Open-source BabyCoder gives small local LLM agents a dreaming sleep loop that generates new questions

Illustrious_Matter_8 · reddit · 2026-09-30

A Reddit developer released BabyCoder, an MIT-licensed open-source toolkit for local agents running small models on 12GB VRAM or less (via LM Studio or any OpenAI-compatible server), after a major rewrite. Python owns the structure (files, memory, loop, rules) while the model only makes the next decision; on failed tool calls, Python suggests what was probably meant, keeping small models on track. Tested with Gemma 12B and Qwen models.

The new Lisa-agent runs continuously: she responds when spoken to, works on her own goals otherwise, gets tired and sleeps. During sleep she dreams about random memories, merges duplicates, and sometimes wakes with a question she researches the next day — e.g., after a conversation about the Dutch word "Herfst", she dreamed about seasons and words, woke wondering whether different words for one thing change its meaning, and independently found the Sapir-Whorf hypothesis.

Key differences from other sleeping-agent projects:

Original post →

More from coding & agent

coding & agent channel →