Offline imagination
Consolidation replays actual experience to form semantic memories. Curiosity identifies unknown states worth exploring. The dreaming system combines both capabilities into something neither provides alone: offline generation of hypothetical experiences to test abstractions, policies, and plans without acting in the environment.
The term "dreaming" is intentionally analogical rather than literal. The computational mechanism is synthetic replay and adversarial scenario generation, not a model of biological sleep. The analogy is useful because it captures the key properties: the system generates internal experiences during an offline phase, those experiences are marked as synthetic rather than observed, and they can influence future behavior through the same learning pathways as real experience.
Synthetic replay
The dream engine samples from the memory base: memories, goals, contradictions, curiosity targets, and causal model fragments. It recombines these into hypothetical scenarios that test how existing abstractions and policies perform under conditions that may not have occurred in actual experience.
Synthetic events carry a dream tag and a synthetic-replay tag so later systems can distinguish imagination from observation. This tagging is essential for epistemic hygiene: a system that cannot distinguish its own simulations from observed reality is vulnerable to confabulation, where imagined scenarios become indistinguishable from facts and reinforce incorrect beliefs.
Adversarial stress testing
The most valuable dreaming scenarios are not the ones that confirm existing beliefs but the ones that challenge them. Adversarial imagination asks: how would this strategy fail? What assumption is least well-supported? Which combination of events would most severely test the current policy?
Experiment 27 demonstrated stress testing through memory pairing. Two pairs of memories were tested:
High-contradiction pair: outage memory (contradictionRisk = 0.85) combined with db-timeout memory (contradictionRisk = 0.80). AdversarialPressure: 0.825. StressScore: 0.775. The stress formula adversarialPressure x 0.6 + (1 - min(stabilityScore)) x 0.4 produces high stress when both contradiction risk is high and minimum stability is low. This pair triggered a stress-failure flag.
Low-contradiction pair: normal memory (contradictionRisk = 0.05) combined with cache memory (contradictionRisk = 0.03). AdversarialPressure: 0.040. StressScore: 0.064. Twelve times lower stress.
The stress-failure flag from the high-contradiction pair is not a system failure; it is an informative finding. The dream engine has identified a memory combination that the system cannot currently reconcile. That is useful: it is a curiosity target (resolve the contradiction between these memories), a reflection target (why did the system form contradictory memories about this situation?), and a world-model input (scenarios involving this combination should be predicted with lower confidence).
Abstraction recombination
Dreams can combine concepts that have never co-occurred in actual experience. Some combinations are nonsensical. Others reveal general principles, novel hypotheses, or edge cases that observed experience alone might not have surfaced for years.
The value of abstraction recombination is transfer. If the system has learned a principle from one domain (authentication outages follow a specific pattern of latency, then error rate, then user reports), recombining that principle with a different domain (database connection pool exhaustion) might generate a plausible synthetic scenario where the same pattern applies. If the scenario is consistent with causal model predictions, it becomes a hypothesis worth testing. If inconsistent, it is an edge case that reveals a gap in the causal model.
Stress testing memory stability
Synthetic scenarios are compared against consolidated memories and causal models. If the system cannot explain why a synthetic outcome is plausible or impossible given its current models, that gap becomes a learning target.
This makes dreaming self-referential in a useful way: the system generates scenarios to test its own models, finds gaps, and those gaps become the targets for grounding, retrieval, and consolidation. The dreaming system is not an end in itself; it is a generator of questions that other systems then address.
The relationship to curiosity is direct: a high-stress dream scenario that the system cannot reconcile produces the same kind of uncertainty signal that the curiosity engine targets. The gap between the synthetic experience and the current model is an information-gain opportunity. The curiosity engine can then prioritize observations or experiments that would close the gap.
What dreaming cannot do
Dreaming generates hypothetical experience, not real experience. A synthetic scenario that tests a policy decision cannot produce real outcomes; it can only produce simulated ones based on the current world model. If the world model is wrong about the domain the scenario involves, the dream is testing the policy against incorrect predictions.
This means dreaming is only as useful as the models it draws on. A well-calibrated world model with rich memory and strong causal structure produces useful dreams. A poorly-calibrated world model with sparse memory and no causal structure produces scenarios that may be internally consistent but do not reflect reality.
The relationship between dreaming and grounding is therefore important. Grounding keeps the world model calibrated against observed reality. That calibration is what makes synthetic dreams meaningful. A system that dreams extensively but grounds rarely will generate elaborate scenarios that diverge progressively from actual behavior.