Recurmind
System overview · the AI being improved

Nous

Nous is the AI that Hebb improves. It is not one big model. It is a brain made of many tiny networks, a body it acts through, and a small language part that only speaks.

01Architecture

Components

Nous: one day-and-night cycle Aegis chooses a practice task. REFLEX, the brain, acts in the body, a browser in a hardened Linux sandbox. An exact checker scores the result. Only verified outcomes go to memory. At night, sleep trains a small LoRA in the cloud on those memories, and the next day of practice begins. Separately, the speaker, a small language model, speaks; its character grows from the memories. Aegis · manager, no voice Chooses practice; invents new goals practice task Brain · REFLEX Tiny networks, ~1.3M trained params action Body Browser in a hardened Linux sandbox outcome Exact checker Success or failure, no partial credit verified only Memory Nous remembers its days night Sleep Small LoRA, trained in the cloud next day · play Speaker · Qwen3.5-4B Speaks; does not decide how Nous acts character
Fig. 1 — One day-and-night cycle. Solid lines: the learning loop. Dashed: Aegis and the speaker, which sit outside the acting path.
BrainTiny networks
Many small, separate networks instead of one large model. The one that acts is called REFLEX.
BodySandboxed browser
Nous acts in a web browser inside a hardened Linux sandbox, cut off from everything it shouldn't touch.
SpeakerVoice
A small language part based on the open Qwen3.5-4B model. It speaks; it doesn't decide how Nous acts.
AegisSilent manager
A manager model with no voice. It chooses what Nous practises and invents new practice goals.
02REFLEX

REFLEX, the part that acts

About 1.3 million trained parameters sitting on a frozen small text encoder. It runs on a weak laptop CPU.

PolicySuccess
REFLEX~1.3M trained params; about 7× faster56.6%
4B language modelSame job, same test57.5%

Fixed test: 87 MiniWoB++ web tasks, each run with 5 seeds. REFLEX ties the 4B model while running about 7× faster.

03Cycle

A day in Nous's life

DayPlay
Nous practises web tasks all day. Every outcome is checked by an exact checker: it worked, or it didn't.
NightSleep
A small LoRA is trained in the cloud on Nous's own memories of the day.
Memory
Nous remembers its days.

Nous learns only from its own verified outcomes. It is not trained on demonstrations.

04Aegis
A/B test in progress

Aegis, the silent manager

Aegis chooses what Nous practises and invents new practice goals. Right now it is being A/B tested against a fixed curriculum, to find out whether it actually helps.

05Character

Character

Nous's character isn't written by anyone. It is growing from its own memories. Hebb won't put words in its mouth here.

In Nous's own wordsComing soon.
Talk to Nous