Scents and algorithms

Anemoia Device is an experimental device developed at the MIT Media Lab (Massachusetts) by researcher Cyrus Clarke, capable of translating images into fragrances. Smellscapes is a participatory project created in Turin by artist Elena Mazzi, in which the olfactory memory of a territory emerges through collective practices of exploration and creation. The comparison highlights two complementary approaches to the construction of sensory memory: one mediated by technology and the other rooted in community experience. Together, they offer a starting point for reflecting on how memories may be produced, interpreted and shared in the future.

At the MIT Media Lab (Cambridge, Massachusetts) there is a machine that promises to turn photographs into smells. But the Anemoia Device, developed by researcher and designer Cyrus Clarke, is not merely a curious experiment in Artificial Intelligence applied to perfume. It is an attempt to rethink memory in the digital age. Memories are increasingly “externalised” in clouds and hard drives, yet this device distils them into fragrances, exploring that nostalgia for a time never lived which dominates our present. «I do not think everything has to be a machine that steals your attention, and I believe we can and should create new forms of technology that make you stop, breathe, and notice the world around you again», Clarke explains. The real focus is therefore not only AI, but the future of our memory: what we are preserving and how we are remembering. And if the future of memory will be digital and will also have a smell, the political and social question is clear. Will we recognise it as ours, or will it be the product of an algorithm that selects and returns a past more seductive than reality, standardising memories and influencing collective narratives?

The Anemoia Device

The Anemoia Device looks like a retro-futuristic console: a metal structure, a green screen, three knobs and a glass beaker that collects the fragrance. Behind this almost poetic form, however, lies a precise process.

First comes visual analysis. The user inserts a photograph and a computer vision model analyses the image, identifying what Clarke calls “subjects and objects of interest”. This means the system does not simply recognise what it sees in general terms, but selects elements it considers relevant: people, objects, luminous details, materials, atmospheres. It is the first level of interpretation: the AI reorganises the image according to its own semantic logic.

Then comes narrative co-creation. This is where the user interacts with three physical knobs used to guide the desired subject, time and mood. Clarke explains that he did not want an interaction based on writing a prompt, where the user must “guess” the right words. «You are not trying to say the right thing to the algorithm, it is more like tuning an instrument», he says. «The system proposes, you orient it, and the result feels co-authored. More embodied, more intentional, and less like a chatbot trying to guess what you meant». The knobs also reduce the complexity typical of generative AI: that overly open space which creates a certain “cognitive load”. In other words, it is difficult to know what to write when faced with a blank page.

Finally, there is olfactory distillation. The last phase in which «you take a dense, layered memory artefact and transform and compress it into something». In other words, the image becomes a narrative, and that narrative is translated into a chemical formula. The system draws from a library of nearly fifty essences. Interestingly, the device does not explicitly follow the classical perfume structure (top, heart and base notes): «I was curious to see what the algorithm would do without that framework», Clarke explains.

The idea of “mapping” olfactory experience is already a well-established direction in scientific and cultural research. In Europe, for instance, projects such as ODEUROPA – Negotiating Olfactory and Sensory Experiences in Cultural Heritage Practice and Research (2021–2023) use AI to analyse texts, images and historical archives, identifying olfactory references and building a knowledge graph: an immense semantic database of European smell capable of tracing links between fragrances, places and cultures over time. Another project, Digitising Smell:From Natural Statistics of Olfactory Perceptual Space to Digital Transmission of Odors (2024–2030), addresses the lack of data on human olfactory experiences in order to digitise what we smell and create something akin to an olfactory equivalent of Google Street View.

With the Anemoia Device, however, Clarke starts from a slightly different premise:

«I always try to begin “outside” technology, with a vision of the world I want to live in», he says.

His research revolves around the idea of “materialising memory”, imagining a world where information is expressed not only through screens and silicon but also through molecular and sensory media. Drawing on the intuition of American computer scientist Mark Weiser about computing woven into everyday life¹, Clarke asks what it would mean for memory to once again be intertwined with materials, surfaces, even the air itself. The Anemoia Device was born precisely to explore this possibility.

Smell becomes a medium for generating sensory narratives that are co-created, immediately inhabitable and deeply lived. The machine shifts from the register of scientific mapping to that of poetic and political experience: an experiential, domestic and almost ritual object. The very name of the device derives from a term coined by writer John Koenig in The Dictionary of Obscure Sorrows²: “Anemoia” is the nostalgia for a time never lived. Clarke describes it as «one of the most relevant words for an era increasingly characterised by generative AI», noting how we now live immersed in digital archives, reconstructed images and retro aesthetics we can inhabit without ever truly having experienced them.

The algorithmic risk: who decides what and how we remember?

«I would say that the machine, in its current form, works “best” from the user’s perspective when an image contains elements that can be semantically aligned with the fragrance library», Clarke explains. «A great example is an image of an old bakery with four women and many pastries. The fragrance library contains several references linked to this scene, including artisan bread, sweet biscuits, vanilla pod, old books, fresh cotton and cinnamon. All these elements reliably transport the user into that environment, whether they focus on the pastries, the counter or the women in the image». The Anemoia Device does not reproduce a real memory. Rather, it evokes sensations similar to a memory, involving a degree of algorithmic mediation. Each image becomes a “memory” to be transformed, and the selection of what becomes meaningful no longer depends on direct experience, but on the algorithm.

This aspect has clear political and cultural implications. Personal memories feed collective history: traditions, relationships and shared identities are built through individual experiences and small everyday narratives. When the selection of memories passes to an algorithm, as in the Anemoia Device, this network risks being simplified, favouring recognisable elements at the expense of unique details. Machines of this kind therefore have a potentially political effect. By offering an interpretative framework of the past, they influence how a group perceives and shares its own memory. Once filtered and standardised, memory leaves little room for marginal details and experiences, both individual and collective. Delegating to an algorithm the choice of what is relevant means accepting the risk of transforming something complex and plural into a simplified storyline. What is not selected by the algorithm risks disappearing, while emphasised elements become the “correct”, and therefore dominant, version of the past.

Two ecologies of memory

The comparison with participatory experiences such as the artistic project Smellscapes in Turin highlights an alternative: memory mediated by the community rather than by the algorithm, where the selection of memories remains a collective and democratic act. The initiative was led by artist Elena Mazzi, who collaborated with the PAV – Parco Arte Vivente and with perfumery experts to design a sensory exploration of the territory. In 2022, as part of the project Smellscapes, conceived by Elena Mazzi for her exhibition at PAV – Parco Arte Vivente in Turin, the local community of the Filadelfia neighbourhood was involved in a participatory process dedicated to the sensory exploration of the area. The project began with exploratory walks through the Filadelfia and Lingotto districts, during which participants collected materials and scent stimuli characteristic of the territory, while also investigating its urban, social and historical dimensions. In the following workshop, the collected materials were analysed and transformed through the Frantoio Sociale – a machine designed by Studio GISTO / Alessandro Mason in collaboration with Matteo Giustozzi and Hund Studio – capable of pulverising urban artefacts and preparing them for further olfactory processing. During the final meeting, participants, guided by perfumer Roberta Conzato, transformed the powders into essences. This process allowed them to observe how the collected smells could be translated into fragrances. In this way, the project turned collective memory into a shared and participatory process: the stories, places and sensations of the neighbourhood emerged through direct interaction between artists, experts and citizens.

One could say that in Turin the power to select smells and narratives remained in the hands of the community, allowing the diversity of lived experiences and the richness of local details to be preserved, while contributing to the creation of a possible olfactory mapping of the territory.

These two approaches, however, do not necessarily exclude one another. On the contrary, they can be read as complementary perspectives through which to question the relationship between technology, perception and memory. What becomes essential is maintaining a critical awareness of the limits and risks of both: on the one hand, the danger that algorithmic systems privilege more evident or semantically recognisable elements, flattening the diversity of memories. On the other, the challenge of translating intimate individual experiences into collective processes without losing their most personal nuances.

---

1. The reference is to the concept of ubiquitous computing developed by American computer scientist Mark Weiser in his article The Computer for the 21st Century (1991). In this text, Weiser argues that the most advanced technologies are those that become invisible because they are integrated into the fabric of everyday life: «The most profound technologies are those that disappear. They weave themselves into the fabric of everyday life until they are indistinguishable from it». ↩︎
2. Visit the official website: https://www.thedictionaryofobscuresorrows.com/ ↩︎