Imagine if you could smell memories. Imagine if you could smell memories. The Anemoia Device is a scent-memory machine that uses generative AI to distill analogue photographs into bespoke fragrances. Each image becomes a singular olfactory composition, a multisensory artifact designed to evoke anemoia –– nostalgia for a moment you never lived. Most AI systems today live on screens. They see, they write, they generate, but they rarely materialize. Rather than producing images, text, or sound alone, this system outputs something far less explored in computing: smell, a sense deeply entangled with memory, emotion, and time. smell This work was developed at the MIT Media Lab’s Tangible Media Group and is described in our NeurIPS 2025 paper: The Anemoia Device: A Tangible AI System for the Co-creation of Synthetic Memories through Scent. Authors: Cyrus Clarke, Nomy Yu, Melo Chen, Yuen Zou, Hiroshi Ishii. The Anemoia Device: A Tangible AI System for the Co-creation of Synthetic Memories through Scent Smelling Memories of Pasts You Never Lived Smelling Memories of Pasts You Never Lived Anemoia is a strange kind of nostalgia. It’s an emerging neologism that describes the feeling of longing for a time you never lived, a memory that doesn’t belong to you but still feels emotionally resonant. This cultural phenomenon is gaining traction and growing in relevance as more of the past is being shared (and generated) online. Old photographs are especially good at triggering anemoia: scenes from another era, frozen moments that hint at lives just out of reach. You can study them, imagine them, even narrate them, but they remain somewhat inaccessible. Anemoia Scent offers a way into those memories. Unlike images or text, smell does not describe the past; it immerses you in it. Of course it is highly subjective, a bit fuzzy, and difficult to pin down exactly, but then that’s closer to how our reconstructive memory systems actually work. A scent doesn’t tell you what happened. It places you somewhere and lets your mind fill in the rest. This makes it particularly suited to anemoia, where perhaps the emotional pull matters more than factual accuracy. reconstructive memory From Photograph to Fragrance From Photograph to Fragrance At a high level, the device translates an analogue photograph into a custom fragrance through a multi-stage AI pipeline from initial perception to narrative reasoning, semantic mapping, and physical actuation. Step 1: Perceiving the Image Step 1: Perceiving the Image The process begins with a vision-language model (VLM) that analyses an analogue photograph and generates a descriptive text caption. This caption is displayed on the device’s integrated screen, creating a shared interpretation of the image. vision-language model (VLM) Step2: Human-in-the-Loop Memory Synthesis Rather than allowing the AI to fully determine meaning, the Anemoia Device introduces a co-creative process designed to increase human agency in the process through three physical rotary dials. These dials shape how the image and caption are interpreted without overwhelming the user with free-form input. Dial options for the Anemoia version of the device are designed to immerse the user in an unfamiliar scene and connect them to a memory of a moment they never lived: Dial 1 — Subject (Perspective): Selects a point of view within the image (e.g. old man, tree, bicycle) and classifies it as living or non-living, which defines parameters in the subsequent step: Dial 2 — Time (Temporal Context): Places the subject within a lifecycle: a) Living: childhood, youth, adulthood, elderly b) Non-living: raw material, manufacture, in-use, decay Dial 3 — Mood (Affective Tone): Assigns an emotional tone: "happy", "sad", "calm", or "angry" Dial 1 — Subject (Perspective): Selects a point of view within the image (e.g. old man, tree, bicycle) and classifies it as living or non-living, which defines parameters in the subsequent step: Dial 2 — Time (Temporal Context): Places the subject within a lifecycle: a) Living: childhood, youth, adulthood, elderly b) Non-living: raw material, manufacture, in-use, decay Dial 3 — Mood (Affective Tone): Assigns an emotional tone: "happy", "sad", "calm", or "angry" Together, these dial inputs transform an image into a situated, temporal, emotional perspective, something closer to memory than metadata. Materializing Memories Materializing Memories Prompting But Not As You Know It Prompting But Not As You Know It The dial selections and image caption are synthesised into a structured prompt created by a large language model. The LLM produces this concise narrative as an interpretive bridge between vision and smell. This narrative is displayed to the user, giving them something to read, while the cross-modal translation from image-to-text-to-scent takes place. Using few-shot in-context learning, the system maps narrative meaning from the prompt to an olfactory output. Semantic-to-Olfactory Translation Semantic-to-Olfactory Translation To enable this translation, we combine a curated scent library and olfactory knowledge base. Our scent library is comprised of 45 scents, with each scent annotated with a set of semantic descriptors, including primary olfactory notes, associated concepts, and emotional qualities. The olfactory knowledge base captures higher-level relationships between language and scent. It contains example pairings of short narratives, their corresponding ideal scent vectors, and a textual rationale explaining the semantic mapping. During generation, the system uses few-shot in-context learning: a small number of these examples are dynamically prepended to the prompt, guiding the LLM’s reasoning. The model then analyses its own generated narrative and, following these examples, outputs a proportional blend of up to four scents that best balances semantic alignment, emotional tone, and thematic coherence. This works to enhance the consistency and coherence of pairings. Once a scent formula is generated, it is physically rendered by a custom olfactory display. Pump timings are computed from the scent proportions and executed via peristaltic pumps that draw from glass reservoirs into a shared blending vessel. The vessel can then be removed and the scent smelled, Design and Technical Overview The physical design of the device is based around the metaphor of distillation. As a result, the device is vertically organised around this flow, image at the top, memory processed through the middle, scent emerging below. At the centre is a XIAO ESP32S3 microcontroller handling timing, sequencing, and coordination. On the input side, a standard USB webcam captures images for processing. User interaction is handled through three rotary encoders. These inputs are read by a microcontroller and reflected back to the user via a 20×4 I2C LCD. Vision and language models operate on a separate host machine. Scent is rendered using a set of peristaltic pumps connected to glass reservoirs, each containing a base fragrance. Pump runtimes are calculated directly from the generated formula, with volumes drawn through silicone tubing into a shared blending vessel, where individual components merge into a single, blended fragrance. A custom PCB CNC-milled in-house forms the backbone, while 3D-printed mounts position pumps, tubing, and reservoirs without enclosing them completely. These internal systems sit within a modular aluminium frame with laser-cut frosted acrylic panels. The translucency allows the inner workings to remain partially visible, partly for aesthetics, but also reinforcing the device’s focus on transformation. On the software side, the system operates as a multi-stage translation pipeline rather than a single generative model. Multimodal image analysis first extracts semantic structure from the photograph. This representation is then combined with the user’s dial selections and passed to a language model, which generates the prompt used for the scent-formula creation. Towards Multisensory AI Systems By using scent as a core output modality, the device is able to evoke a sense of presence and embodiment that audiovisual media alone cannot. In parallel, the tangible interface enables a more engaging experience. This combination creates the ideal conditions for encountering and incorporating a multi-sensory artifact of a fictional yet resonant past. The system in its current form does not attempt to create precise scentrealism. Instead it treats scent as an interpretive medium. The goal for this stage of the work is to create a scent palette that feels plausibly connected to the image and its imagined history. incorporating scentrealism By translating photographs into fragrances, the device generates something closer to a synthetic memory than a representation. In this sense, scent becomes a way of co-creating a memory rather than retrieving it. The photograph provides a visual anchor, AI models provides a narrative structure, and the fragrance completes the loop by engaging the body. What emerges is not a recollection of your past, but a felt experience of a past you never experience – the feeling of anemoia. If you would like to try the device, please register your interest. please register your interest
