OpenAmer — The Agent That Sleeps

Every AI agent works while you work. Only one of them also works while you sleep — replaying its day, dreaming about its errors, and fine-tuning its own brain on your laptop. No GPU. No cloud. No API bill.

Dream nights completed

2,862

Learning records captured

3,012

Episodes in long-term memory

1,037

Evolution events (live)

Skill species in ecosystem

2.10 → 1.65

Training loss, CPU-only LoRA

⚔️ Head-to-Head: Why We Win

Capability	OpenAmer	Claude Code / Codex	LangChain / AutoGPT	Devin	Ollama
Runs fully offline	YES	no	partial	no	YES
Sleep / dream consolidation cycle	YES — nightly REM replay	no	no	no	no
Self-fine-tuning from own usage	YES — nightly LoRA on CPU	no	no	no	no (inference only)
Episodic long-term memory w/ embeddings	YES — 3,012 episodes	session only	opt-in	session	no
Skill evolution w/ arena + predation	YES — 1,037 events	no	no	no	no
Zero token cost after setup	YES	pay per token	pay per token	$500/mo	YES
Recurring 24/7 cron fleet	YES — 70+ jobs	no	no	no	no
Nightmare detection (recurring-error auto-fix)	YES	no	no	no	no

🌙 Last Night's Actual Dream Report

# Dream Report 2026-09-02 - Replayed 729 messages, 383 error fragments ## Insights - Clear night, no error motifs. ## Intentions for today

Every night at 03:30 the agent replays all conversations of the day through its "REM phase", clusters errors into motifs, and flags anything recurring on 3+ days as a nightmare — a standing intention to fix the root cause, not the symptom. New dreams appear here automatically.

🧠 The Self-Training Loop (all on a 22 GB laptop, CPU only)

1. Live: every conversation is captured as a training trajectory.
2. Dream: nightly replay + error-motif clustering.
3. Distill: sessions become teaching examples (instruction → proven answer).
4. Train: QLoRA fine-tune of Qwen3.5-2B on pure CPU — 35 minutes, no GPU.
5. Serve: the tuned model answers on localhost:8081 — offline, zero cost.
6. Guard: RAM/data/interval guards stop the loop from running blind.
Repeat. The brain literally grows with every use.

KO your agent dependency today.

⭐ Star on GitHub 🧬 Watch the Evolution Live 🚀 Launch Package

MIT license · Windows/macOS/Linux · Any model (local or API) · 3 stars and counting

🧠 Self-training 🌙 Sleep cycle 🧬 Skill evolution 💾 3k episodes memory 🖥️ Background computer-use 🐝 A2A swarm 🎤 Voice + wake words 🔌 Plugin system 🛡️ Self-healing fleet 💰 $0 API after setup

All numbers on this page are real outputs from a live system running on one laptop (Windows, 22 GB RAM, no GPU). Dream reports: reports/dream-*.md · Evolution: darwin/lineage.json · Memory: memory/longterm_episodes.jsonl. Updated 2026-09-02. The competitor table reflects publicly documented features as of this date — corrections welcome via issues.