What is dream-cycle-lab?

Dream Cycle Lab is a research test bench for a specific problem in AI: how to make a language model actually learn new facts by updating its own internal weights, instead of relying on retrieval, meaning looking facts up at query time as most systems do today. The project calls this the write side problem, since the AI field has good solutions for reading from external memory but almost no tooling for writing new knowledge into model weights without breaking what the model already knows. The core metaphor is a dream cycle, based loosely on how biological memory gets consolidated during sleep. Experiences, which may be noisy or even contradictory facts, accumulate in a fast store called a replay buffer. Periodically, a sleep cycle evaluates those facts and selects a subset to train into a LoRA adapter, a small set of extra weights layered on top of the base model that updates only a tiny fraction of its parameters rather than retraining the whole thing. The lab tests five different policies for deciding which facts get selected for consolidation, ranging from random sampling to picking whichever facts currently surprise the model most, and it tests four defense strategies meant to prevent catastrophic forgetting, the tendency for new training to overwrite knowledge learned earlier. Those defenses include a penalty that protects weights considered important for past knowledge, keeping a running average of adapter weights across training cycles, and mixing old facts back into new training batches. Evaluation is closed book: at test time the model has no access to retrieval, so any correct answer has to come from what is actually stored in its weights. The README also documents a series of experiments and results, including a finding that an earlier evaluation method was overstating how well the model learned facts, because it was only checking whether the model repeated back the exact training sentence rather than genuinely answering a rephrased question. The project is written in Python and uses MLX for in-process training on Apple hardware. It is aimed at researchers interested in continual learning and memory consolidation for language models rather than at general users.