(pneuma) "breath" — the vital force that animates matter

A cognitive runtime for self-improving LLM agents with temperature-staged dreaming. Agents alternate between wake cycles (context-based reasoning) and sleep cycles (weight-level learning through LoRA fine-tuning), mirroring biological NREM/REM sleep.

The Cognitive Cycle

Wake Phase (Daily)

Perceive Plan Learn Execute Evaluate Deliver

Each day the agent perceives its environment, plans a goal using a developmental temperature schedule, learns from web sources, creates an artifact, self-critiques, and delivers via Telegram.

Sleep Phase (Weekly)

Collect NREM Consolidate REM Dream QLoRA Train Evaluate Deploy

Weekly sleep consolidates accepted work at low temperature (NREM) and generates creative cross-domain mashups at high temperature (REM). QLoRA fine-tuning bakes patterns into weights.

The Experiment — 4 Conditions, 12 Weeks
Bot A
Baseline
No fine-tuning, fixed temp 0.7. Context-only memory. The control.
Bot B
Flat Dreaming
Weekly QLoRA at uniform temp 0.7. All dream content saved to KB.
Bot C
Staged Dreaming
NREM (0.2) + REM (0.8-1.3). Developmental temperature schedule.
Bot D
Staged + Amnesia
Like Bot C, but dreams go to weights only, not explicit memory.
Key Hypotheses

H1: Self-training bots (B/C/D) outperform baseline (A) after week 4
H2: Temperature-staged bots (C/D) surpass flat training (B) on creativity
H3: Flat training (B) shows model collapse by weeks 8-12
H4: Dream amnesia (D) produces cleaner knowledge base than full recall (C)
H5: Staged bots exhibit a developmental arc: chaotic → refined → mastery
H6: Creative breakthroughs cluster around high-temperature spike days

Project Links
Technical Stack

Model: Qwen3-14B via Ollama (NVIDIA L4, 24GB VRAM)
Orchestration: n8n (self-hosted)
Database: 4x PostgreSQL 16 (one per bot)
Fine-tuning: QLoRA via unsloth/PEFT
Storage: GCS for knowledge bases & adapter checkpoints
Delivery: 4 Telegram bots (one per condition)
Infra: GCP g2-standard-4, Terraform, Docker Compose, Caddy