I'm Vladimir from Toronto. This is the Sleeping LLM:
A system that gives LLMs persistent memory by injecting facts directly into model weights during wake (via MEMIT) and consolidating them into LoRA during sleep. No RAG, no database — the knowledge is in the weights. MEMIT is short-term memory; LoRA is long-term memory. Sleep is the transfer between them.
Key results: 100% recall at 60 facts on 70B with zero perplexity drift. Per-fact graduated consolidation achieves 100% advancement and 1.00 chat recall at all scales (5–20 facts), with MEMIT edits dissolving on schedule as LoRA absorbs each fact — making effective lifetime capacity unbounded.
6 research papers covering the full trajectory from LoRA prototype through the alignment tax discovery to per-fact graduated consolidation:
| Paper | Topic | DOI |
|---|---|---|
| 1. Sleep-Wake Consolidation | LoRA sleep-wake on 3B | 10.5281/zenodo.18778760 |
| 2. The Alignment Tax | RLHF suppresses continual learning at scale | 10.5281/zenodo.18778762 |
| 3. Dual-System Memory | MEMIT+LoRA, null-space constraints | 10.5281/zenodo.18778764 |
| 4. Sleeping LLM: Two-Phase | SWS+REM sleep, per-fact staged consolidation | 10.5281/zenodo.18778766 |
| 5. Sleep-Wake Convergence | MEMIT-only, convergence proof, pruning death spiral | 10.5281/zenodo.18778768 |
| 6. Per-Fact Graduated Consolidation | Resolves capacity ceiling, unbounded lifetime memory | 10.5281/zenodo.18779159 |
Code + experiments: vbario/sleeping-llm
Other projects:
- ResonantMail.com — Get your first 100 customers with 1:1 personalized email
- DefaultTools.com — Collapse 10-20 tabs into one place
- ChatSetter.ai — Instagram DM tool that books appointments
- ClawzBot.com — Open source AI robot (coming soon)
Fields of interest: continual/lifelong machine learning, intelligence orchestration, digital biology, cognitive ergonomics