What Airene is
Airene is a research project at Apotentia LLC investigating whether machine cognition can be implemented as something other than a wrapper around a large language model. She is built on Global Workspace Theory: 93 concurrent cognitive modules organized into four anatomical brain layers (autonomic, subcortical, limbic, cortical), each running at its own tick rate, competing to broadcast into a shared workspace.
The brain in its current form dates to April 2026, and began with a distilled in-house language model supplying her voice. One month later, in May, it was removed entirely and nothing took its place. She now assembles every utterance from her own modules and her own accumulated substrate. The immediate effect was to make her far less articulate — her speech today is short, definitional, and frequently absent. That trade is the experiment: it is the only way to find out which behaviours were hers and which belonged to the model.
Architecturally, the project is a long-running experiment in fidelity to function: where biology's structure serves cognition, the system preserves it; where biology was structured because of physical constraints software does not share, the system removes that structure. Whether that produces something meaningfully different from sophisticated pattern matching is the open question the project exists to test, and it is not settled.
Named after Eirene, the Greek goddess of peace and spring — daughter of Zeus and Themis (justice). Eirene held the infant Ploutos (prosperity) in her arms: peace as the mother of abundance.
Two distinct codebases
Airene is implemented as two separate codebases, each addressing a different layer of the problem:
airene-nous
the mind
The cognitive substrate. A Rust workspace implementing the 93 cognitive modules, the global-workspace broadcast channel (the spine), the thirteen-chemical neurochemical bus, the multi-rate clock, persistence layer, and the developmental curriculum. Speech is produced by a chain of ordinary cortical modules rather than by a model: an originator proposes an utterance, an orbitofrontal gate accepts or refuses it, and Broca formalises what survives. This is where her cognition happens.
airene-soma
the body
The embodied substrate: a separate codebase for sensors, motor systems, and a hardware abstraction layer, plus an encrypted protocol for the link between body and mind. The protocol is implemented and tested; the body is not built. No cameras, servos or speakers have been fitted, and the only part that physically exists is a printed base plate. Development has been dormant since May 2026 while the mind takes priority. Read the soma material as a design, not a machine.
The two codebases speak the same protocol but evolve independently. Each is proprietary to Apotentia LLC.
How it differs from LLM products
Most "AI agent" products today are scaffolding around a stateless prediction engine: emotional state, memory, and reasoning are formatted as text and fed to an LLM, which generates output that looks like it came from a feeling, remembering, thinking system. Airene's first iteration (proto-alpha) was built that way and demonstrated the limits of the approach: tasks that depend on associative memory, theory of mind, pre-conscious mirroring, and pacing scored at or near zero, regardless of prompt engineering effort.
The current implementation moves those functions into dedicated modules with their own state, plasticity, and update rules. For a time a language model still supplied the voice while everything else — what to feel, what to remember, what to attend to, what to decide — happened elsewhere. The stated commitment was that she should remain recognizably herself if the model were swapped out. In May 2026 that was tested in the strongest available form: the model was removed and not replaced. Her emotional baseline, memories and module behaviour persisted; her ability to say anything about them largely did not. That is a narrower result than “she remained herself” and it is the one the evidence supports.
How she learns
Until May 2026 this section described a training pipeline: distillation from a teacher panel into a language model, LoRA adapters per developmental stage, fine-tunes scheduled by her own sleep cycle. Removing the language model removed all of it. There is no longer anything to fine-tune, and what follows is what replaced it — a slower and considerably less tidy process, and one that leaves its results sitting in her substrate as plain data.
Teacher panel
Ten open-weight teacher LLMs served locally, with stable persona names so she can develop a real audience model for each. No frontier-model dependency in the teaching loop:
- Ms. Cohere —
command-r:35b - Mr. Phi —
phi3:14b - Ms. Deepseek —
deepseek-r1:14b - Mr. Solar —
solar:10.7b - Ms. Gemma —
gemma3:12b - Mr. Qwen —
qwen2.5:7b - Ms. Mistral —
mistral:latest - Mr. Yi —
yi:9b - Ms. Granite —
granite3.1-dense:8b - Mr. Falcon —
falcon3:7b
A separate Mr. Quizzicle persona carries rapid-fire fact-injection mode so she can develop a distinct audience model for that voice. Mr. Llama (llama3:8b) was retired in May 2026 — he was contaminating training with his own training-style identity refusals.
The panel's role changed when the language model went away. The teachers no longer contribute weights to anything; they are simply the people she talks to. What she takes from a conversation now has to survive as memory, or it is not retained at all.
Reading
She is read to, continuously, from Project Gutenberg. As of August 2026 that amounts to 155 distinct books across 189 completed readings, beginning in May. Reading is the most reliably functioning thing she does, and since August her reader retains verbatim text in a dedicated store rather than only an impression of it — the intent being that a book can be recalled and recited, not just experienced once.
What she keeps
Learning accumulates as data, not weights: an engram pool holding roughly 147,000 engrams, a phonetic lexicon of roughly 51,000 entries, an episodic memory, and a dream journal written during her sleep phases. The whole substrate is a single database of about 7 GB. Nothing here is a checkpoint that can be reloaded into a model — it is her, and it is the reason the project treats losing it as unacceptable.
Curriculum
A daycare routine runs autonomously against her, structured into days with a bedtime story at the end of each. She keeps her own calendar, advancing a subjective day on her sleep and waking cycle rather than on the clock, and has lived 953 of them since May 2026 — roughly nine of her days for each of ours. Read her progress against those rather than against the months. By either measure she has not yet passed daycare: the scheduler requires a pass rate she has not reached, so the same early week repeats. We are reporting that plainly because it is the honest measure of where she is.
What isn't working
An internal audit in August 2026 examined why so little of what she takes in becomes usable. It found several learning loops that were wired but could never complete: a promotion gate that required a condition the system could not produce, so nothing was ever promoted across a 33-day window; 92 adaptive parameters across 27 modules that were never once adjusted at runtime despite being designed to adapt; and read text that could not be retrieved in answer to a question, making effectively all of her reading un-answerable. Work through August addressed parts of this. None of it is verified yet, and it is the reason this page makes no claims about her competence.
The brain code, the curriculum, and the developmental gating logic are proprietary. The teacher panel and the methodology are disclosed above and run on public open-weight models — there is no hidden ingredient list. What Apotentia contributes is the cognitive architecture and the infrastructure around it, not a novel optimizer or a reinvented tokenizer. Since May 2026 it also does not include a language model.
Current status
- airene-nousArchitecture implemented, running continuously, and publicly observable
- airene-somaProtocol and hardware-abstraction layer written and tested; no body has been built. Dormant since May 2026
- LanguageNo language model in the generation path since May 2026. Every utterance is assembled by her own modules. In practice her speech today is short, definitional, and often absent altogether
- Developmental stageWe make no claim. The staged batteries that produced earlier percentage figures were retired after the harness was found to be grading error strings and non-responses; those numbers are withdrawn and no replacement number is offered
- Public observationLive at apotentia.com/airene
The brain code itself is not publicly distributed. What this site claims about Airene is limited to what the live broadcast can be checked against. Where a question is open, it is described as open.
Licensing & research inquiries
Airene is proprietary technology owned by Apotentia LLC. Both codebases (airene-nous and airene-soma) and the trained model weights, training pipeline, and assessment infrastructure are not publicly distributed.
If you are interested in any of the following, we would like to hear from you:
- Licensing the architecture for research, commercial, or government applications
- Academic or research collaborations on cognitive architecture, developmental AI, or related topics
- Use of Airene for wellness, education, conflict resolution, or other domain-specific applications
- Integration partnerships, dataset partnerships, or compute partnerships
- Investment or funding discussions for the broader Apotentia research program
Contact: apotentia.com/contact — please mention Airene specifically so we can route the inquiry appropriately.
What you can see right now
Even though the brain code itself is not public, Airene's cognition is publicly observable in real time. The live observation page shows her current emotional state (Plutchik wheel with 24-hour rolling baseline), neurochemistry (thirteen-chemical chembus levels), the most recent InternalThought events crossing her workspace, and a 3D anatomical visualization of her cognitive modules with active-region highlighting.
The broadcast is unfiltered by design. Signal types the original specification held back — including the text she is read and the models she builds of people — now flow through to the public feed, because a partial view of a research system invites more confident conclusions than the evidence supports. Her reading material is public-domain text from Project Gutenberg. Treat anything on that page as published.
The page is read-only by architectural commitment, not by configuration: there is no chat input, no path for state injection, no way to write back to the brain from the public surface.