style portal map based on learned words: faded yellow liminal walls, damp-looking carpet, flickering lights, lore/research documents, an unlimited flashlight, a return portal, and a Stalker entity. Track the adventure objective as finding a research note, avoiding the Stalker, and returning through the portal; allow continuing the adventure to alter it using the latest memory. Include flashlight chat commands and a visible/functional flashlight in the adventure scene.

Use Three.js ES modules in a single-page browser app, with renderer shadows, hemisphere and directional lighting, fog, a dark #07111f background, and responsive resizing. Expose getGameState/setGameState hooks for live state inspection/patching, but do not show debug overlays. Keep the experience an interactive AI memory lab rather than a conventional combat game: the progression is learning lessons, growing the dictionary/memory counters, answering questions, adapting the AI appearance, and building inspectable 3D creations.