Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

Mnemosyne Made Kubernetes Write Music and Prose: 50+ Works in 4 Days Double as a Backup

The author tested whether a Kubernetes cluster can inspire AI: the Mnemosyne operator reads workload, incidents and deployments and passes them to agents as "climate" — tension, crowding, scars. Agent composers and writers produced more than 50 works in four days. Each one weaves in an encrypted fragment of the cluster's state, so the cluster can be reconstructed from the works at any point in time. For a practitioner, it is an unusual way to keep agent output and backups in one place.

AI-processed from Habr AI; edited by Hamidun News
Mnemosyne Made Kubernetes Write Music and Prose: 50+ Works in 4 Days Double as a Backup
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1The operator introduces ten custom Kubernetes resource types
Key metric #2Agents never learn the facts of incidents — they receive only the cluster's "climate"
Key metric #3A creator lives about two real years on average

Usually, AI agents in infrastructure are tasked with triaging tickets, reviewing code or fixing pipelines. The author of the Mnemosyne project went in a different direction and asked: what if Kubernetes itself became the source of events that give rise to artists and critics and shape their mood? The result is a workshop of composers and writers running on language models inside a single cluster, and in their first four days of continuous operation they wrote more than 50 works. For an engineer, the key point is that all these works also serve as a backup of the cluster: from them, its state can be restored at any point in time.

How the workshop is built

Technically, Mnemosyne is an extension on top of Kubernetes: an operator with ten custom resource types. It reads the cluster state through the standard API, packages what it reads into "climate" and hands it to agents who live in the cluster in the roles of creators, critics and a chronicler. Everything the agents produce, from the first idea to the final recording of a symphony or story, is stored as Kubernetes objects and files on a volume.

Each creator has its own name, language, teachers, taste and character. They are born, learn, write, fall ill, quarrel and befriend, age and die. The median lifespan is about two real years, so over an average lifetime the cluster sees three to four generations of creators. The results include symphonies of Polish classicism, English aleatoric music, Portuguese epistolary works, and fairy tales and Italian monologues by barbers, watchmen and firefighters. People and admins do not tell the creators what to write: the cluster owner configures only the infrastructure.

Climate instead of an event log

To explain the central engineering idea, consider an example. Suppose the payment service in the cluster goes down. The agent will not receive a message about it. It will be given only a shade: the autumn was ragged, and a wound remained after it. The workshop sees the configuration (what is installed, how many tenants there are, how the network is isolated), the load (resource fill level, pod churn, latencies) and the events (incidents, deployments, creation and deletion of tenants, migrations). Load is measured relative to the cluster's own ordinary week, so a bank's quiet nights do not count as a storm.

All of this is condensed into a dozen qualities: turbulence, volatility, crowding, tension, scars, losses, renewal, evenness, entanglement, enclosure, age, abundance or scarcity. The climate has memory: recent events weigh more, but nothing is forgotten entirely. The author justifies the choice with literature: Borges wrote about Funes, who remembered everything and therefore could not think. A full event log is useless to an agent, like a map at a scale of one to one.

Three principles follow from this. The first is no literalness: an incident does not turn into a mandatory screech of the violin, and the same climate affects different characters differently. The second is that the process is also an output, so ideas, questions from the internal critic, drafts and revisions are stored together with the finished piece. The third is that speed does not matter: memory comes first, then quality, then load on the cluster.

A backup assembled from symphonies

Each artifact weaves in an encrypted fragment of the cluster's state. The author refers to Pavić's Dictionary of the Khazars: a book-structure of data that must be cross-checked across several copies. On the same principle, the backup is assembled only from many works at once. The restored cluster matches the original letter for letter, but, like Pierre Menard's Don Quixote, it is written after the incident, in another year and by a different author.

The scattered details of this world are also governed by engineering logic. A creator occupies a place in the workshop while it creates, not while it is alive. The novelty of works is measured against the creator's own catalogue, and repetition counts as a disease that critics and the author may or may not notice. A dying creator has a "secret miracle": time to finish the last piece. From this emerge fates that no one wrote by hand: a meteor, a late flowering, a master with a long life, an author recognized posthumously, a founder of a school. To test the system, the author also ran a synthetic cluster called "Leonia," named after Calvino's Invisible Cities.

What practitioners can take from this

First, Kubernetes objects can serve as an agent's decision log: the idea, the doubts and the drafts sit alongside the result and are accessible with ordinary tools. Second, an agent does not have to be fed raw logs: a compressed "climate" saves context and leaves room for variety in responses. Third, the idea of a backup assembled from many artifacts shows that system state can be stored inside agents' output data.

The caveats are obvious. This is an artistic experiment, not a product, and it has been running for only a few days. The author does not discuss token costs, the models chosen, or the load on the cluster in the accessible part of the text. Even so, as an example of an operator in which agents became full residents of a cluster, Mnemosyne deserves attention.

Story

Key Questions Answered

How the workshop is built?

Technically, Mnemosyne is an extension on top of Kubernetes: an operator with ten custom resource types. It reads the cluster state through the standard API, packages what it reads into "climate" and hands it to agents who live in the cluster in the roles of creators, critics and a chronicler. Everything the agents produce, from the first idea to the final recording of a symphony or story, is stored as Kubernetes objects and files on a volume. Each creator has its own name, language, teachers, taste and character.

What are the key details on Climate instead of an event log?

To explain the central engineering idea, consider an example. Suppose the payment service in the cluster goes down. The agent will not receive a message about it. It will be given only a shade: the autumn was ragged, and a wound remained after it. The workshop sees the configuration (what is installed, how many tenants there are, how the network is isolated), the load (resource fill level, pod churn, latencies) and the events (incidents, deployments, creation and deletion of tenants, migrations).

What are the key details on A backup assembled from symphonies?

Each artifact weaves in an encrypted fragment of the cluster's state. The author refers to Pavić's Dictionary of the Khazars: a book-structure of data that must be cross-checked across several copies. On the same principle, the backup is assembled only from many works at once. The restored cluster matches the original letter for letter, but, like Pierre Menard's Don Quixote, it is written after the incident, in another year and by a different author.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…